A64 CPUs, chipsets, motherboards

[hitechjb1]

(Long post. Please scroll down. It contains hyper-links to various subjects about the A64.)

Introduction

This thread attempts to share, highlight key and latest informations for overclocking and performance optimziation, ..., for A64 systems.

As of May 2004, there are two main second generation chipsets for the A64 CPU's, namely the Nforce3 250 (GB) from nVidia and K8T800 Pro from VIA, the Nforce3 for 939 is called Nforce3 Ultra. Currently, both are competing head to head in performance (such as HT bus speed), features (RAID, networking, ...).
- Both support 754 and 939 CPU's.
- Both announced having PCI/AGP lock
- Most Nforce3 250/250 GB motherboards reported PCI/AGP lock working
**** June 02, 04 Anandtech reported that VIA confirmed that there are PCI/AGP lock problems in some K8T800 Pro motherboards. So check test results for specific motherboards and bios carefully.

In general, the reviews for Nforce 3 250 GB are more favorable than the first generation Nforce3 150 and K8T800 (w/o Pro), in terms of stability, bios overclocking, faster HT bus (150 HT bus max around 600 MHz w/ 8 bit upstream, whereas 250 around 1000 MHz w/ 16 bit up/down-stream), HT bus overclocking, chipset built-in features such as RAID, networking, bug fixes, .... At this stage, do not get motherboards with Nforce3 150 and K8T800 (without Pro) for cost saving.

The 250 GB (compared to non-GB) chipset allows less device dependence on the PCI bus, whose bandwidth is way imbalance compared to an A64 system bandwidth (max_HT_BW to max_PCI_BW = 60:1), until PCI-express becomes main stream.

Since Sept 2004, 90 nm 939 Winchester 3000+, 3200+, 3500+ 512 KB L2 have become available. These 90 nm 939 CPU with new revision D0 (with some core enhancements ?) should be cooler and perform better than a 130 nm 754/939 NewCastle, in most cases even a 754 ClawHammer, at same clock frequency.

With a 939 motherboard, IMO, for new build, the 939 combo should be a better choice than a 754 system with a NewCastle, and even a 754 ClawHammer, especially taking into account for future compatibility, uniformity and memory intensive applications. Pricewise, a 90 nm 939 system is also as cost effective as a then 754 system. A 90 nm 939 3200+ Winchester should be a good choice for a cost effective, high performance, high bandwidth, overclocking A64 system with AGP or PCI-e.
Low PR 90 nm 939 Winchester (Sept 2004)

Two 90 nm 939 CPUs revision E3 Venice (512 KB L2) and revision E4 San Diego (1 MB L2) with SSE3 support will debut in 2005 April.
Venice and San Diego (April 2005)
About Rev E and SSE3 instructions
Major difference between Venice (E3)/San Diego (E4) and Winchester (D0)/NewCastle (CG)/ClawHammer (CG)

Socket 939 based dual core X2 Toledo and Manchester which are derived from SanDiego and Venice respectively are available since mid 2005.
Dual-core
Very preliminary dual core performance analysis
Dual core overclocking estimation from single core statistic

The 1xx 939 Opteron's are based on SanDiego (for single core) and Toledo/Denmark (for dual core), all have 1MB L2 per core. The high average stable overclockability makes them very popular.
Opteron A64 939 (90 nm SOI DSL)

Venice and San Diego are socket-compatible with Winchester, as well as the dual cores Toledo and Manchester. Most 939 platforms that apply to NewCastle 939, Winchester, FX can work with Venice, San Diego, Toledo and Manchester. Bios upgrade may be needed to correct some compatibility issues, consult manufacturers and various forums for details.

Q2 06 (?) would be expected another major upgrade for A64, the AM2 processor, based on the revision F core, with DDR2 memory support, new 940-pin socket (possibly requiring new heat sink), requiring new motherboard w/ DDR2 support.
The 940-pin socket AM2 rev. F CPU and DDR2

Details about overclocking a Winchester 3000+, DFI LP UT Nforce4 Ultra-D, TCCD memory on air:
- Memtest86 boot at 3006 MHz 1.65 V, memory 334 MHz 3-5-5-10 1T 2.8 V
- Memtest86 pass at 2970 MHz 1.62 V, memory 330 MHz 3-5-5-10 1T 2.8 V
- Windows XP boot at 2.95 GHz 1.60 V
- Sandra CPU run at 2.94 GHz 1.60 V
- SuperPI run at 2.90 GHz 1.60 V, memory 322 MHz 3-5-5-10 1T 2.8 V
- 3DMark01, ScienceMark 2.0 run at 2.85 GHz 1.55 V, memory 317 MHz 2.5-4-4-8 1T 2.8 V
- SuperPI 1M 30 sec, SuperPI 32M 27 min 40 sec
- Prime95 run stably at 2.73 GHz 1.50-1.52 V, memory 303 MHz 2.5-3-3-7 1T 2.8 V, 23 C idle, 38 C load (24+ hours, user aborted)
Overclocking sandbox for A64 939 systems with Winchester and beyond

From the specification and bios settings, I think the current DFI Nforce4 SLI-DR/Ultra-D boards have already provided most if not all the overclocking features one would want (stability, wide voltage range especially for CPU, memory, extensive memory tweaks), including chipset cooling, Vcore regulator cooling, SATA and SATA2 channels and various RAID configurations, .... IMO, barring surprise from bugs of chipset, hardware, reliability issues, most would think the wait for a nice overclocking 939 system is "over" and it would last us through Venice, San Diego, Toledo (dual core) under the 939 platform. I do not know what other manufacturers' would offer in the future for 939, but I would think any improvement in overclocking and features over the current DFI boards would be marginal and not fundamental.

The DFI, MSI Nforce4 SLI and Ultra-D are distinct from the Nforce3 Ultra in that they provide further optional dual video cards mode for boosting existing video card performance (at a cost) and/or future video performance enhancement when price of video cards drops.


New motherboards with chipsets, such as Nforce4 (regular, Ultra, SLI) from Nvidia, K8T890 from VIA, for PCI-express support are available towards end of 2004, early 2005.
Nforce4 chipsets with PCI-e, SLI features

A64 Nforce4 939 Motherboards

Memory modules (for 754 and 939 platforms) (post 16)

Typical Overclocking Systems for 939, 754 (post 2)

Overclocking setting for various bus frequencies (with a memory divider table)

Some overclocking scenarios for 939 Winchester/Venice/San Diego


Eventually, it comes down to specific motherboard implementation of the chipset, value-added features, overclocking and performance optimization features and user-friendlyness, such as
- board stability (bios, voltage line, ...)
- various voltage ranges for overclocking
- component cooling and noise

- memory, HT bus speed and overclocking potential
- number of SATA/PATA drives supported and RAID features
- (eventual) support of PCI-express (such as with Nforce4 chipset)
- sound quality

When more CPU (models) are available and are being used, the overall picture of overclockability, voltage/frequency/temperature characteristic, stepping, ... will become clearer, as price and yield become better. That would be a good time to get CPU. Eventually, one or two motherboards may become more popular for overclockers than the rest, ....

Here will attempt to highlight key informations that are important for overclocking and performance optimization, not duplicating non-essential, obvious informations, ..., keeping them up-to-date and accurate, ....


Detailed technical informations
(Please click the hyper-link, will jump to the post about a subject.)

A64 CPU's
- A64 940, 754, 939 CPU Models, OPN code, rating (post 5)
- Different CPU and system platforms (940, 754, 939) (post 4)
- A64 main features, AMD technical documents (post 3)
- Revisions and steppings (under construction) (post 6)
- Note on Absolute Max Voltage rating (post 25)
- A64 max temperature (post 71)
- Difference between a ClawHammer and a NewCastle (post 53)
- Power state of A64 desktop, mobiles (DTR, 1.4V, 1.2V) for 754 and 939 (post 58)
- A64 mobiles (How different are A64 desktop and mobiles) (post 67)
- Comments on Integrated Heat Spreader (IHS) (post 80)
- How to identify the physical core of an A64 (post 86)
- How to read A64 part number for 940, 754, 939 (post 26)
- Some links about latest silicon technology, Silicon on Insulator (SOI), Strained Silicon (SS), Dual Stress Liner (DSL)
- MOS scaling, voltage, power and leakage current
- Low PR 90 nm 939 Winchester (Sept 2004)
- Venice and San Diego (April 2005)
- About Rev E and SSE3 instructions
- Major difference between Venice (E3)/San Diego (E4) and Winchester (D0)/NewCastle (CG)/ClawHammer (CG)
- Overclocking frequecy and voltage of various A64
- Opteron A64 939 (90 nm SOI DSL)
- What are some of the differences between revisions E3, E4 and E6
- The 940-pin socket AM2 rev. F CPU and DDR2

A64 Memory subsystem
- Memory bus, cache and memory bandwidth (for 940, 754, 939)
- Memory bus frequency setting, SYNC/ASYNC mode
- Memory modules (for 754 and 939 platforms) (post 16)
- About ECC memory (for A64)
- Memory bandwidth and efficiency in terms of CPU frequency, memory frequency, CPU_memory_divider, CPU_multiplier, memory_HTT_ratio

Chipsets for A64
- Chipsets (with previews) (post 9)
- Main difference between Nforce3 250 GB/Ultra, K8T800 Pro and Sis 755/964 (post 10)
- Nforce4 chipsets with PCI-e, SLI features
- ATI Radeon XPRESS 200 PCI-express chipset for A64 Platforms
- Why native device support from chipset is better (post 73)

Motherboards for A64
- A64 Nforce4 939, Nforce3 754, 939 Motherboards (post 11)
- A64 K8T800 Pro Motherboards (754, 939) (post 12)
- (Opteron) MP motherboards (post 13)

Key technologies of A64
- Some key features of the A64 platforms (post 2)
- (eventual) 64-bit software, 32-bit X86 code compatibility
- 130/90 nm SOI silicon technology
- 512 KB/1 MB L2 cache, 2 MB L2 (future ?)
- Separation of memory bus and HT system bus at CPU level
- HyperTransport (post 15)
... spec 1600 MT/s (800 MHz w/ DDR), 6.4 GB/s max (32-bit one way) for 940, 754
... spec 2000 MT/s (1000 MHz w/ DDR), 8 GB/s max (32-bit one way) for 939
- On-chip north bridge and memory controller (Opteron, FX and 939 have 128-bit dual channel, 754 has 64-bit)
- Opteron is basically FX with extra capabilites (such as coherent HT links) for MP/mission-critical systems
- PCI-express and devices (post 17)
- DDR2 memory module support (late 05 ?) (post 18)
- Dual-core A64 processor (post 48)

System setup and performance evaluation
- Typical Overclocking Systems for 939, 754 (post 2)
- Overclocking setting for various bus frequencies (post 8)
- Relationship between CPU_memory_divider and CPU_multiplier, memory_HTT_ratio (post 60)
- Example to setup frequencies for CPU and memory
- How to choose memory divider and memory_HTT_ratio
- PSU rating estimate for some 939 CPU and system
- Some overclocking scenarios for 939 Winchester/Venice/San Diego
- Overclocking frequecy and voltage of various A64
- Very preliminary dual core performance analysis


- The tradeoff between CPU and memory bus frequencies on performance (post 62)
- Performance analysis of various A64 systems (including Barton, P4's) (post 7)
- How to compare ClawHammer and NewCastle (post 55)
- Comparing Sempron (754 256 KB L2, pre-release?) with A64 (754 512 KB L2) (post 61)
- Stability testing using memtest, SuperPI (32M) and Prime95


Overclocking techniques
- Overlocking A64's 101 by Gautam

A64 system: setups, results and experiences
- Exemplary A64 systems: setups, results and experiences (post 63)

Please enter your A64 result(s) here, it will help us all long term.
A64 overclocking result collection


Anticipated major milestones
- Nforce3 150, K8T800 for A64 754 platform (past)

- 2H 04 (starting May/June) would be for good price performance system
... A64 754
... Nforce3 250 GB motherboard
... K8T800 Pro + VT8237 (may have PCI/AGP lock problem)
... AGP 8X video card, supports unbuffered DDR 400/500 memory module

- 2H 04 (starting June/July ?) would be for top memory bandwidth system
... A64 939 130 nm SOI
... Nforce3 250 GB/Nforce3 Ultra motherboard
... K8T800 Pro + VT8237 (may have PCI/AGP lock problem)
... AGP 8X video card, supports unbuffered DDR 400/500 memory module

- Towards year end 04/early 05 would be for
... A64 939, better CPU price and yield for 130 nm SOI, 90 nm SOI with new revision
... new motherboard w/ PCI-express, such as with Nvidia Nforce4 (regular, Ultra, SLI), VIA K8T890 chipsets
... new PCI-express video cards
... supports unbuffered DDR 400/500 memory module
... rev E A64 with SSE3 instruction supports
... rev E dual core for socket 939 (2H 05)

- Q2 06 would be expected another major upgrade, the M2 (AM2) processor
... dual-core A64
... revision F cores
... DDR2 memory
... new 940-pin socket
... new motherboard w/ DDR2 support
... http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2476


These are the major milestones for upgrade/new build. It does not imply one has to upgrade everytime along the way and right at the beginning. Make the decision based on need, price-performance, .... Listed time is for reference only and may be off.



Links of interest (links may change over time for currency)

AMD CPU roadmap
http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_608,00.html

AMD A64 Performance Guide
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/31783.pdf (FX-55)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/31366.pdf (4000+)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/31225.pdf (3400+)

C627627's user friendly AMD CPU roadmap
http://www.c627627.com/AMD/Athlon64/
http://www.c627627.com/AMD/OpteronAthlon64/
http://www.c627627.com/AMD/AthlonXP/


AMD Athlon 64 3800+ CPU: E3 Processor Core aka Venice at the Door

FX-55 and 4000+
http://www.ocforums.com/showthread.php?p=3161495#post3161495
http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2249&p=1
The 4000+ is basically a FX-53 with locked multiplier above x12. It has dual channel, 1 MB L2, and is rated at 2.4 GHz.

Review of 90 nm 939 from Anandtech (10/14/04)
http://www.xbitlabs.com/articles/cpu/display/athlon64-90nm.html
90nm Processors from AMD: Athlon 64 3500+, 3200+ and 3000+
http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2242
Review of 90 nm 3200+ Winchester, comparing with 3400+ CH, 3500+ 939, FX-53 all at same clock frequencies (10/11/04)
http://www.madshrimps.be/?action=getarticle&articID=230
AMD 90nm power consumption measured
http://techreport.com/onearticle.x/7417

939 CPU review (3800+, FX-53) (06/01/04)
http://www.anandtech.com/cpu/showdoc.html?i=2065

Benchmarks comparing
939 FX-53, 939 3800+ (512KB L2), 754 3700+ (1MB L2), 939 3500+ (512KB L2), XP 2400+ (256KB L2)
P4 3.4EE (1MB L2, 2MB L3), P4 3.4 Prescott (1MB L2), P4 2.4C (512KB L2)
http://hardocp.com/articleprint.html?article_id=626

Benchmarking various A64's, Barton, P4's, .... (03/18/04)
http://www.techreport.com/reviews/2004q1/athlon64-fx53/index.x?pg=1

Benchmarking of a A64 FX-53 (rated 2.4 GHz, 1 MB L2, dual channel) w/ a Barton 2500+, Prescott 3.2E (03/17/04)
http://bit-tech.net/review/309/1

Benchmarking of a A64 754 3000+ (rated 2.0 GHz, 512 KB L2, rev C0) (01/21/04)
http://www.techreport.com/reviews/2004q1/athlon64-3000/index.x?pg=1

...

 

[hitechjb1]

Some key features of the A64 platforms

For A64 platform, it is a new generation of
- CPU architecture (64-bit and associated features, on-chip north bridge and memory controller, larger L2 cache, ...)
- silicon technology (130 nm SOI and then 90 nm SOI, SSDOI)
- system technology (separate memory and HT system bus, and many new devices into the future, e.g. PCI-express)
- OS and software (64-bit OS and applications, the A64 works perfectly w/ x86-32 bit software)
- more features in chipset and motherboards (e.g. faster serial link, native raid 0/1/0+1/JBOD, faster native network supports, ...)
....

- A64 has many new CPU architectural features, more raw power, higher stock CPU frequency, scalable to next generation 90 nm SOI silicon technology (overclockability)

- A64 platform replaces the single system bus (aka FSB) of XP by two SEPARATE buses, namely a memory bus and a HyperTransport HT system bus (connecting to all system devices via the on-chip north bridge).

- As a result, the system bandwidith would be two to four times that of an XP system running the same bus frequency. The HT bus specification to 800 MHz w/ DDR for 754, to 1000 MBz w/ DDR for 939, 32-bit, a max bandwidth of 6.4 GB/s. The dual channel 128-bit memory bus (for 939/940) has a max bandwidth of 6.4 GB/s. The single channel 64-bit memory bus (for 754) has a max bandwidth of 3.2 GB/s.

- The memory controller for A64 is on the CPU chip, hence access time for memory read/write is reduced compared to having the memory controller on the chipset. As such and the separate memory bus and system bus, there is less bus conflict and the effective memory latency between the CPU (after L2 miss) and the memory (L3) is reduced.

Differences between the XP FSB and the A64 buses (separate memory bus and HyperTransport bus) (page 19)


Typical Overclocking Systems for 939, 754

As of end of July 2004, due to CPU price drop of 754 and 939 CPU, the price difference between a starting 754 system and a starting 939 system is about $150 - 200 (the price difference of 2800+/3000+/3200+ and a 3500+). I think building a 939 system (non PCI-e) is becoming attractive, especially for high memory bandwidth applications and for 939 compatibility. Lower PR 939 may be coming too for further lowering of price/performance.

Low PR 90 nm 939 Winchester (Sept 2004)

Venice and San Diego (April 2005)

Major difference between Venice (E3)/San Diego (E4) and Winchester (D0)/NewCastle (CG)/ClawHammer (CG)

Venice and San Diego are socket-compatible with Winchester. Most 939 platforms that apply to NewCastle 939, Winchester, FX can work with Venice and San Diego.

New motherboards with chipsets (such as Nforce4 from Nvidia, K8T890 from VIA) for PCI-express support will be available towards end of 2004.

Options:
1. Need to build system now and have AGP video card: It is true that 754 has more choice of motherboards, but the MSI K8N Neo2 Platinum, though may not be perfect in every aspect, is a viable choice with a 90 nm Winchester 3000+/3200+ for current build (3200+ is perferred for its flexible in setting memory and allowing lower HTT to overclock CPU in case).
2. Using PCI-e video card: If not reusing existing AGP video card, better to wait for a NF4 PCI-e board (either SLI or Ultra) so the latest PCI-e video cards can be used down the road.


939 Price performance system (AGP or PCI-e) (Sept 2004)

- A64 939 Venice 512KB L2 (3000+/3200+/3500+/3800+) or San Diego 1MB L2 (3700+/4000+) (April 2005)

- A64 939 3000+ (x9) or 3200+ (x10) 90 nm 512 KB L2 Winchester (3500+ Winchester available also)
.... preferably week 044x or after CBBHD

- two choices for motherboards: AGP or PCI-e
.... AGP based motherboard: if staying with AGP video card, Nforce3 Ultra motherboard can be used, e.g. MSI K8N Neo2 Platinum
.... PCI-e based motherboard:
........ ASUS Nforce4 A8N SLI-deluxe
........ MSI Neo4 Platinum (Nforce4 Ultra), MSI Neo4 Diamond (Nforce4 SLI)
........ DFI LanParty UT Nforce4 Ultra-D, DFI LanParty Nforce4 SLI-DR
........ PCI-e video card, e.g. Nvidia 6800 GT/6600 GT, or ATI X800XL
........ the SLI version supports dual video card configuration, some Ultra boards (e.g. MSI, DFI) can potentially run dual video cards
.... PSU for PCI-e system:
........ requires 24-pin power connector, higher 12 V current of estimated 28 A for single video card, 30-33 A for dual video card
........ Fortron Blue Storm AX500-A, Antec True Control II 550
PSU rating estimate for some 939 CPU and system

- 2x512 MB DDR500+ dual channel or overclock equivalent,
e.g. modules w/ Samsung TCCD DRAM chips
G. Skill PC4400/4800, PQI Turbo PC3200, OCZ rev.2 Platinum, Corsair 4400C25
(some TCCD from low 220-230 MHz with 2/2.5-2-2-x 1T, to 250 MHz with 2.5-3-2/3-x 1T, to 280-300 MHz 2.5-3-3-x 1T, to 300+ 2.5/3-3/4-3/4-x 1T at ~2.8 V)
BH-5/UTT based modules, 250-260 MHz 2-2-2-x 1T 3.3+ V

- SLK-948U or XP-90 or XP-90C (copper) or XP-120 (check for motherboard compatibility first)

e.g. 3200+ Winchester, x10, HTT >= 250 MHz, memory (1:1) >= 250 MHz 1T, CPU >= 2.5 GHz (to 3 GHz).
A 3200+ with x10 multiplier is more flexible than a 3000+ with x9 (x9 is doable) for setting up CPU (between 2.5 - 3 GHz), HTT and memory bus.

Price performance system (June 2004)

- A64 754 1 MB L2 ClawHammer, CPU revision CG
- A64 754 512 KB L2 NewCastle, CPU revision CG
(as price moves down, move to higher PR and multiplier with small price differential)
- with x9 multiplier and below (for air/water cooling), e.g. 3000+ ClawHammer rated 1.8 GHz, 2800+ NewCastle rated 1.8 GHz
- with x10 multiplier and below (for air/water/extreme cooling), e.g. 3200+ ClawHammer rated 2.0 GHz, 3000+ NewCastle rated 2.0 GHz
- 1.5V desktop or 1.5V mobile DTR or 1.4V mobile CPU
- 250 GB motherboard <- such as MSI K8N Neo Platinum, EPOX 8KDA3+, ABIT KV8 Pro rev 1.1
- 2x512 MB DDR500 single channel or overclock equivalent
- SLK-948U
- As of June 04, (2800/3000/3200+HS+mb+memory) cost around $550 - 650.
(assume 2800+/3000+/3200+ $180 - $280, HS $40, mb $130, memory $200)
- As of end of July 04, 2800+/3000+/3200+ price reduction by $50-100, so (2800/3000+HS+mb+memory) cost around $500-550.

Target setup:
BIOS HTT setting: 250 - 300 MHz
CPU 1 MB L2 ClawHammer CG revision
- x9, target to 2.25 - 2.7 GHz (3000+ 754 ClamHammer, 2800+ 754 NewCastle), memory at 250 - 300 MHz, effective BW = 3800 - 4560 MB/s
- x10, target to 2.5 - 3 GHz (3200+ 754 ClawHammer, 3000+ 754 NewCastle)
memory at 250 - 300 MHz, effective BW = 3800 - 4560 MB/s
HT at x3, at least 750 - 900 MHz w/ DDR, or x4 from 1000 and up

For x8 2800+ ClawHammer, the 300 MHz HTT may not be able to max out CPU
- x8, target to 2.0 - 2.4 GHz (2800+ 754 ClawHammer)

High performance 939 system (June 2004)

- A64 939 512 KB L2, CPU revsion CG
(as price moves down, move to higher PR and multiplier with small price differential)
- 3500+, rated 2.2 GHz with x11 multiplier and below
- 1.5V desktop
- 939 motherboard <- such as MSI K8N Neo2 Platinum, ABIT AV8, ASUS A8V deluxe rev. 2, Gigabyte GA-K8NSNXP
- 2x512 MB DDR500 dual channel or overclock equivalent
- SLK-948U
- As of June 04, (3500+HS+mb+memory) cost around $850. (assume 3500+ $500)
- As of end of July 04, 939 3500+ price reduction by $150, (3500+HS+mb+memory) cost around $700 (assume CPU $350).

High performance 939 FX system (June 2004)

- A64 FX 939 1 MB L2, CPU revsion CG
(as price moves down, move to higher PR and multiplier with small price differential)
- FX53, rated 2.4 GHz, with multiplier unlocked
- 1.5V desktop
- 939 motherboard <- such as MSI K8N Neo2 Platinum, ABIT AV8, ASUS A8V deluxe rev. 2, Gigabyte GA-K8NSNXP
- 2x512 MB DDR500 dual channel or overclock equivalent
- SLK-948U
- As of June 04, (FX53+HS+mb+memory) cost around $1200.
(assume FX53 - $800, +$50 on memory)

*** K8T800 Pro may still have PCI/AGP lock problem. So check test results for specific motherboards and bios carefully.


Links to reviews of other 754, 939 boards:

A64 Nforce3 Motherboards (754, 939)

A64 K8T800 Pro Motherboards (754, 939)

Overclocking setting for various bus frequencies (post 8)

 

[hitechjb1]

A64 main features

QUOTE from AMD Tech Doc

• Compatible with Existing 32-Bit Code Base
– Including support for SSE, SSE2, MMX™, 3DNow!™ technology and legacy x86 instructions
– Runs existing operating systems and drivers
– Local APIC on-chip
• AMD64 Technology
– AMD64 technology instruction set extensions
– 64-bit integer registers, 48-bit virtual addresses, 40-bit physical addresses
– Eight new 64-bit integer registers (16 total)
– Eight new 128-bit SSE/SSE2 registers (16 total)
• Integrated Memory Controller
– Low-latency, high-bandwidth
– 72-bit DDR SDRAM at 100, 133, 166, and 200 MHz (FOR 754)
– 144-bit DDR SDRAM at 100, 133, 166, and 200 MHz (FOR 939 / 940)
– Supports up to four registered DIMMs or up to three unbuffered DIMMs
• HyperTransport™ Technology to I/O Devices
– One 16-bit link supporting speeds up to 800 MHz (1600 MT/s) or 3.2 Gigabytes/s in each direction (for 940/754)
- One 16-bit link supporting speeds up to 1000 MHz (2000 MT/s) or 4.0 Gigabytes/s in each direction (for 939)
• 64-Kbyte 2-Way Associative ECC-Protected L1 Data Cache

– Two 64-bit operations per cycle, 3-cycle latency
• 64-Kbyte 2-Way Associative Parity-Protected L1 Instruction Cache
– With advanced branch prediction
• 16-Way Associative ECC-Protected L2 Cache
– Exclusive cache architecture—storage in addition to L1 caches
– 256-Kbyte, 512-Kbyte, and 1-Mbyte options
• Machine Check Architecture
– Includes hardware scrubbing of major ECC-protected arrays
• Power Management
– Multiple low-power states
– System Management Mode (SMM)
– ACPI compliant, including support for processor performance states

Omit Electrical Interfaces and Packaging, can be found in Data Sheet.



A64 CPU technical documents

Spec quick finder
http://www.amdcompare.com/us-en/desktop/

Revision Guide for AMD AthlonTM 64 and AMD OpteronTM Processors
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/25759.pdf

AMD Athlon 64 Processor Power and Thermal Data Sheet (30430)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf
It contains A64, A64 FX, mobile A64 specifications.

AMD Athlon 64 Processor Data Sheet for 754 and 939 (24659)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/24659.PDF
AMD Athlon 64 FX Processor Data Sheet for 940 and 939 (30431)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30431.pdf

AMD Functional Data Sheet, 754 Pin Package (31410)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/31410.pdf
AMD Functional Data Sheet, 939 Pin Package (31411)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/31411.pdf
AMD Functional Data Sheet, 940 Pin Package (31412)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/31412.pdf

AMD Socket 940 Design Specification (25766)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/25766.pdf
Clock Generator Specification for AMD64 Processors (24707)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/24707_PUB.PDF

AMD Athlon 64 and AMD Opteron Processors Thermal Design Guide (26633)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/26633_5649.pdf
Builder’s Guide for AMD Athlon™ 64 Processor-Based Desktops and Workstations (31684)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/31684.pdf

BIOS and Kernel Developer's Guide for AMD AthlonTM 64 and AMD OpteronTM Processors (26094)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/26094.PDF

AMD Processor Pricing
http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_609,00.html


Link to AMD Athlon™ 64 Processor Tech Docs
http://www.amd.com/us-en/Processors/TechnicalResources/0,,30_182_739_7203,00.html
Link to AMD Product Information
http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118,00.html


Link to Intel Products and Services
http://www.intel.com/products/index.htm?iid=HPAGE+header_products&
http://www.intel.com/products/processor_number/processor_numbers.htm

Link to an interesting article on AMD and Intel CPU code names
http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2178

....

 

[hitechjb1]

Different CPU and system platforms (754, 939, 940)

There are three main platforms, namely CPU's with 754, 939 and 940 sockets. They are literally refer to the pin count of the CPU/socket, but ther are implications on system cost, performance, scalability into the future, .... 940, 754 were first introduced. 939 was introduced in the middle of 04.

Socket 754
- 512 KB L2 or 1 MB L2
- 64-bit based memory bus (single channel)
- Unbuffered and registered DIMMs with a 64-bit data bus with optional 8 bits of Error Correcting Code (ECC)
(can work with regular, unbuffered memory modules for non-mission critical system)
- Up to three unbuffered DIMMs and four registered DIMMs

Socket 940 (FX or Opteron)
- 1 MB L2 (only)
- 128-bit based memory bus (dual channel)
- Stacked registered DIMMs (requires registered memory modules)
- ECC checking with single-bit correction and double-bit detection
- Opteron has provision for MP configuration, via coherent HT links

Socket 939 (A64 or A64 FX)
- 512 KB L2 or 1 MB L2
- 128-bit based memory bus (dual channel)
- Up to four unbuffered DIMMs in a 128-bit configuration, or up to two unbuffered DIMMs in a 64-bit configuration
(can work with regular, unbuffered memory modules for non-mission critical system)
- ECC checking with double-bit detect with single-bit correct (optional)
- Speculated to include dual-core configuration in 2005.

Opteron is included in the context of socket 940 and A64 FX.
An Opteron is basically an A64 FX (1 MB L2, dual channel memory controller w/ 128-bit memory bus) with the same internal core (SledgeHammer) with a few extra capabilities such as the coherent HT links used for connecting to other processors via HT bus for building multi-processor system.


Possible usage

- Socket 754 can be considered for price/performance and early adaption

- Socket 940 covers the Opteron and FX CPU. The socket 940 Opteron is for MP and servers applications. Since connecting for MP, the HT links and ports of the CPU require coherent links and they are different than that of a FX. The internal core of a FX is the same as that of an Opteron.

- Socket 939 can be considered for higher end, its scalability and future compatibility of motherboard features and CPU (price and yield) into the future, speculated to include dual-core configuration in 2005.

- Theoretically, 939/940 always delivers better peformance than a 754, especially on memory bandwidth and memory intensive applications, if price-performance is not a major consideration factor.

- Both 754 and 939 have L2 cache size 512 KB or 1 MB, 940 has only 1 MB L2. Eventually max L2 size will be increased to 2 MB.

- The total system bandwidth (memory + HT) is about two times that of XP system for 754, four times for 939.

- A major difference between 754 and 939/940 is the memory bus and memory bandwidth.
... memory bus
...... 128-bit based dual channel memory bus for 939/940 (actual 128 + 16-bit ECC, total 144-bit)
...... 64-bit based memory bus for 754 (actual 64 + 8-bit ECC, total 72-bit)
... memory module
...... Non ECC and unbuffered memory modules supported by both 939 and 754
...... 940 requires registered memory modules, and optional ECC memory preferred for mission-critical servers
... Memory bandwidth efficiency of a typical 754 system is 95%+.
... Memory bandwidth efficiency of a typical 939 system is 86%+, recent 939 to 90%+ seen.
... Effective memory bandwidth of 939/940 is estimated to be 80+% higher than that of 754.

 

[hitechjb1]

This is a summary of some Desktop A64, Mobile A64 DTR and Mobile A64, A64 FX, A64 X2 (dual core), Opteron

A64 940, 754, 939, AM2 940 CPU Models, OPN code, rating


Desktop A64 X2 AM2 940 (90 nm SOI DSL) Windsor (UNDER CONSTRUCTION)

FX-62: ADAFX62IAA6CS 1.30/1.35V (JH F2 rev, 00020F32h) Windsor, 2x1 MB L2, 2.8 GHz, x14, 125 W
5200+: ADA5200IAA6CS 1.30/1.35V (JH F2 rev, 00020F32h) Windsor, 2x1 MB L2, 2.6 GHz, x13, 89 W
4800+: ADA4800IAA6CS 1.30/1.35V (JH F2 rev, 00020F32h) Windsor, 2x1 MB L2, 2.4 GHz, x12, 65 W
4400+: ADA4400IAA6CS 1.30/1.35V (JH F2 rev, 00020F32h) Windsor, 2x1 MB L2, 2.2 GHz, x11, 65 W
4000+: ADA4000IAA6CS 1.30/1.35V (JH F2 rev, 00020F32h) Windsor, 2x1 MB L2, 2.0 GHz, x10, 65 W

5000+: ADA5000IAA5CU 1.30/1.35V (BH F2 rev, 00020FB2h) Windsor, 2x512 KB L2, 2.6 GHz, x13, 89 W
4600+: ADA4600IAA5CU 1.30/1.35V (BH F2 rev, 00020FB2h) Windsor, 2x512 KB L2, 2.4 GHz, x12, 65 W
4200+: ADA4200IAA5CU 1.30/1.35V (BH F2 rev, 00020FB2h) Windsor, 2x512 KB L2, 2.2 GHz, x11, 65 W
3800+: ADA3800IAA5CU 1.30/1.35V (BH F2 rev, 00020FB2h) Windsor, 2x512 KB L2, 2.0 GHz, x10, 65 W

Opteron A64 939 (90 nm SOI DSL)
- The 144, 146, 148, 150, 152, 165. 170, 175 Opterons are for non-SMP.
- The corresponding 244, 246, 248, 250, 252, 265, 270, 275 and 844, 846, 848, 850, 852, 865, 870, 875 Opterons are the respectively 2-way and 4 to 8-way SMP versions.
- The specification of the various 1xx and 2xx and 8xx are the same, except the 1xx's are for non-SMP. With coherent HT links in 2xx and 8xx, the 2xx's are validated for 2-way SMP, and the 8xx's are validated for 4-8 way SMP.
- Should be able to work with non-ECC unbuffered memory, ECC and buffered memory modules.

Single core:
144: OSA144DAA5BN 1.35/1.4V (SH E4 rev, 00020F71h) SanDiego, 1 MB L2, 1.8 GHz, x9, 67 W
146: OSA146DAA5BN 1.35/1.4V (SH E4 rev, 00020F71h) SanDiego, 1 MB L2, 2.0 GHz, x10, 67 W
148: OSA148DAA5BN 1.35/1.4V (SH E4 rev, 00020F71h) SanDiego, 1 MB L2, 2.2 GHz, x11, 85(?) W
150: OSA150DAA5BN 1.35/1.4V (SH E4 rev, 00020F71h) SanDiego, 1 MB L2, 2.4 GHz, x12, 85(?) W
152: OSA152DAA5BN 1.35/1.4V (SH E4 rev, 00020F71h) SanDiego, 1 MB L2, 2.6 GHz, x13, 104(?) W

Dual core:
165: OSA165DAA6CD 1.35/1.4V (JH E6 rev, 00020F32h) Toledo (Denmark), 2x1 MB L2, 1.8 GHz, x9, 110 W
170: OSA170DAA6CD 1.35/1.4V (JH E6 rev, 00020F32h) Toledo (Denmark), 2x1 MB L2, 2.0 GHz, x10, 110 W
175: OSA175DAA6CD 1.35/1.4V (JH E6 rev, 00020F32h) Toledo (Denmark), 2x1 MB L2, 2.2 GHz, x11, 110 W

Desktop A64 X2 939 (90 nm SOI DSL) Toledo
3800+: ADA3800DAA5CD 1.35/1.4V (JH E6 rev, 00020F32h) Toledo, 2x512 KB L2, 2.0 GHz, x10, 89 W (512 KB L2 per core "disabled")
4400+: ADA4400DAA6CD 1.35/1.4V (JH E6 rev, 00020F32h) Toledo, 2x1 MB L2, 2.2 GHz, x11, 110 W
4800+: ADA4800DAA6CD 1.35/1.4V (JH E6 rev, 00020F32h) Toledo, 2x1 MB L2, 2.4 GHz, x12, 110 W
FX-60: ADAFX60DAA6CD 1.35/1.4V (JH E6 rev, 00020F32h) Toledo, 2x1 MB L2, 2.6 GHz, x13, 110 W

Desktop A64 X2 939 (90 nm SOI DSL) Manchester
3800+: ADA3800DAA5BV 1.35/1.4V (BH E4 rev, 00020FB1h) Manchester, 2x512 KB L2, 2.0 GHz, x10, 89 W
4200+: ADA4200DAA5BV 1.35/1.4V (BH E4 rev, 00020FB1h) Manchester, 2x512 KB L2, 2.2 GHz, x11, 89 W
4600+: ADA4600DAA5BV 1.35/1.4V (BH E4 rev, 00020FB1h) Manchester, 2x512 KB L2, 2.4 GHz, x12, 89 W

Desktop A64 939 (90 nm SOI DSL) San Diego
3500+: ADA3500DAA4BN 1.35/1.4V (SH E4 rev, 00020F71h) SanDiego, 512 KB L2, 2.2 GHz, x11, 67 W (512 KB L2 "disabled")
3700+: ADA3700DAA5BN 1.35/1.4V (SH E4 rev, 00020F71h) SanDiego, 1 MB L2, 2.2 GHz, x11, 89 W
4000+: ADA4000DAA5BN 1.35/1.4V (SH E4 rev, 00020F71h) SanDiego, 1 MB L2, 2.4 GHz, x12, 89 W

Desktop A64 939 (90 nm SOI DSL) Venice
Rev E3 Venice
3000+: ADA3000DAA4BP 1.35/1.4V (DH E3 rev, 00020FF0h) Venice, 512 KB L2, 1.8 GHz, x9, 67 W
3200+: ADA3200DAA4BP 1.35/1.4V (DH E3 rev, 00020FF0h) Venice, 512 KB L2, 2.0 GHz, x10, 67 W
3500+: ADA3500DAA4BP 1.35/1.4V (DH E3 rev, 00020FF0h) Venice, 512 KB L2, 2.2 GHz, x11, 67 W
3800+: ADA3800DAA4BP 1.35/1.4V (DH E3 rev, 00020FF0h) Venice, 512 KB L2, 2.4 GHz, x12, 89 W
Rev E6 Venice
3000+: ADA3000DAA4BW 1.35/1.4V (DH E6 rev, 00020FF2h) Venice, 512 KB L2, 1.8 GHz, x9, 67 W
3200+: ADA3200DAA4BW 1.35/1.4V (DH E6 rev, 00020FF2h) Venice, 512 KB L2, 2.0 GHz, x10, 67 W
3500+: ADA3500DAA4BW 1.35/1.4V (DH E6 rev, 00020FF2h) Venice, 512 KB L2, 2.2 GHz, x11, 67 W
3800+: ADA3800DAA4BW 1.35/1.4V (DH E6 rev, 00020FF2h) Venice, 512 KB L2, 2.4 GHz, x12, 89 W

Desktop A64 939 (90 nm SOI) Winchester
3000+: ADA3000DIK4BI 1.4V (DH8 D0 rev, 00010FF0h) Winchester, 512 KB L2, 1.8 GHz, x9, 67 W
3200+: ADA3200DIK4BI 1.4V (DH8 D0 rev, 00010FF0h) Winchester, 512 KB L2, 2.0 GHz, x10, 67 W
3500+: ADA3500DIK4BI 1.4V (DH8 D0 rev, 00010FF0h) Winchester, 512 KB L2, 2.2 GHz, x11, 67 W

Desktop A64 939 (130 nm SOI)
3500+: ADA3500DEP4AW 1.5V (DH7 CG rev, 00000FF0h) "NewCastle 939", 512 KB L2, 2.2 GHz, x11, 89 W
3500+: ADA3500DEP4AS 1.5V (SH7 CG rev, 00000F7Ah) "ClawHammer 939", 512 KB L2, 2.4 GHz x11, 89 W ** (512 KB L2 "disabled")
3800+: ADA3800DEP4AW 1.5V (DH7 CG rev, 00000FF0h) "NewCastle 939", 512 KB L2, 2.4 GHz, x12, 89 W
4000+: ADA4000DEP5AS 1.5V (SH7 CG rev, 00000F7Ah) "ClawHammer 939", 1 MB L2, 2.4 GHz x12, 89 W **

Desktop A64 754 (130 nm SOI)
2800+: ADA2800AEP4AP 1.5V (SH7 C0 rev, 00000F48h) ClawHammer, 512 KB L2, 1.8 GHz, x9, 89 W (512 KB L2 "disabled")
3000+: ADA3000AEP4AP 1.5V (SH7 C0 rev, 00000F48h) ClawHammer, 512 KB L2, 2.0 GHz, x10, 89 W (512 KB L2 "disabled")
3200+: ADA3200AEP5AP 1.5V (SH7 C0 rev, 00000F48h) ClawHammer, 1 MB L2, 2.0 GHz, x10, 89 W
3400+: ADA3400AEP5AP 1.5V (SH7 C0 rev, 00000F48h) ClawHammer, 1 MB L2, 2.2 GHz, x11, 89 W
2800+: ADA2800AEP4AR 1.5V (SH7 CG rev, 00000F4Ah) ClawHammer, 512 KB L2, 1.8 GHz, x9, 89 W (512 KB L2 "disabled") **
3000+: ADA3000AEP4AR 1.5V (SH7 CG rev, 00000F4Ah) ClawHammer, 512 KB L2, 2.0 GHz, x10, 89 W (512 KB L2 "disabled")
3200+: ADA3200AEP5AR 1.5V (SH7 CG rev, 00000F4Ah) ClawHammer, 1 MB L2, 2.0 GHz, x10, 89 W
3400+: ADA3400AEP5AR 1.5V (SH7 CG rev, 00000F4Ah) ClawHammer, 1 MB L2, 2.2 GHz, x11, 89 W
3700+: ADA3700AEP5AR 1.5V (SH7 CG rev, 00000F4Ah) ClawHammer, 1 MB L2, 2.4 GHz, x12, 89 W

2800+: ADA2800AEP4AX 1.5V (DH7 CG rev, 00000FC0h) NewCastle, 512 KB L2, 1.8 GHz, x9, 89 W
3000+: ADA3000AEP4AX 1.5V (DH7 CG rev, 00000FC0h) NewCastle, 512 KB L2, 2.0 GHz, x10, 89 W
3200+: ADA3200AEP4AX 1.5V (DH7 CG rev, 00000FC0h) NewCastle, 512 KB L2, 2.2 GHz, x11, 89 W
3400+: ADA3400AEP4AX 1.5V (DH7 CG rev, 00000FC0h) NewCastle, 512 KB L2, 2.4 GHz, x12, 89 W **

Mobile A64 754 (DTR) (130 nm SOI)
3000+: AMA3000BEX5AR 1.5V (SH7 CG rev, 00000F4Ah) ClawHammer, 1 MB L2, 1.8 GHz, x9, 81.5 W
3200+: AMA3200BEX5AR 1.5V (SH7 CG rev, 00000F4Ah) ClawHammer, 1 MB L2, 2.0 GHz, x10, 81.5 W
3400+: AMA3400BEX5AR 1.5V (SH7 CG rev, 00000F4Ah) ClawHammer, 1 MB L2, 2.2 GHz, x11, 81.5 W
3700+: AMA3700BEX5AR 1.5V (SH7 CG rev, 00000F4Ah) ClawHammer, 1 MB L2, 2.4 GHz, x12, 81.5 W

Mobile A64 754 1.4V (130 nm SOI)
2800+: AMN2800BIX5AR 1.4V (SH7 CG rev, 00000F4Ah) ClawHammer, 1 MB L2, 1.6 GHz x8, 62 W
3000+: AMN3000BIX5AR 1.4V (SH7 CG rev, 00000F4Ah) ClawHammer, 1 MB L2, 1.8 GHz x9, 62 W
3200+: AMN3200BIX5AR 1.4V (SH7 CG rev, 00000F4Ah) ClawHammer, 1 MB L2, 2.0 GHz x10, 62 W

Mobile A64 754 1.2V (130 nm SOI)
2700+: AMD2700BQX4AX 1.2V (DH7 CG rev, 00000FC0h) NewCastle, 512 KB L2, 1.6 GHz, x8, 35 W
2800+: AMD2800BQX4AX 1.2V (DH7 CG rev, 00000FC0h) NewCastle, 512 KB L2, 1.8 GHz, x9, 35 W

A64 FX 939 (90 nm SOI DSL)
FX57: ADAFX57DAA5BN 1.35/1.4V (SH8 E4 rev, 00020F71h) SanDiego, 1 MB L2, 2.8 GHz, x14, 104 W
FX55: ADAFX55DAA5BN 1.35/1.4V (SH8 E4 rev, 00020F71h) SanDiego, 1 MB L2, 2.6 GHz, x13, 104 W

A64 FX 939 (130 nm SOI DSL)
FX55: ADAFX55DEI5AS 1.5V (SH7 CG rev, 00000F7Ah) "ClawHammer 939", 1 MB L2, 2.6 GHz x13, 104 W

A64 FX 939 (130 nm SOI)
FX53: ADAFX53DEP5AS 1.5V (SH7 CG rev, 00000F7Ah) "ClawHammer 939", 1 MB L2, 2.4 GHz x12, 89 W

A64 FX 940 (130 nm SOI)
FX51: ADAFX51CEP5AK 1.5V (SH7 C0 rev, 00000F58h) SledgeHammer, 1 MB L2, 2.2 GHz x11, 89 W
FX51: ADAFX51CEP5AT 1.5V (SH7 CG rev, 00000F5Ah) SledgeHammer, 1 MB L2, 2.2 GHz x11, 89 W
FX53: ADAFX53CEP5AT 1.5V (SH7 CG rev, 00000F5Ah) SledgeHammer, 1 MB L2, 2.4 GHz x12, 89 W

** Not in AMD released tech doc (yet), but seems to be available

How to identify the physical core of an A64 (post 86)


Spec quick finder from AMD:
http://www.amdcompare.com/us-en/desktop/

Revision Guide for AMD AthlonTM 64 and AMD OpteronTM Processors (25759)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/25759.pdf

The following links detail about the models (OPN) and specifications:

A64 Model (OPN) for 754, 939, FX, including desktop, mobile DTR, low voltage:
How to read A64 part number for 940, 754, 939 (post 26)

http://www.amdboard.com/amd64_opn.html

A64 939 spec:
http://www.ocforums.com/showpost.php?p=2760928

A64 754 spec:
http://www.ocforums.com/showpost.php?p=2760930

A64 Mobile 754 DTR, 1.4V, 1.2V spec:
http://www.ocforums.com/showpost.php?p=2760959

A64 FX (940, 939) spec:
http://www.ocforums.com/showpost.php?p=2761277

 

[hitechjb1]

A64 Revisions and steppings

Officially,
- C0, CG are referred to revision in the AMD technical doc for ClawHammer, NewCastle, SledgeHammer, ....
- D0 is the revision for the 90 nm Winchester 939 CPU, available since Sept 04.
- E3 is the revision for the 90 nm 939 Venice.
- E4 is the revision for the 90 nm 939 SanDiego, X2 Manchester, SanDiego based Opteron.
- E6 is the revision for the 90 nm 939 X2 Toledo, new Venice, Toledo based Opteron.
- F2 is the revision for the 90 nm AM2 940 Windsor

Regarding to
- the famous 5-letter codes (such as JIUHB, IQYFA, IQYHA, ...) that we have seen in XP/Barton, and
- the popular 4-letter code (such as MPMW, RPMW, SPMW, TPMW, UPMW, VPMW, WPMW, XPMW),
I have seen the following codes in the desktop A64, mobile A64, A64 FX, ....
MPMW, RPAW, RPMW, SPCW, SPMW, TPAW, TPMW, UPAW, UPMW, VPMW, WPMW, XPMW, ....
So these familiar 4-letter codes are still being used in the A64 CPU.


According to the AMD document on "Codes and carrier options".
http://www.amd.com/us-en/assets/content_type/DownloadableAssets/01codes-packoptions.pdf
For the 4-letter codes:
First letter represents the day of week manufactured.
- R for first day of week
- S for second day of week
...
- X for last day of week
- M for combined lots.
Second letter is for assembly location.
Third letter is for wafer lot sequence of the day, A - Z, with M for combined lots.
Fourth letter, nothing or W which designates combining wafer lots is prohibited.

Since Venice/San Diego, instead of just R, S, T, U, V, W, X for day of week, A, B, C, D, E, F, G are seen.


E.g. From a 939 Winchester, CFFBD 0431 UPAW
CFFBD - stepping code
0431 - week of 31, year 2004
UPAW - as above


Regarding to the 5-letter codes, I have seen

AAAPC (CH)
AAARC
AAASC FX-53
AAAXC FX-55
ABASC
ABBLE (Venice E3) 0504 1.4V
ABBWE (939 single core E6) 05xx 1.35V
ACBWE (Toledo) 0513, 4800+
ACB2E (Toledo) 0536, FX-60
ACBYF (AM2 940 Windsor 1MB L2) 0604, FX-62, ...
ADBHE (Manchester) 0509 (1.30V), 0514, 4200+
ADB4F (AM2 940 Windsor 512KB L2) 0603, 5000+, ...
CAAAC
CAABC (CH)
CAACC (CH)
CAAGC (CH 939) FX-53
CAAJC (CH C0)
CAAMC (CH C0, 940) FX-51
CAAOC (CH C0)
CAAPC (CH)
CAAWC
CAAXC (CH 939) 4000+
CAAZX (CH) 0501 ???
CAA2C (CH 939, NC 939) 0449, 4000+, FX-55 (AS)
CABCE (SanDiego) FX-57, FX-55 (BN)
CABGE (SanDiego) 3700+, 4000+
CABHE (SanDiego) 0505, 3700+, 4000+, FX-55 (BN), (L2 cut SanDiego -> Venice)
CABNE (SanDiego) 0528, Opteron 939 (BN), FX-57 (BN)
CABYE (SanDiego) 0536, Opteron 939 (BN)
CAB1E (SanDiego) 0550, Opteron 939 (BN)
CAB2E (SanDiego) 0540, Opteron 939 (BN), FX-57 (BN)
CACJE (SanDiego) 0546, Opteron 939 (BN)
CBAEC (NC, NC 939)
CBASC (NC, NC 939)
CBAUC (NC, NC 939)
CBAVC (NC)
CBAXC (NC)
CBAZC (NC) 0433
CBBFD (Winchester) 0431
CBBGD (Winchester) 0439
CBBHD (Winchester) 0444
CBBID (Winchester) 0446
CBBLE (Venice E3) 0504
CBB1E (Toledo) 0550, Opteron 939
CCBBE (Toledo) 0610, Opteron 939
CCBKE (Toledo) 0643, Manchester (L2 cut Toledo -> Manchester)
CCBWE (Toledo) 0515, 4400+/4800+, Opteron 939, (L2 cut Toledo -> Manchester)
CCB1E (Toledo) 0550, Opteron 939
CCB2E (Toledo) 0536, FX-60
CDBHE (Manchester) 0526, 3800+
DAAOC (CH)
LBBID (Winchester) 0509
LBBLE (Venice E3) 0515
LBBWE (Venice E6) 0525
YBBLE (Venice E3) 0524

The week listed (in some entries) is the earliest week that the stepping found appeared (may not be the real first week of existence).

It is still too early to derive any correlation (if any) between these letters and overclockability, since there is not much data around yet, ....

If there are more known eventually and it becomes more meaningful, I may add more.

940, 754, 939 CPU models and specifications (post 5)
How to read A64 part number for 940, 754, 939 (post 26)
How to identify the physical core of an A64 (post 86)

A64 overclocking result collection

 

[hitechjb1]

Summary

If everything clocked to the SAME frequencies of CPU, memory bus, HT bus, same memory timing setting, same revision (CG preferred), I think the performance order should be like this:

939 A64 1 MB L2 (unbuffer memory) >=
939 FX 1 MB L2 (registered memory) >=
754 1 MB L2 (ClawHammer) >=
754 512 KB L2 (NewCastle)

939 A64 512 KB L2 ~ 754 A64 1 MB L2 (non-memory intensive programs)
939 A64 512 KB L2 >= 754 A64 1 MB L2 (memory intensive programs)
939 A64 512 KB L2 >= 754 A64 512 KB L2

90 nm 939 512 KB L2 >= 130 nm 939 512 KB L2

>= stands for better or equal performance.


On performance between various CPU variants and platforms

The numbers on performance made here are based on two benchmark results and analysis of some A64 FX, A64 754 ClawHammer (1MB L2), A64 754 NewCastle (512KB L2), Barton, P4's, ...

http://www.ocforums.com/showthread.php?s=&postid=2762781#post2762781
http://www.ocforums.com/showthread.php?s=&postid=2766934#post2766934

For 939 with 1 MB L2 (such as FX), when available, it would perform similar to a 940 FX/Opteron, running at the same frequenies.

For 939, if used with unbuffered, non-ECC memory module, can be slightly faster than a 940 due to less memory overhead.

Both of 939/940 dual channel have 80% more effective memory bandwidth than the 754, hence they are always better in performance, especially for memory intensive programs such as video and image streaming, applications using spatially structured data as in scientific computation, up to 20-80% higher performance (e.g. PCmark02 memory test, Sandra memory bandwidth, Sciencemark Stream, many scientific programs). For video, image streaming, data needs to be refreshed constantly from the main memory (L3) to the on chip L2 via the memory bus as size of data >> L2 size at any given time. Under such situation, the high dual channel memory bandwidth delivers a marked performance advantage.

From a few gaming benchmarks, a A64 FX/939 at 2.4 GHz performs 12-20% better than an A64 754 with 1MB L2 at 2.0 GHz, and 15-29% better than an A64 754 with 512 L2 at 2.0 GHz (memory bus, HT bus same frequencies). Not clear if 754 CPU's were clocked to same speed, what would the performance difference be, as the performance difference can be attributed to both memory bandwidth and CPU raw power, but these numbers put an upper bound on gaming performance of 939 over 754. From looking at another set of game benchmarks with both a 939 (512 KB L2) and a 754 (1 MB L2), both running at same frequencies, they are about tie. So I would conclude that for gaming, if both 939 and 754 have the same L2 cache size, running at the same frequencies of CPU, memory and HT, a 939 performs few % (say 5%) on the average better than a 754 for most games.

If not counting memory intensive programs, the advantage of 939 FX over 754 with 1 MB L2 is few % (say 5%) on the average.

Running at the same frequencies of CPU, memory (and bandwidth), HT, the 1M L2 CPU would perform better than the 512 KB L2 CPU. The performance difference between a 512 KB L2 (such as NewCastle) and a 1 MB L2 (such as ClawHammer) would average around 5%, and between 2-10%+ depending on the application.

As seen from many non-memory intensive programs, a 939 with smaller 512 KB L2 roughly tie with a 754 with larger 1 MB L2, within a few % point either way.


Comparing 754 to Barton

Roughly speaking, when clocking to the same frequencies of CPU, HT, memory,
the top line FX-53 (which 939 would resemble) is better than a barton by about 24-32% (average over a range of progarms, should look at the detailed breakdown as listed in the links).

An A64 754 with 1 MB L2 would be close to the FX-53 except for memory bandwidth and memory intensive programs. For memory intensive applications, the 939/940 would have an edge on performance over the 754 (with same L2 size and running same frequencies) ranging from 20-80%, as seen from those benchmarks.

An A64 754 with 512 KB L2 would be 2-10% worse than an A64 754 w/ 1 MB L2.
....

A 754 512 KB L2 A64 (2800/3000) + 250 GB motherboard is around $100 more (as of May 04) compared to a Nforce2 + mobile Barton, but in return, one gets a NEW system with the A64 technologies + 15-25% average gain over a Barton (at same frequencies).
After July 2004 price drop for 754 and 939, the $100 difference is only about $50.

A 754 1 MB L2 A64 (3200) + 250 GB motherboard is around $150 more (as of May 04) compared to a Nforce2 + mobile Barton, but in return, one gets a NEW system with the A64 technologies + 20-30% average gain over a Barton (at same frequencies).
After July 2004 price drop for 754 and 939, the $150 difference is only about $100.


Low PR 90 nm 939 Winchester (Sept 2004)

Since Sept 04, 90 nm 939 3000+, 3200+, 3500+ 512 KB L2 become available.
These 90 nm 939 CPU with new revision D0 (with some core enhancements ?) should be cooler and perform better than a 130 nm 754/939 NewCastle, in most cases even a 754 ClawHammer, at same clock frequency.

The test result of this link shows that a 90 nm Winchester (3200+) performed better than a 130 nm 939 (3500+) and in many cases a 130 nm 3400+ with larger 1 MB L2, when clocked at same frequencies, from less than 1% to few % over a range of benchmarks.
http://www.madshrimps.be/?action=getarticle&articID=230
Review of 90 nm 939 from Anandtech
http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2242

With a 939 motherboard, IMO, for new build, the 939 combo should be a better choice than a 754 system with a NewCastle, and even a 754 ClawHammer, especially taking into account for future compatibility, uniformity and memory intensive applications. Pricewise, a 90 nm 939 system is also as cost effective as a then 754 system. A 90 nm 939 3200+ Winchester should be a good choice for a cost effective, high performance, high bandwidth, overclocking A64 system with AGP or PCI-e.


Peformance Analysis of various A64 Platforms

Estimation and importance of 939 platform memory bandwidth (page 19)

Differences between the XP FSB and the A64 buses (separate memory bus and HyperTransport bus) (page 19)

Some remarks on cache latency, cache size, memory latecny and memory bandwidth (for A64's) (page 19)


Additional links:

Memory bandwidth, latency, cache size, ... on A64 gaming performance
http://anandtech.com/cpuchipsets/showdoc.aspx?i=2330&p=1

Benchmarks comparing
939 FX-53, 939 3800+ (512KB L2), 754 3700+ (1MB L2), 939 3500+ (512KB L2), XP 2400+ (256KB L2)
P4 3.4EE (1MB L2, 2MB L3), P4 3.4 Prescott (1MB L2), P4 2.4C (512KB L2)
http://hardocp.com/articleprint.html?article_id=626

Performance scaling of memory, HT and CPU (post 76)

DOOM3 and performance comparison of DC/SC, cache size (post 75)

 

[hitechjb1]

Overclocking setting for various bus frequencies

In A64 system, there is no external FSB per se, there are the CPU, the memory bus, the HT system bus. The HTT is a signal of the CPU, and is NOT a physical bus for data transfer. The HTT is used to set various frequencies.

Some bios, CPUID utilities still call HTT as FSB, so the name HTT is used to set CPU frequency, memory bus frequency, HT bus frequency until everybody agree and use the same terms.

The system bus external to the CPU is the HT (HyperTransport bus). The rest of the subsystems such as video, hard drives (IDE, SATA, RAID), optical drives, networking, serial links, multi-CPU communication (for multi-processor board), ..., are comunicating with the CPU via the HyperTransport bus to the chipset and various bridges down stream.

HT system bus or HyperTransport bus is not the same as HTT (see 3 below).


1. CPU_frequency = HTT x CPU_multiplier

CPU mulitplier is CPU specific, some can be adjusted from below up to the stock multiplier of the CPU.
E.g. A64 754 3200+, stock 2GHz = 200x10, multiplier X10, X9, X8, X7, ....

CPU multiplier

The FX are unlocked. But the A64 754, 939 are unlocked up to the default stock multiplier.
E.g. for a A64 754 3200+, 2 GHz stock frequency, default multiplier is x10, so it is unlocked for x10, x9, x8, ....
This is good since one can still overclocking the HTT which is related to the HT system bus and the memory bus via their respective multipliers.

2. Nforce3 150, 250/GB/Ultra, Via K8T800/Pro chipsets support the AGP bus. Since late 2004, motherboards with chipsets such as Nforce4/Ultra/SLI, K8T890, ATI XPRESS 200 support only PCI-e bus.

2a. AGP_frequency

Can be set independently if motherboard/chipset have PCI/AGP lock such as Nforce3 250 GB.
E.g. AGP can be adjusted to hold constant at 66, 67 MHz independent of the HTT overclocking.

2b. PCI-e_frequency

For motherboard with PCI-e,
Set PCI-e_frequency to 100 MHz.


3. HT_bus_frequency = HTT x HT_multiplier

The specification calls for 800 MHz max for 754, 1000 MHz for 939.

HT_multiplier (also called LDT multiplier), most bios has it ranged between 1X and 5X.
E.g. HTT = 200 MHz, 1X to 5X for HT from 200 MHz to 1000 MHz


4. memory_bus_frequency

memory_bus_frequency = CPU_frequency / CPU_memory_divider
or
memory_bus_frequency = CPU_multiplier x HTT / CPU_memory_divider
where
CPU_memory_divider = ceiling(CPU_multiplier / memory_HTT_ratio)
where memory_HTT_ratio = 1/1, 1/2, 2/3, 3/4, 4/5, 5/6, 7/8, 9/10, ... (availability of somel settings is bios/motherboard dependent)


Using this formula, the CPU_memory_divider can be calculated based on CPU_multiplier and memory_HTT_ratio.

Some common cpu_memory_dividers generated by spreadsheet are listed in this table.

Relationship between CPU_memory_divider and CPU_multiplier, memory_HTT_ratio
How to determine memory bus frequency
From which the memory bus frequency can be determined
(memory_bus_frequency = CPU_multiplier x HTT / CPU_memory_divider).



Some planning/estimation based on the CPU multiplier, memory rating to ensure the CPU_memory_divider is an integer. Otherwise some round down of memory bus speed would occur as the divider would round up to next integer.

E.g,
HTT = 270 MHz
CPU_multiplier = 9
memory_HTT_ratio = 5/6 (= 166/200 bios setting)
So
CPU_frequency = 270 x 9 = 2430 MHz
CPU_memory_divider = ceiling(9/(5/6)) = 11 or from table in link
memory_bus_frequency = 2430 /11 = 221 MHz
HT_bus_frequency = 270 x 3 = 810 MHz (using x4 may be too high)
AGP/PCI is locked (AGP = 66, 67 MHz)

- Additional settings are the RAS/CAS timings of tCAS, tRCD, tRP, tRAS.
- Can run any speed (ASYNC) using the CPU_memory_divider, to match memory module speed.
- Slow or faster memory can be used to get whatever bandwidth allowed in the memory modules.
- ASYNC has minimum impact on memory bandwidth efficiency
- 754 efficiency 95%+, 939 efficiency 86%+

For desirable overclocking control, one should be able to set all these frequencies of CPU, HT, AGP/PCI, memory independently as shown above. Exact bios naming may vary.


Appendix: Memory SYNC vs ASYNC

SYNC: HTT = memory_bus_frequency (or memory_HTT_ratio = 1/1)

Without a higher ratio of memory_HTT_ratio > 1, running SYNC is the easiest way to max out a memory module that can take high bus frequency.
E.g. with PC4000 or above, by running the HTT (HTT) as high as possible, to 250 - 300+ MHz. Many boards are able to support 300 MHz HTT (HTT) stable.

If HT bus is limited to around 1000 MHz,
- when HTT at 250 MHz, HTT_multiplier can be set at x4,
- when HTT is far beyond 250 MHz towards 300 MHz, HTT_multiplier may have to be dropped to x3 so the HT bus is stable under 1000 MHz.
This is a tradeoff between memory bus frequency and the HT bus frequency.

E.g. assume CPU maxed around 2500 MHz, with PC4000-4400 memory
Case 1 (PC4000):
HTT = 250 MHz, CPU_multiplier = 10, CPU_frequency = 2500 MHz
HT_bus_frequency = 1000 MHz (w/ x4 HT_mulltiplier)
CPU_memory_divider = 10/1 = 10
memory_bus_frequency = 250 MHz
Case 2 (~PC4400):
HTT = 280 MHz, CPU_multiplier = 9, CPU_frequency = 2520 MHz
HT_bus_frequency = 840 MHz (w/ x3 HT_mulltiplier)
CPU_memory_divider = 9/1 = 9
memory_bus_frequency = 280 MHz

ASYNC: In general, HTT != memory_bus_frequency.

max memory_bus_frequency = HTT x memory_HTT_ratio
where memory_HTT_ratio = ..., 6/5, 5/4, 4/3, 3/2, 2/1, 1/2, 2/3, 3/4, 4/5, 5/6, ...
In some bios, only a few can be selected. DFI boards have the most setting choices.

With reference to the rated HTT 200 MHz,
- HTT = 200, max memory frequency = 200, memory_HTT_ratio = 1/1
- HTT = 200, max memory frequency = 183, memory_HTT_ratio = 11/12
- HTT = 200, max memory frequency = 180, memory_HTT_ratio = 9/10
- HTT = 200, max memory frequency = 175, memory_HTT_ratio = 7/8
- HTT = 200, max memory frequency = 166, memory_HTT_ratio = 5/6
- HTT = 200, max memory frequency = 160, memory_HTT_ratio = 4/5
- HTT = 200, max memory frequency = 150, memory_HTT_ratio = 3/4
- HTT = 200, max memory frequency = 140, memory_HTT_ratio = 7/10
- HTT = 200, max memory frequency = 133, memory_HTT_ratio = 2/3
- HTT = 200, max memory frequency = 120, memory_HTT_ratio = 3/5
- HTT = 200, max memory frequency = 100, memory_HTT_ratio = 1/2

E.g. assume CPU maxed at around 2500 MHz
Case 1 (CPU x10 max, ~ PC3200):
HTT = 250 MHz, CPU_multiplier = 10, CPU_frequency = 2500 MHz
memory_HTT_ratio = 5/6
CPU_memory_divider = 10/(5/6) = 12
memory_bus_frequency = 208 MHz
Case 2 (CPU x9 max, ~ PC3200):
HTT = 280 MHz, CPU_multiplier = 9, CPU_frequency = 2520 MHz
memory_HTT_ratio = 3/4
CPU_memory_divider = 9/(3/4) = 12
memory_bus_frequency = 210 MHz
Case 3 (CPU x8 max, ~ PC4000):
HTT = 310 MHz, CPU_multiplier = 8, CPU_frequency = 2480 MHz
memory_HTT_ratio = 4/5 (may not be available, may need 3/4 or 5/6)
CPU_memory_divider = 8/(4/5) = 10
memory_bus_frequency = 248 MHz (assuming 4/5 ratio)


Memory modules (for 754 and 939 platforms)

 

[hitechjb1]

Chipset

Oct 2004:
Nforce4 chipsets with PCI-e, SLI features

The chipsets NFORCE3 250 GB and VIA K8T800 PRO are replacing the old chipsets NFORCE3 150 and K8T800.
There is another SiS 755 chipset.

The newer chipsets are much better for
- richer features (RAID 0/1/0+1/JBOD, Gb networking, bios tweak, ...)
(250 GB has 2 native raid channels + 2 on-board raid channel (total 4), native gigabit networking, native firewall, ...)
- higher HT bus bandwidth (both data width and frequency)
- more HT device supports
- stability (bios, driver bug fixes)
- The NFORCE3 250 GB supports both 754 and 939
- Both announced having PCI/AGP lock.
- Most Nforce3 250/250 GB motherboards reported PCI/AGP lock working
**** June 02, 04 Anandtech reported that VIA confirmed that there are PCI/AGP lock problems in some K8T800 Pro motherboards. So check test results for specific motherboards and bios carefully.

IMO, the NFORCE3 250 GB is a better choice than the 150, and will soon be available (May - June 04, I think). For K8T800 Pro, have to see how the PCI/AGP lock problem plays out.

So the current/soon-to-be possibilities are, IMO:
- motherboard with Nforce3 250 GB + 754 CPU (price performance)
- motherboard with Nforce3 Ultra/K8T800 Pro + 939 CPU (high performance, high memory bandwidth system)
Typical Overclocking Systems for 754, 939

As of today (early May 04), both the chipsets K8T800 Pro and 250 GB (and the associated motherboard) seem to be head to head competitive with each other, will need to see more results about how each performs based on individual motherboard implementation, features, performance, OC friendness, ....


Previews

nVidia chipset for A64 939 with PCI-e and motherboards (preview)
http://www.anandtech.com/mb/showdoc.html?i=2075&p=2

ATI RS480 for AMD Socket 939 Athlon 64 with PCI-e (preview)
http://www.anandtech.com/mb/showdoc.html?i=2075&p=4

New Athlon64 Chipset from SiS for 939 with PCI-e (preview)
http://www.anandtech.com/mb/showdoc.html?i=2075&p=8

 

[hitechjb1]

Main difference between Nforce3 250 GB/Ultra, K8T800 Pro and Sis 755/964

This is still evolving as more info are coming daily, ...

- K8T800 Pro (+ VT8237)
.... Two-chip solution for chipset
.... setting to 1000 MHz HT bus w/ DDR (i.e. 2000 MT/s)
.... SATA w/ RAID 0/1/0+1/JBOD
.... K8T800 Pro may still have PCI/AGP lock problem
**** June 02, 04 Anandtech reported that VIA confirmed that there are PCI/AGP lock problems in some K8T800 Pro motherboards. So check test results for specific motherboards and bios carefully. http://www.anandtech.com/chipsets/showdoc.html?i=2069&p=3

- Nforce3 250 GB
.... Single chipset solution
.... 800 MHz HT bus w/ DDR (i.e. 1600 MT/s)
.... latest to 1000 MHz HT bus w/ DDR (i.e. 2000 MT/s)
.... Some motherboard (e.g. MSI K8N Neo support 5X HT_multiplier to 1000 MHz HT bus)
.... native supports for 1 Gb/s ethernet
.... native supports for SATA and PATA w/ RAID 0/1/0+1/JBOD
........ native IDE controller in chipset supporting 4 drives
........ 2 native SATA ports + 2 ext SATA ports (total 4 SATA ports)
........ RAID can span over SATA + PATA (total 8 drives max)
.... chipset w/ built-in networking firewall

- Nforce3 Ultra is for 939 platform. Seems to have same feature as 250 GB. Not clear at this time what is the internal difference (circuit performance, ...) between Nforce3 Ultra and Nforce3 250 GB.

- SiS 755/964
.... Two-chip solution for chipset
.... 800 MHz HT bus w/ DDR (i.e. 1600 MT/s)
.... native supports 2 SATA ports w/ RAID 0, 1, JBOD

- Nforce4 chipset for A64 940/754/939, with supports for
.... PCI-e (up to 32 lanes)
.... MP (up to 8-way)
.... 2 x Gigabit networking
.... 8 SATA, 6 PATA (with CK8-04)
.... SP-10 sound storm 2
....
will arrive towards end of 2004.
Nforce4 chipsets with PCI-e, SLI features


This is an article comparing chipsets for 939 - NVIDIA Nforce3 Ultra, SiS 755FX and VIA K8T800 Pro.
The article puts the Nforce3 Ultra on top in terms of functionality, no winner performance-wise,
AGP/PCI lock of K8T800 Pro may not be reliable enough, ....
http://www.tomshardware.com/motherboard/20040901/index.html


VIA K8T800 Pro Chipset

http://www.via.com.tw/en/k8-series/k8t800pro.jsp

http://www.anandtech.com/chipsets/showdoc.html?i=2046&p=1

http://www6.tomshardware.com/motherboard/20040505/index.html


Nforce3 Chipset 250 GB

http://www.nvidia.com/page/nf3.html

http://www.tweaktown.com/document.php?dType=review&dId=636

http://www.motherboards.org/articlesd/hardware-reviews/1380_1.html

http://www6.tomshardware.com/motherboard/20040420/

http://www.techreport.com/reviews/2004q2/nforce3-250gb/index.x?pg=1

http://www.gamers-depot.com/hardware/motherboards/nf3/n250/001.htm

http://www.anandtech.com/chipsets/showdoc.html?i=2004
http://www.anandtech.com/chipsets/showdoc.html?i=2009


Nforce3 Ultra Chipset for 939

This is for 939 platform. Seems to have same feature as 250 GB. Not clear at this time what is the internal difference between Nforce3 Ultra and Nforce3 250 GB.
http://www.nvidia.com/object/IO_13738.html


SiS 755 Chipset

http://www.anandtech.com/chipsets/showdoc.html?i=1922&p=1


CURRENTLY UNDER CONSTRUCTION

 

[hitechjb1]

A64 Nforce4 939, Nforce3 754, 939 Motherboards
Order of listing does not imply preference or popularity.

A64 Nforce4 939 Motherboards

About the Nforce4 chipset, refer to this link:
Nforce4 chipsets with PCI-e, SLI features

Anandtech Nforce4 Ultra motherboard roundup
motherboards included: Abit Fatal1ty AN8, Biostar NF4UL-A9, Chaintech VNF4-Ultra, DFI LANParty UT nF4 Ultra-D, ECS KN1 Extreme, Epox 9NPA+, Winfast NF4UK8AA (Foxconn)
Gold editor choice:
- Epox 9NPA+
- DFI LANParty UT nF4 Ultra-D
http://www.anandtech.com/mb/showdoc.aspx?i=2465

Anandtech Nforce4 SLI motherboard roundup
motherboards included: Asus A8N-SLI Deluxe, DFI LANParty nF4 SLI-DR, Gigabyte K8NXP-SLI, MSI K8N Neo4/SLI Platinum
Gold editor choice:
- DFI LANParty nF4 SLI-DR
- MSI K8N Neo4/SLI Platinum
http://www.anandtech.com/mb/showdoc.aspx?i=2358


Epox 9NPA+ Ultra and SLI
Epox 9NPA+ Ultra
http://www.epox.com/USA/product.asp?id=EP-9NPAplusULTRA
Epox 9NPA+ SLI
http://www.epox.com/USA/product.asp?id=EP-9NPAplusSLI

http://www.hardocp.com/article.html?art=NzQ4
http://www.xbitlabs.com/articles/mainboards/display/epox-9npa.html

DFI LanParty NForce4
The SLI version has 2 x16 PCI-e slots for dual video cards each running x8 bandwidth. The Ultra version has 2 x16 PCI-e slots for dual video cards running x16 and x2 bandwidth respectively, reported about 10% less video performance than the former SLI configuration, but about $50 less. Vcore can be adjusted to 1.85 V, Vdimm to 4 V.

DFI nForce4: SLI and Ultra for Mad Overclockers
http://www.anandtech.com/mb/showdoc.aspx?i=2337

http://www.xbitlabs.com/news/mainboards/display/20050112132310.html
UT NForce4 SLI-D
http://www.dfi.com.tw/Product/xx_pr....jsp?PRODUCT_ID=3469&CATEGORY_TYPE=LP&SITE=NA

UT Nforce4 Ultra-D
http://www.dfi.com.tw/Product/xx_pr....jsp?PRODUCT_ID=3471&CATEGORY_TYPE=LP&SITE=NA
Dual Xpress Graphics (DXG) mode in DFI UT Nforce4 Ultra-D:
http://www.dfi.com.tw/Press/press_h...&TITLE_ID=4890&LINKED_URL=arch344.jsp&SITE=NA

Nforce4 SLI-DR
http://www.dfi.com.tw/Product/xx_pr....jsp?PRODUCT_ID=3449&CATEGORY_TYPE=LP&SITE=NA

MSI K8N Neo4
The SLI version has 2 x16 PCI-e slots for dual video cards each running x8 bandwidth. The Ultra version has 1 x16, 1 x4 PCI-e slots for dual video cards running x16 and x2 bandwidth respectively.

MSI K8N Neo4 Platinum Ultra (Nforce4 Ultra)
http://www.msicomputer.com/product/p_spec.asp?model=K8N_Neo4_Platinum&class=mb
http://www.msi.com.tw/program/products/mainboard/mbd/pro_mbd_detail.php?UID=637
http://www.hexus.net/content/reviews/review.php?dXJsX3Jldmlld19JRD05NTc=
http://www.tbreak.com/reviews/article.php?cat=mobos&id=339&pagenumber=1

MSI K8N Neo4 Platinum SLI (Nforce4 SLI)
http://www.msicomputer.com/product/p_spec.asp?model=K8N_Neo4_Platinum/SLI&class=mb
http://www.anandtech.com/video/showdoc.aspx?i=2258&p=1
MSI K8N Neo4 Diamond SLI (Nforce4 SLI)
http://www.msi.com.tw/program/products/mainboard/mbd/pro_mbd_detail.php?UID=638
http://www.hardwarezone.com/articles/view.php?cid=6&id=1364
MSI K8N Neo4 Platinum/SLI (for US) and Diamond/SLI are basically the same thing.

Asus A8N-SLI Deluxe Nforce4
http://www.asus.com/products/mb/socket939/a8nsli-d/overview.htm
Extreme Overclocking: http://www.vr-zone.com/?i=1582&s=1
Chipset Vmod: http://www.vr-zone.com/?i=1624&s=1
http://www.amdzone.com/modules.php?...ns&file=index&req=viewarticle&artid=96&page=1
http://www.hothardware.com/viewarticle.cfm?articleid=612&cid=2
http://www.pcper.com/article.php?aid=98

Gigabyte GA-K8NXP-9 (Nforce4 Ultra)
http://www.hardcoreware.net/reviews/review-250-1.htm
http://www.anandtech.com/mb/showdoc.aspx?i=2273
Gigabyte GA-K8NXP-SLI (Nforce4 SLI)
http://www.tbreak.com/reviews/article.php?id=341
http://www.anandtech.com/mb/showdoc.aspx?i=2285


A64 Nforce3 (754, 939) Motherboards
This section includes motherboards for 754, 939 systems based on Nforce3 250, 250 GB, Nforce3 Ultra.

Nforce3 Ultra 939 Motherboards (for A64 939 and A64 FX)

A good Nforce3 Ultra 939 motherboard guide, with known bugs, tweaks and fixes
http://xtremesystems.org/forums/showthread.php?t=46730

Anandtech tested some Nforce3 Ultra and K8T800 Pro 939 boards,
including:
- ABIT AV8 939 (VIA K8T800 PRO)
- ECS KV2 Extreme 939 (VIA K8T800 PRO)
- Gigabyte K8NSNXP-939 (Nforce3 Ultra)
- MSI K8N Neo2 Platinum 939 (Nforce3 Ultra)
- MSI K8T Neo2 Platinum 939 (VIA K8T800 PRO)
- ASUS A8V Deluxe 939 Revision 2 (VIA K8T800 PRO) (revised and added this later as A8V rev 2 apparently fixes AGP/PCI lock problem).
The top three are:
- MSI K8N Neo2 Platinum 939 (Nforce3 Ultra) (1st)
- ABIT AV8 939 (VIA K8T800 PRO) (2nd)
- ASUS A8V Deluxe 939 Revision 2 (VIA K8T800 PRO) (2nd)
http://www.anandtech.com/mb/showdoc.aspx?i=2128&p=1
[Put both 939 Nforce3 Ultra and K8T800 Pro here since they came from same testing source].

DFI LANParty UT nF3 250Gb
Review from Anandtech - DFI LANParty UT nF3-250Gb: Overclocker's Dream
http://www.anandtech.com/mb/showdoc.aspx?i=2198

EPOX 9NDA3+ (Nforce3 Ultra)
http://www.ocworkbench.com/2004/epox/9nda3+/g1.htm

MSI K8N Neo2 (nForce3 Ultra)
http://www.viperlair.com/reviews/cpu_mobo/msi/amd/s939/k8npe/index.shtml
It was tested with FX-53, 2x512MB Corsair 3200LL, ATI 9800XT.
Assuming it was running at rated,
2.4 GHz CPU, 200 MHz HTT/memory bus (could not confirm these numbers in review), it achieved
Sandra memory bandwidth: 6058/6001 (94% efficiency)
3DMark 2001 SE v330: 26851 (640x480x32), 21411 (1024x768x32)
http://www.pcper.com/article.php?aid=57&type=expert&pid=1

Gigabyte K8NSNXP-939 (Nforce3 Ultra)
http://www.giga-byte.com/MotherBoard/Products/Products_GA-K8NSNXP-939.htm
http://www.short-media.com/review.php?r=248&p=1
http://www.linuxelectrons.com/article.php?story=20040603104541941


A64 Nforce3 754 Motherboards

A review comparing Nforce3 250 GB motherboards.
- Gigabyte K8NSNXP (nVidia nForce3-250)
- MSI K8N Neo Platinum (nVidia nForce3-250Gb)
- ASUS K8N-E Deluxe (nVidia 250 GB 754)
- Shuttle AN51R (nVidia 250 GB 754)
It ranked the boards in this order: MSI K8N, ASUS K8N-E, Gigabyte K8NSNXP, Shuttle AN51R.
It did not test the EPOX 8KDA3+/J which is very popular here.
http://www.tweaktown.com/document.php?dType=article&dId=689&dPage=1

A review comparing some K8T800 Pro and Nforce3 250 GB motherboards.
- Abit KV8 PRO (VIA K8T800 PRO)
- Chaintech VNF3-250 (nVidia nForce3-250)
- Epox 8KDA3+ (nVidia nForce3-250Gb)
- Gigabyte K8NSNXP (nVidia nForce3-250)
- MSI K8N Neo Platinum (nVidia nForce3-250Gb)
- nVidia nForc3-250Gb Reference Board
It gave editor choice to the Epox 8KDA3+, 2nd choice to both Chaintech VNF3-250 and MSI K8N Neo Platinum
http://www.anandtech.com/chipsets/showdoc.html?i=2063
Another roundup on
- Asus K8N-E (nF3-250Gb)
- Soltek K8AN2E-GR and (nF3-250Gb)
- DFI LanParty UT (nF3-250Gb)
putting the DFI LanParty UT as Gold Editor's choice.
http://www.anandtech.com/mb/showdoc.aspx?i=2206&p=1
[Put both 754 Nforce3 250GB and K8T800 Pro here since they came from same testing source].


DFI LanParty UT (250 GB 754)
http://www.dfi-street.com/forum/index.php
http://www.dfi.com.tw/Product/xx_pr....jsp?PRODUCT_ID=2840&CATEGORY_TYPE=MB&SITE=NA
http://www.ocworkbench.com/2004/dfi/LP-UT-NF3-250GB/g1.htm
http://www.anandtech.com/mb/showdoc.aspx?i=2198&p=1

EPoX 8KDA3+ (250 GB 754)
- Not come with IEEE 1394, has to be via PCI card ?
- Current HT_multiplier up to 4X
- HTT setting to 350 MHz
- CPU voltage to 1.7 V, VDIMM to 2.8 V, VAGP 1.8 V, Chipset voltage 1.75V
- From pic, only two phase vcore regulator. It comes with MOSFET heat sinks, user to install.
- Has 6 channels (2 native chipset, 4 from Silicon Image) of RAID 0/1/0+1, JBOD (?)
- May require bios update to boot mobile. Bios update may require a regular desktop A64.
http://www.epox.com/USA/product.asp?id=EP-8KDA3plus
http://www.digital-daily.com/motherboard/epox-8kda3+/
http://www.neoseeker.com/resourcelink.html?rlid=80052
http://www.tbreak.com/reviews/article.php?cat=mobos&id=302&pagenumber=10
http://www.hexus.net/content/reviews/review.php?dXJsX3Jldmlld19JRD03NTc=

NOTE:
- Bios settings may vary from version to version.
- The number of phases (2, 3, 4, ...) refers to the number of pulse-width modulation channels used to generate/regulate the Vcore. Usually three-phrase is considered better than two-phase if component quality and rating (such as capacitors, MOSFET, ...) and design parameters (such as switching frequency) are equal.


MSI K8N Neo Platinum (250 GB 754)
- Have 5X HT_multiplier
- HTT setting to 300 MHz
- CPU voltage to 1.81 V, VDIMM to 2.85 V, VAGP to 1.85 V
- Three phrase vcore regulator, heat sink on MOSFET's.
- Has 4 channels (2 native from chipset) of RAID 0/1/0+1, no JBOD (?)
- Current bios can boot mobile 1.4 V and mobile DTR.
MSI K8N Neo Platinum FAQ
http://forum.msi.com.tw/board.php?boardid=28&sid=
http://www.msi.com.tw/program/products/mainboard/mbd/pro_mbd_detail.php?UID=572
http://www.vr-zone.com/?i=826&s=1
http://www.tbreak.com/reviews/article.php?id=290
http://www.hardocp.com/article.html?art=NjA3
http://www.anandtech.com/mb/showdoc.html?i=2036


ASUS K8N-E Deluxe (250 GB 754)
- Three phrase vcore regulator
- Has 6 channels (2 native chipset, 4 from Silicon Image) of RAID 0/1/0+1, JBOD (more RAID features)
http://usa.asus.com/prog/spec.asp?m=K8N-E Deluxe&langs=09


Gigabyte K8NSNXP (250 GB 754)
http://www.pcstats.com/articleview.cfm?articleID=1612


Chaintech VNF3-250 (250 754, without GB)
http://www.ocworkbench.com/2004/chaintech/vnf3-250/vnf3-250-g1.htm


Chaintech ZNF3-250 ZENITH (250 754, without GB)
http://www.chaintechusa.com/tw/eng/product_spec.asp?MPSNo=13&PISNo=266
http://www.neoseeker.com/resourcelink.html?rlid=74533



Related links:
Chipsets (with previews)

Main difference between Nforce3 250 GB, K8T800 Pro and Sis 755

Overclocking setting for various bus frequencies

 

[hitechjb1]

A64 K8T890 Motherboards with PCI-e (939) (preview only)

http://www.abit-usa.com/products/mb/techspec.php?categories=1&model=215


A64 K8T800 Pro Motherboards (754, 939)
This section includes motherboards for 754, 939 systems based on the chipset K8T800 Pro + VT8237.
Order of listing does not imply preference or popularity.

K8T800 Pro may still have PCI/AGP lock problem
**** June 02, 04 Anandtech reported that VIA confirmed that there are PCI/AGP lock problems in some K8T800 Pro motherboards. So check test results for specific motherboards and bios carefully.


K8T800 Pro 939 Motherboards (for A64 and A64 FX)

MSI K8T Neo2 (Via K8T800 Pro + VT8237)
http://www.msicomputer.com/product/p_spec.asp?model=K8T_Neo2-FIR&class=mb

ABIT AV8 (Via K8T800 Pro + VT8237)
http://www.abit-usa.com/products/mb/techspec.php?categories=1&model=175
http://www.abit-usa.com/news/2004/20040617.php
http://www.ocworkbench.com/2004/abit/av8/g1.htm

ASUS A8V Deluxe (Via K8T800 Pro + VT8237)
Revision 2 seems to fix the AGP/PCI lock problem. Do not get the previous revisions.
http://www.asus.com/products/mb/socket939/a8v-d/overview.htm
http://www.motherboards.org/articlesd/motherboard-reviews/1417_1.html
http://www.hothardware.com/viewarticle.cfm?articleid=534&cid=3


A64 K8T800 Pro 754 Motherboards
Order of listing does not imply preference or popularity.

ABIT KV8 Pro (Via K8T800 Pro + VT8237)
Some boards have AGP/PCI lock working, some not. Heard that Rev 1.0 boards not working, Rev 1.1 boards working.
http://www.abit-usa.com/products/mb/techspec.php?categories=1&model=176
http://www.abit-usa.com/news/2004/20040504.php


Related links:
Chipsets (with previews)

Main difference between Nforce3 250 GB, K8T800 Pro and Sis 755

Overclocking setting for various bus frequencies

 

[hitechjb1]

MP motherboards

ABIT WN-2S+: Dual AMD Opteron™, 8GB DDR memory, 2 Gigabit Ethernet, 4-channel SATA II , Dual PCI-E x16 Slots
http://www2.abit.com.tw/page/en/ser...?pMODEL_NAME=WN-2S+&fMTYPE=Workstation+Boards

Thunder K8WEX (S2895) Dual Opteron PCI Express Workstation Motherboard
http://www.amdboard.com/tyan_s2895_opteron_board.html

A Quad Opteron MP motherboard (preview)
http://www.anandtech.com/mb/showdoc.html?i=2075&p=5

nVidia chipset and dual Opteron motherboard (preview)
http://www.anandtech.com/mb/showdoc.html?i=2075&p=2

 

[hitechjb1]

Memory bus, cache and memory bandwidth (for 940, 754, 939)

A64 platform replaces the single system bus (aka FSB) of XP by two SEPARATE buses, namely a memory bus and a HyperTransport HT system bus (connecting to all system devices via the on-chip north bridge).

As a result, the system bandwidith would be two to four times that of an XP system running the same bus frequency. The HT bus specification to 800 MHz w/ DDR for 754, to 1000 MBz w/ DDR for 939, 32-bit, a max bandwidth of 6.4 GB/s. The dual channel 128-bit memory bus (for 939/940) has a max bandwidth of 6.4 GB/s. The single channel 64-bit memory bus (for 754) has a max bandwidth of 3.2 GB/s.

The memory controller for A64 is on the CPU chip. As such and the separate memory bus and system bus, there would be less bus conflict and the effective memory latency would be reduced.

- Theoretically, 939/940 always delivers better peformance than a 754, especially on memory bandwidth and memory intensive applications, if price-performance is not a major consideration factor.

- Both 754 and 939 have L2 cache size 512 KB or 1 MB, 940 has only 1 MB L2. Eventually max L2 size will be increased to 2 MB.

- The total system bandwidth (memory + HT) is about two times that of XP system for 754, four times for 939.

- A major difference between 754 and 939/940 is the memory bus and memory bandwidth.
... memory bus
...... 128-bit based dual channel memory bus for 939/940 (actual 128 + 16-bit ECC, total 144-bit)
...... 64-bit based memory bus for 754 (actual 64 + 8-bit ECC, total 72-bit)
... memory module
...... Non ECC and unbuffered memory modules supported by both 939 and 754
...... 940 requires registered memory modules, and optional ECC memory preferred for mission-critical servers
... Memory bandwidth efficiency of a typical 754 system is 95%+.
... Memory bandwidth efficiency of a typical 939 system is 86%+, recent 939 to 90%+ seen.
... Effective memory bandwidth of 939/940 is estimated to be 80+% higher than that of 754.

Estimation and importance of 939 platform memory bandwidth (page 19)

Some remarks on cache latency, cache size, memory latecny and memory bandwidth (for A64's) (page 19)

Cache and CPU performance

Differences between the XP FSB and the A64 buses (separate memory bus and HyperTransport bus) (page 19)


Memory bus frequency setting, SYNC/ASYNC mode

memory_bus_frequency = CPU_frequency / CPU_memory_divider
or
memory_bus_frequency = CPU_multiplier x HTT / CPU_memory_divider
where
CPU_memory_divider = celing(CPU_multiplier / memory_HTT_ratio)

Using this formula, the CPU_memory_divider can be calculated based on CPU_multiplier and memory_HTT_ratio.
Some common cpu_memory_dividers generated by spreadsheet are listed in this table.
From which the memory bus frequency can be determined
(memory_bus_frequency = ceiling(CPU_multiplier x HTT / CPU_memory_divider)).

In the table, for most motherboards and bios, 1/2 multipliers may not be supported. If that is the case, 1/2 multipliers should be round to the next integer. E.g. 13.5 should become 14.


Ref:
Relationship between CPU_memory_divider and CPU_multiplier, memory_HTT_ratio
How to determine memory bus frequency (post 60)

Some planning/estimation based on the CPU multiplier, memory rating to ensure the CPU_memory_divider is an integer. Otherwise some round down of memory bus speed would occur as the divider would round up to next integer.


Memory SYNC vs ASYNC

- Additional settings are the RAS/CAS timings of tCAS, tRCD, tRP, tRAS.
- Can run any speed (ASYNC) using the CPU_memory_divider, to match memory module speed.
- Slow or faster memory can be used to get whatever bandwidth allowed in the memory modules.
- ASYNC has minimum impact on memory bandwidth efficiency
- 754 efficiency 95%+, 939 efficiency 86%+, recent 939 to 90%+ seen (all under 1T setting assumed).

SYNC: HTT = memory_bus_frequency (or memory_HTT_ratio = 1/1)

Without a higher ratio of memory_HTT_ratio > 1, running SYNC is the easiest way to max out a memory module that can take high bus frequency.
E.g. with PC4000 or above, by running the HTT as high as possible, to 250 - 300+ MHz. Many boards are able to support 300 MHz HTT stable.

If HT bus is limited to around 1000 MHz,
- when HTT at 250 MHz, HTT_multiplier can be set at x4,
- when HTT is far beyond 250 MHz towards 300 MHz, HT_multiplier may have to be dropped to x3 so the HT bus is stable under 1000 MHz.
This is a tradeoff between memory bus frequency and the HT bus frequency.

E.g. assume CPU maxed around 2500 MHz, with PC4000-4400 memory
Case 1 (PC4000):
HTT = 250 MHz, CPU_multiplier = 10, CPU_frequency = 2500 MHz,
memory_bus_frequency = 250 MHz,
HT_bus_frequency = 1000 MHz (w/ x4 HT_mulltiplier)
CPU_memory_divider = 10/1 = 10
Case 2 (~PC4400):
HTT = 280 MHz, CPU_multiplier = 9, CPU_frequency = 2520 MHz,
memory_bus_frequency = 280 MHz,
HT_bus_frequency = 840 MHz (w/ x3 HT_mulltiplier)
CPU_memory_divider = 9/1 = 9

ASYNC: In general, HTT != memory_bus_frequency.

memory_bus_frequency = HTT x memory_HTT_ratio
where memory_HTT_ratio = ..., 6/5, 5/4, 4/3, 3/2, 2/1, 1/2, 2/3, 3/4, 4/5, 5/6, ...
In some bios, only a few can be selected. DFI boards have the most setting choices.

With reference to the rated HTT 200 MHz,
- HTT = 200, max memory frequency = 200, memory_HTT_ratio = 1/1
- HTT = 200, max memory frequency = 183, memory_HTT_ratio = 11/12
- HTT = 200, max memory frequency = 180, memory_HTT_ratio = 9/10
- HTT = 200, max memory frequency = 175, memory_HTT_ratio = 7/8
- HTT = 200, max memory frequency = 166, memory_HTT_ratio = 5/6
- HTT = 200, max memory frequency = 160, memory_HTT_ratio = 4/5
- HTT = 200, max memory frequency = 150, memory_HTT_ratio = 3/4
- HTT = 200, max memory frequency = 140, memory_HTT_ratio = 7/10
- HTT = 200, max memory frequency = 133, memory_HTT_ratio = 2/3
- HTT = 200, max memory frequency = 120, memory_HTT_ratio = 3/5
- HTT = 200, max memory frequency = 100, memory_HTT_ratio = 1/2

E.g. assume CPU maxed at around 2500 MHz
Case 1 (CPU x10 max, ~ PC3200):
HTT = 250 MHz, CPU_multiplier = 10, CPU_frequency = 2500 MHz,
memory_HTT_ratio = 5/6, memory_bus_frequency = 208 MHz,
CPU_memory_divider = 10/(5/6) = 12
Case 2 (CPU x9 max, ~ PC3200):
HTT = 280 MHz, CPU_multiplier = 9, CPU_frequency = 2520 MHz,
memory_HTT_ratio = 3/4, memory_bus_frequency = 210 MHz,
CPU_memory_divider = 9/(3/4) = 12
Case 3 (CPU x8 max, ~ PC4000):
HTT = 310 MHz, CPU_multiplier = 8, CPU_frequency = 2480 MHz,
memory_HTT_ratio = 4/5, memory_bus_frequency = 248 MHz, <-- 4/5 ratio may not be available, may need 3/4 or 5/6
CPU_memory_divider = 8/(4/5) = 10


Memory modules (for 754 and 939 platforms) (post 16)


Tradeoff between memory bus frequency, timing, command rate

There are tradeoffs in overall performance (not just raw memory bandwidth) between low latency 2-2-2-x at 230-250 MHz with 1T (command rate), and higher latency between 2.5-3-3-x and 3-4-4-3-x at 270-300 MHz with 2T (command rate), test out these two cases for specific applications and benchmarks. Modules based on TCCD can run 1T to 280+ MHz at 2.5-3/4-3/4-x.

For 939 system, as the dual channel 128-bit memory bus can already provide 80+% bandwidth over the 754, a tight 2-2-2-x running 240 - 250 MHz may provide sufficient memory bandwidth for the whole system using overclocking memory such as with BH-5 chips (with 3.x V). Still, high memory bus above 250 MHz is preferred.

I've seen 939 system achieving 90%+ memory bandwidth efficiency.
Assuming 90% bandwidth efficiency, with good memory modules for high bus frequency (probably at relaxed timing),
at 250 MHz, the effective bandwidth would be 7200 MB/s,
at 275 MHz, the effective bandwidth would be 7920 MB/s,
at 300 MHz, the effective bandwidth would be 8640 MB/s (doable for 4400 module).

Based on Anandtech 939 run-up test using Mushkin PC3500 Level 2 and OCZ PC 3500 Platinum (both use BH5 DRAM chips), with 2 dimms (512MBx2), at 200 MHz memory bus, 2-2-2-10 1T timing, the Sandra (SP2) memory banwdith test obtained was 6100 MB/s (95.3% efficiency) for MSI K8N neo2, 6000 MB/s (93.8% efficiency) for ABIT AV8 and became 5000 MB/s (78.1% efficiency) for 2T.

If the above result for 939 can be further confirmed, there is a difference of 15% in bandwidth efficiency between 1T and 2T.

For a typical 754 system, the effective memory bandwidth is around 3000-4000 MB/s.


Towards year end of 2005, A64 platforms may move to DDR2 memory module (discussed in other posts in this thread) which would require new motherboards, maybe new CPU socket (not clear as of May 04).



Reviews on memory modules for A64 platforms

 

[hitechjb1]

HyperTransport bus


HyperTransport is for point to point connecting the CPU to peripheral subsystems such as networking, storage, serial links, chip to chip communication, I/O, ....

(HyperTransport bus does exist in nforce2 chipset, linking NB and SB.)

It is based on packet switching with a packet size of multiple of 32 bit (4 byte), with a max packet size of 64 bytes. HyperTransport allows for bi-directional transfer.

Data width can be configured in 2, 4, 8, 16, 32 bit.

Its specification is 200-800 MHz with DDR, hence
max bit rate = 1600 Mb/s (per bit)
For 939, its specification has been increased to 1000 MHz with DDR, hence
max bit rate = 2000 Mb/s (per bit)

Since it is packet switching, the switching rate is usually referred to as number of transfer per sec (T/s). So
the maximum transfer rate = 1600 MT/s
For 939, the maximum transfer rate = 2000 MT/s

At maximum 32-bit transfer,
the max bandwidth for 32-bit = 32 x 800 x 2 / 8 = 6400 MB/s = 6.4 GB/s
(peripheral bandwidth already higher than the XP FSB)
For 939, the max bandwidth for 32-bit = 32 x 1000 x 2 / 8 = 8000 MB/s = 8.0 GB/s

Since transfer is allowed for bi-direction, for 32-bit transfer,
the max throughput for 32-bit = 12.8 GB/s.
For 939, the max throughput for 32-bit = 16.0 GB/s.

Compared this speed with 33 MHz/32-bit PCI which is 133 MB/s, it is 48X. Compared to the 1 GB/s for PCI-express, it is 12X.

The max bandwidth (for peripheral communication) of 6.4 GB/s (8.0 GB/s for 939) is comparable to that offered by the current system bus (FSB) used for both memory, video and peripheral at 200 MHz, the max bandwidth for dual channel quad pump P4 is 6.4 GB/s, and DDR for AMD is 3.2 GB/s.


Differences between the XP FSB and the A64 buses (separate memory bus and HyperTransport bus) (page 19)

Memory bus, cache and memory bandwidth (for 940, 754, 939)

Estimation and importance of 939 platform memory bandwidth (page 19)

This thread discusses the confusion between HTT and HT (hypertransport), their relationship, and clarifies their difference: Is HTT an acronym for Hypertransport bus?


Links to HyperTransport Overview, Whitepapers, Specifications, ...

http://www.hypertransport.org
http://www.hypertransport.org/tech_overview.html

 

[hitechjb1]

Memory modules (for 754 and 939 platforms)

Socket 940 Opteron and FX platforms require the use of registered memory modules, in ECC mode for mission-critical systems.

A nice thing about 754 and 939 CPU's is that they support, for non-mission critical system, existing unbuffered, DDR 400/500 (PC 3200/4000) memory modules that most of us have.
Both 754 and 939 support either ECC or non-ECC memory modules.
754 supports up to 3 dimms. 939 supports up to 4 dimms.

PC3200 or DDR 400 is rated at 200 MHz, max bandwidth = 200 x 2 x 8 = 3200 MB/s
PC3500 or DDR 438 is rated at 219 MHz, max bandwidth = 219 x 2 x 8 = 3500 MB/s
PC3700 or DDR 462 is rated at 231 MHz, max bandwidth = 231 x 2 x 8 = 3700 MB/s
PC4000 or DDR 500 is rated at 250 MHz, max bandwidth = 250 x 2 x 8 = 4000 MB/s
PC4400 or DDR 550 is rated at 275 MHz, max bandwidth = 275 x 2 x 8 = 4400 MB/s

For A64 system, the memory bus frequency can be flexibly set at various (SYNC/ASYNC) ratio to CPU frequency and HTT frequency (see post on overclocking setting) with minimal impact on memory bandwidth efficiency.

Existing DDR400 (PC3200) modules would work fine with A64 754 and 939 system.

If getting new memory modules, memory modules that can run speed of DDR 500, or even DDR 550 are preferred, even with higher tRCD and CAS latencies. Getting 1T to run is more important than tRCD and CAS on bandwidth and overall performance. With the DDR500+ (PC4000+) memory modules, memory bus can be overclocked to the level 250 - 300+ MHz. Memory bus at such 250 - 300+ MHz level delivers significantly higher raw memory bandwidth than that at the low 200 - 230 MHz even with lower latancy of 2-2-2-x by most PC3200 modules. Though some existing programs, drivers may not be able to take full advantage of the high raw bandwidth, have to see whether future programs, OS can benefit such excess bandwidth.

Some low latency modules with DRAM chips such as BH5 (new memory modules with BH5 are hard to get currently (2004)) may be able to overclock to 230-250 MHz with tight timing 2-2-2-x and enough high Vdimm (if choosing to use 3.0 - 3.3 V).

Here lists some memory modules that work well in an A64 system, further research recommended to suit exact system and need.
- Micron 256Mb DRAM chips (-5B C), found in some Buffalo PC3700, OCZ EB, Crucial Ballistix, to 250-275+ MHz (2.x V)
- Hynix chips (BT-D43), found in some Kingston PC3200, to the 250-275+ MHz (2.x V)
- HyperX PC4000 (SamSung TCCC chips) rated 2.6 V to at least 250 MHz
- Samsung TCCD chips (4ns) to 250 MHz level with CAS 2.5
- A-DATA PC4200 hyperram to 250-300 MHz (2.x V), at 1T reported
- OCZ EB series, e.g. PC3700, based on Micron DRAM chip, overclock to high bus frequency (DDR550 speed) with low equivalent latency.
- Crucial Ballistix PC3200, based on Micron DRAM chip, overclock to DDR500 speed with 2.5-2-2-5, at 2.x V
- G. Skill TCCD or PQI Turbo PC3200, based on TCCD DRAM chips, reported doing 2-2/3-2-x 1T to 230 MHz, 2.5/3-3/4-3-x 1T to 250-280+ MHz using 2.x V http://www.anandtech.com/memory/showdoc.aspx?i=2235&p=1
- OCZ PC3200 Platinum Rev. 2, based on TCCD DRAM chips, reported doing 2-2/3-2-x 1T to 230 MHz, 2.5/3-3/4-3-x 1T to 250-280+ MHz using 2.x V
- Corsair 4400C25: Based on Samsung TCCD, rated 2.5-4-4-8 275 MHz 2.75V http://www.anandtech.com/memory/showdoc.aspx?i=2312
- BH-5/UTT based modules (250-260+ MHz 2-2-2-x 1T 3.3 V), e.g. OCZ VX, Twinmos Speed Premium 3200
- Samsung UCCC based modules for 1GB x 2, oc 260-280 MHz 3-4-3/4-8 1T, 2.6-2.7 V, if lucky ~300 MHz 3-4-4-8 1T, 2.7 V
- Micron 5B F based modules for 1 GB x 2, oc 280 MHz 3-3-3-x 1T, 2.6-2.8 V

Testing TCCD based memory modules on a 939 platform:
# Centon PC3200-C2x512PC3200LL
# Corsair TwinX 1024 PC4400C25
# Geil Ultra X-GLX1GB3200DC
# G.Skill PC3200/PC4400 1024-LE
# G.Skill PC3200/PC4400 1024-LD
# Mushkin PC3200 Level II V2
# OCZ PC3200EL Rev2
# PDP PC3200-PDC1G3200+XBLK
# PQI PC3200-PQI3200-1024DPU (2.6v)
http://www.madshrimps.be/?action=getarticle&number=1&artpage=1094&articID=277

Some of these TCCD modules may have higher rated equivalent of PC4000/4400/4800 with good timing, rated 250/275/300 MHz. If price difference is not big, say less than 10%, it is a good idea to get the one with guaranteed PC4400 2.5-3/4-3/4-x 1T rating.

Running at 2.x V would lessen some concerns about high Vdimm on the CPU memory controller interface. A64 SDRAM IO Ring abs. max voltage (VDDIO) is 2.9 V.

Memory frequency and latency tradeoff
How much frequency increase is needed to break-even with low latency
Testing UTT and TCCD memory modules in Winchester and DFI NF4 Ultra-D setup
Some results about comparing memory frequency, memory timing, memory divider


Tradeoff between memory bus frequency, timing, command rate

There are tradeoffs in overall performance (not just raw memory bandwidth) between low latency 2-2-2-x at 230-250 MHz with 1T (command rate), and higher latency between 2.5-3-3-x and 3-4-4-3-x at 270-300 MHz with 2T (command rate), test out these two cases for specific applications and benchmarks. Modules based on TCCD can run 1T to 280+ MHz at 2.5-3/4-3/4-x.

For 939 system, as the dual channel 128-bit memory bus can already provide 80+% bandwidth over the 754, a tight 2-2-2-x running 240 - 250 MHz may provide sufficient memory bandwidth for the whole system using overclocking memory such as with BH-5 chips (with 3.x V). Still, high memory bus above 250 MHz is preferred.

I've seen 939 system achieving 90%+ memory bandwidth efficiency.
Assuming 90% bandwidth efficiency, with good memory modules for high bus frequency (probably at relaxed timing),
at 250 MHz, the effective bandwidth would be 7200 MB/s,
at 275 MHz, the effective bandwidth would be 7920 MB/s,
at 300 MHz, the effective bandwidth would be 8640 MB/s (doable for 4400, 4800 modules).

Based on Anandtech 939 run-up test using Mushkin PC3500 Level 2 and OCZ PC 3500 Platinum (both use BH5 DRAM chips), with 2 dimms (512MBx2), at 200 MHz memory bus, 2-2-2-10 1T timing, the Sandra (SP2) memory banwdith test obtained was 6100 MB/s (95.3% efficiency) for MSI K8N neo2, 6000 MB/s (93.8% efficiency) for ABIT AV8 and became 5000 MB/s (78.1% efficiency) for 2T.

If the above result for 939 can be further confirmed, there is a difference of 15% in bandwidth efficiency between 1T and 2T.

For a typical 754 system, the effective memory bandwidth is around 3000-4000 MB/s.


Memory bus, cache and memory bandwidth (for 940, 754, 939)

Memory bus frequency setting, SYNC/ASYNC mode

About ECC memory (for A64)

 

[hitechjb1]

PCI-express and devices

PCI-express is a new standard for connecting new generation of devices such as video card, RAID card, networking card, ... (to the chipset).

PCI-express connection is based on point to point, two-way serial communication between devices, compared to the shared bus strucutre of existing PCI. Each device has its own delicated communication channel/bandwidth, hence resolving the traditional shared bus contention and conflict.

- PCI-express has higher bandwidth per pin compared to existing PCI.
- PCI-express is software compatible with PCI 2.2 devices.
- PCI-express is modular and expandable in pin count, in the form of 1X, 2X, 4X, 8X, 16X, 32X slots. E.g. 16X slot has 164 pins.
- PCI-express slot provides more power for high power devices such as video card. 75 W for PCI-express compared to 25 W from AGP 8X.

PCI - 32 bit, 33 MHz, 84 pins, max BW = 132 MB/s
AGP 4X - 32 bit, 4 x 66 MHz, 108 pins, max BW = 1.06 GB/s
AGP 8X - 32 bit, 8 x 66 MHz, max BW = 2.1 GB/s
PCI-X - 64 bit, 133 MHz, 150 pins, max BW = 1.06 GB/s

PCI-express 1X, 40 pins, maxBW = 2.5 Gb/s (basic link)
PCI-express 16X, 168 pins, maxBW = 5 GB/s

Most obvious to us is the up-coming PCI-express video cards (X16 slot form factor) from nVidia and ATI. The NV45 processor (NV40 + PCI-e) from nVidia for the 6800 series, NV43 for the 6600 series; the R423 processor (R420 + PCI-e) for high end ATI X800, the RV410 for mid-range PCI-e ATI X700.

Further, Nvidia is coming up with SLI (Scalable Link Interface) to support multiplie GPU's on the PCI-e bus of a single system to enhance graphic performance.
http://www.nvidia.com/page/sli.html


Some introductory articles for PCI-express.
http://developer.intel.com/technology/pciexpress/devnet/docs/WhatisPCIExpress.pdf
http://developer.intel.com/technology/pciexpress/downloads/3rdGenWhitePaper.pdf
http://www.pcstats.com/articleview.cfm?articleID=1087

PCISIG commitee defines PCI, but this site requires membership for accessing data.
http://www.pcisig.com/specifications


Importance of PCI express in an A64 system

In the pre-nforce2 era, the frequency ratio between the system bus (FSB) and PCI is 4:1, 5:1.

In the nforce2 era, the ratio becomes 6:1, 7:1.

In the A64 era, with the system bus (HT bus) going towards 1000 MHz, such discrepancy between the system bus and the PCI is very huge, a BW ratio of 60 to 1.

That is why PCI-e has to come in to level the bandwidth difference.

PCI 32 bit, 33 MHz, max BW = 132 MB/s
HT bus 1000 MHz, max BW = 8 GB/s (one way)
PCI-express 16X, 168 pins, maxBW = 5 GB/s

Devices (if any) that are connected to PCI in an A64 system create performance and bandwidth imbalance between such devices, the system bus, and in turns the CPU and memory.

That is why 250 GB is preferred (even over the 250) as its networking and RAID (part of) are native in the south bridge, not through the PCI bridge.

 

[hitechjb1]

DDR2 and compatibility with socket 939

DDR2 is a recent standard after DDR 400 (400/500). DDR2 memory module provide higher memory bandwidth (higher clock frequency) beyond DDR 500 but not necessary shorter latency. It also operates at lower voltage. DDR2 delivers equivalent DDR 533 and up (to 800 at least ?). Further it consumes less system power per module.

DDR2 and DDR memory modules are different in terms of module pin count, voltage, signal timing, signal termination, chip package, ..., as far as the memory module and CPU/memory controller interface are concerned. Definitely a new motherboard layout will be required, but this may not be sufficient for the DDR2 change over.

As such difference, the socket 939 pin layout may or may not be able to interface with the DDR2 memory module, given the constraint imposed by the layers of motherboard interconnect, signal to noise consideration, ...., and more engineering details that have to be addressed.

AFAIK, I don't think there is a yes or no answer for socket 939 and DDR2 compatibility yet (as of April 04). The DDR2 memory change over wont't happen until at least late 2005. Whether the above answer being yes or no, I think whoever wants a A64 platform could not wait that long until the DDR2 + socket issue is clear one way or the other. I think by the time if there is a DDR2 change over (18 months ?), the motherbarods will be more feature rich, higher HT bandwidth and better HT devices (such as PCI-e) and much more powerful 90 nm CPU, there would probably be another upgrade cycle.


Micron, Samsung, ... have begun manufacturing DDR2 memory modules (as of 2Q04).
This article gives an overview of DDR2 memory modules from Micron.
http://download.micron.com/pdf/pubs/designline/dl3Q03.pdf

A 256 MB, 512 MB DDR2 spec from Samsung (used for understanding DDR2 only)
http://www.samsung.com/Products/Sem...253FG0/ds_ddr2_256mb_f-die_base_udimm_r10.pdf

 

[c627627]

 

[topslop1]

So 939 motherboards and 939 A64's will be available around the start of June? If that's true then maybe I should wait a month before getting an A64 solution. However, will the 939's want DDR2? Or will they work with both things?