A64 CPUs, chipsets, motherboards

[hitechjb1]

Comparing Sempron (754 256 KB L2, pre-release?) with A64 (754 512 KB L2)

Based on the report and benchmark numbers from
http://www.theinquirer.net/?article=16936
http://www.tweakers.net/nieuws/33100?t=1088594925&,
let's look some benchmark numbers of a Sempron which is a A64 754 with 256 KB L2 compared to some A64 754 with 512 KB L2, running x86-32 bit programs.

CPU compared:
A64 754 3000+, 512 KB L2, at 2.0 GHz
A64 754 2800+, 512 KB L2, at 1.8 GHz
Sempron 754 3100+, 256 KB L2, at 1.8 GHz (pre-release)
From the OPN, the Sempron SDA3100AIP3AX is a 754 lidded OuPGA, 1.4 V, 70C, 256 KB L2, CG rev

HTT: 200 MHz
Motherboard: Nforce3 250
Memory: Kingston DDR 400 256 MB x 2, 8-3-3-3
Video card: Leaktek A310 (GeForce FX 5600)
Windows XP Pro SP1 + DX 9.0b

Numbers are listed in this order: A64 3000+, A64 2800+, Sempron 3100+ (with the A64 2800+ as reference)
3Dmark 2001 SE Pro build 330: 10373, 10200, 9990 --> 1.70%, 0%, -2.06%
PCmark 2004 Pro build 120: 3703, 3450, 3418 --> 7.33%, 0%, -0.93%
Super PI 4M: 234 s, 254 s, 255 s --> 7.87%, 0%, -0.39%
Prime95: 146.715 ms, 153.989 ms, 160.821 ms --> 4.72%, 0%, -4.44%
CPUmark 99: 219, 201, 192 --> 8.96%, 0%, -4.48%

So from these few benchmark runs, the 256 KB L2 A64 Sempron performs worse than, from 0.39% to 4.48%, a 512 KB L2 A64, when both are running under the same conditions of CPU frequency (1.8 GHz) and HTT (200 MHz). This performance difference due to L2 size difference is inline with the average number of 5% that we have been using.

 

[hitechjb1]

The tradeoff between CPU and memory bus frequencies on performance

In many A64 motherboards/systems, memory bus and CPU frequencies are related by one of the following forms:

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

where memory_HTT_ratio is ..., 5/6, 4/5, 3/4, 2/3, 1/2, 1/1, 2/1, 3/2, 4/3, 5/4, 6/5, ....

So in general, the final tuning of CPU and memory would end up with two discrete points (CPU1, memory1) and (CPU2, memory2) around the final, "optimal" overclocking operating point. Question is: which is better (CPU1, memory1) or (CPU2, memory2).

E.g. assuming the max for CPU is 2500 MHz, for memory is 266 MHz,
CPU1 = 2500 MHz, memory1 = 250 MHz (CPU_multiplier = 10, mem_HTT_ratio = 1/1, divider = 10)
CPU2 = 2400 MHz, memory2 = 266 MHz (CPU_multiplier = 9, mem_HTT_ratio = 1/1, divider = 9)

The choice of the two overclocking points (CPU1, memory1) and (CPU2, memory2) for performance depends on specific applications, one may be better for certain programs and the other for some others.

For CPU intensive programs (requiring small cache), CPU1 and memory1 would be better. For memory intensive programs, CPU2 and memory2 should be better.

CPU and memory bus operate and repeat operations at a fixed time interval called cycle time, and so they operate a fixed number of operations per second called frequency.
The unit of time is second, and the unit of frequency is Hz. Hz stands for Hertz, and is the same as cycle per second.

The relationship between frequency and cycle time is
f = 1/T

Today, typical CPU operates around 2.5 - 3.0+ GHz, meaning it repeats 2,500,000,000 - 3,000,000,000+ operations every second.
Typical memory bus and system memory operates around 200 - 300 MHz, about 1/10 times of CPU. (1 MHz = 1,000,000 Hz).

So it is apparent that the faster the frequency (MHz), the more operations the CPU can repeat every second, the more computer instructions it can do per second.
For the same programs with the same amount of instructions to execute, a higher MHz CPU can finish it sooner, hence faster turn around in time.

Same is true for moving memory data, video data, ..., over the memory bus. Higher memory bus frequency would enable programs that require lots of memory access, video subsystem access to finish sooner in time.

Overall_performance = A memory + B CPU + ...

where A and B are some constants, CPU and memory stand for frequency of CPU and memory bus.

For CPU intensive programs where everything can reside in the CPU cache (e.g. small kernel code, inner loop code)
A ~ 0
Overall_performance = B CPU (only CPU frequency is important)

For memory intensive programs such as large matrix computation, video compression in which lots of memory and video access are needed, B is also important, so is memory bus frequency.

In general, A and B are non-zero, so both CPU and memory bus frequency are important. The value of A and B depend on specific applications.

 

[hitechjb1]

Exemplary A64 systems: setups, results and experiences

This is not an exhausted list of all the A64 systems in the universe. Mainly are some typical systems (mostly from members here) for showing typical results and reporting experiences. In case some interesting and important setups are missing, pls let me know. The order is arranged as a chronological stack, FILO. More will be added over time.

Sentential's MSI K8N Neo2 + 939 Winchester
Sen's Ultimate Winchester Thread

Osirus's MSI K8N Neo2 + 939 Winchester
Need help with First time A64 Overclock!

Nuclear_Fuzion's Abit AV8 rev1.1 and 939 FX-53
http://www.ocforums.com/showthread.php?t=318606

Buhammot's Chaintech VNF3-250 and 754 desktop 2800+ CG NewCastle
http://www.ocforums.com/showthread.php?t=318441

Silver's EPOX 8KDA3+ and 754 3000+ CG NewCastle (with H20 and certain vmod)
http://www.ocforums.com/showthread.php?t=312541&highlight=silver

Vod's ABIT AV8 (K8T800 PRO) and 939 3500+ CG 512 KB L2, running Linux
http://www.ocforums.com/showthread.php?t=310755

WA2's Chaintech VNF3-250 and 754 desktop 3400+ C0 ClawHammer in PROMETEIA MACH II
http://www.ocforums.com/showthread.php?t=311577

Michaelkahl's MSI K8N NEO Platinum and 754 desktop 3200+ C0 ClawHammer
http://www.ocforums.com/showthread.php?t=311077

Jess1313's Abit KV8 K8T800 Pro and 754 desktop 3200+ CG ClawHammer
He is building a few different A64 systems, ....
http://www.ocforums.com/showthread.php?t=310646

CandymanCan's EPOX 8KDA3J and 754 desktop 3000+ CG NewCastle
http://www.ocforums.com/showthread.php?t=309671

NiTrO bOiE's MSI K8N Neo Platinum and 754 desktop 3200+ CG NewCastle
http://www.ocforums.com/showthread.php?t=308190

Tedinde's MSI K8N Neo Platinum and 754 desktop/mobile 3000+/3200+ (various C0, CG ClawHammer, NewCastle) testing. Lots of results, experience, extreme cooling (VapoChill PE), ....
http://www.xtremeresources.com/forums/showthread.php?t=32300

Overclocking an A64 754 3200+ (04/26/04)
rated 2.0 GHz 1 MB L2 by 23%, 246x10 MHz, default Vcore 1.5V, stock HS
Motherboard Nforce3 250 GB MSI K8N Neo
Memory setting tested: 200x10 1:1, 270x9 1:1, 300x8 (async)
HT tested: 200x4 (800 MHz) up to HTT at 260-265 MHz
Include also testing an A64 754 3400+ (rated 2.2 GHz 1 MB L2)
http://www.anandtech.com/mb/showdoc.html?i=2036&p=6

D]g[ts's Gigabyte K8N PRO (Nforce3 150) and 754 3200+ CG ClawHammer cooled by Vapochill
http://www.ocforums.com/showthread.php?t=310346

Gautam's Gigabyte GA-K8N Pro (Nforce3 150) and 754 DTR 3200+ CG ClawHammer
Athlon64 3200+ DTR Initial Results


Ref:
Typical Overclocking Systems for 754, 939

Overclocking setting for various bus frequencies

A64 754 1 MB L2 + 250 GB system available

 

[hitechjb1]

place holder

 

[hitechjb1]

place holder for 939

 

[hitechjb1]

Availability of 939 motherboards (as of Jul, 2004)

- MSI K8N Neo2 (nForce3 Ultra)

- Gigabyte GA-K8NSNXP-939 (nForce3 Ultra) ~ $230

- ASUS A8V Deluxe (Via K8T800Pro + VT8237) ~ $ 165

- ABIT AV8 (Via K8T800Pro + VT8237) ~ $ 135


As of 07/31/2004:

Anantech revised the original 939 motherboard runup and included the ASUS A8V deluxe revision 2 which apparently fixes the PCI/AGP lock problem in older revisions.
Anandtech tested some Nforce3 Ultra and K8T800 Pro 939 boards.
http://www.anandtech.com/mb/showdoc.aspx?i=2128&p=1
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 Revision 2 (VIA K8T800 PRO)

The top three are:
- MSI K8N Neo2 Platinum 939 (Nforce3 Ultra) (1st)
- ABIT AV8 939 (VIA K8T800 PRO) (silver)
- ASUS A8V deluxe Revision 2 (VIA K8T800 PRO) (silver)


3500+: ADA3500DEP4AW 1.5V (CG rev, FF0h) <- NewCastle 939, 512 KB L2, 2.2 GHz, x11, ~ $485/$495
End of July 2004, 754/939 price drop, 3500+ (retail box) ~ $350

FX53: ADAFX53DEP5AS 1.5V (CG rev, F7Ah) <- ClawHammer 939, 1 MB L2, 2.4 GHz x12, ~ $800


Links to more info for the motherboards:

A64 Nforce3 Motherboards (754, 939)

A64 K8T800 Pro Motherboards (754, 939)

 

[hitechjb1]

A64 mobiles (How different are A64 desktop and mobiles)

Let's look at the A64 754 3200+ CG ClawHammer which has the models for desktop, mobile DTR, mobile 1.4V.

These are the model and OPN code for the two 3200+:

Desktop A64 754
3200+: ADA3200AEP5AR 1.5V (CG rev, F4Ah) <- ClawHammer, 1 MB L2, 2.0 GHz, x10

Mobile A64 754 DTR
3200+: AMA3200BEX5AR 1.5V (CG rev, F4Ah) <- ClawHammer, 1 MB L2, 2.0 GHz, x10

Mobile A64 754 1.4V
3200+: AMN3200BIX5AR 1.4V (CG rev, F4Ah) <- ClawHammer, 1 MB L2, 2.0 GHz x10

Mobile A64 754 1.2V
2800+: AMD2800BQX4AX 1.2V (CG rev, FC0h) <- NewCastle, 512 KB L2, 1.8 GHz, x9
(There is no mobile 1.2V 3200+ ClawHammer, use 2800+ NewCastle as example.)

All the 754 CG 3200+ ClawHammer variants (desktop and mobiles) have the same default frequency and multiplier, as well as L2 size.


As far as voltage, current, power rating, the AMD tech doc put them differently:
E.g. for max-P state,
754 Desktop CG ClawHammer 3200+/3400+: 1.5 V, 57.8 A, 89 W
754 Mobile DTR CG ClawHammer 3000+/3200+/3400+/3700+: 1.5 V, 52.9 A, 81.5 W
754 Mobile 1.4 V CG ClawHammer 2800+/3000+/3200+: 1.4 V, 42.7 A, 62 W
754 Mobile 1.2 V CG NewCastle 2700+/2800+: 1.2 V, 27.3 A, 35 W
(not implying power rating always uniform across PR rating for each power state)

Further, going more details into low power state and power break down, the mobile DTR and the other mobiles 1.4 V and 1.2 V have different low power states and power dissipation than the desktop 3200+. Among the mobile themselves, the power dissipation ratings are also different.

The mobiles 3200+ DTR, 1.4V, 1.2V dissipate (much) less power than the desktop counterpart per frequency (MHz) and also in absolute terms.
For max-P state,
the ratio for thermal design power = 100% : 91.6% : 69.7% : 39.3%.
the ratio for current = 100% : 91.5% : 73.9% : 47.2%.

Hence indeed the mobile DTR, 1.4V, 1.2V run cooler than the desktop counterpart, question is whether desktop motherboards and bios can handle them properly at some given time.

The above information can be found in this link or the AMD web site.
940, 754, 939 CPU models and specifications


Apart from the obvious voltage, frequency, power rating for the desktop and the various mobile CPU's.

Each type of mobiles has its own voltage, power, low-power states specification per its power up/down sequences. For details, read/study the details in the AMD tech doc for each of them. As such, it requires certain bios and motherboards to handle them, .... As more bios and motherboards mature, I hope/expect more and more bios would be able to handle them.

Actually, based on reading the AMD tech doc (unless there are typo or missing informations),
the low power states of desktop A64 is different from the mobile DTR, mobile 1.4V and mobile 1.2V. The latter three have the same low-power states.
...


Power state of A64 desktop, mobiles (DTR, 1.4V, 1.2V) for 754 and 939

Some mobile rumors:
http://www.theinquirer.net/?article=16966


04/08/31: Added A64 754 DTR 3700+.

 

[hitechjb1]

Due to end of July 04 price drop in 754 and 939 CPU, updated this
Typical Overclocking Systems for 754, 939 (post 2)

Least expensive 754 for general purpose, games
2800+/3000+ 754 CG NewCastle (may be 3000+ ClawHammer) + board ~$250-300
3200+ CG ClawHammer + board ~ $350 (after july CPU price drop)

Least expensive 939 for video encoding/streaming, scientific computation
3500+ 939 + board ~$500 (after july CPU price drop)

 

[hitechjb1]

Sempron

Sempron is a family of AMD low price CPU's (value line) made up of either
- Tbred B core in Socket A or
- A64 core with 64-bit capabilities disabled in Socket 754 (maybe 939 also)

http://anandtech.com/cpuchipsets/showdoc.aspx?i=2139
http://www.tomshardware.com/cpu/20040728/index.html

 

[hitechjb1]

A64 max temperature

In the AMD tech doc "AMD Athlon 64 Processor Power and Thermal Data Sheet (30430)":

The max temperature for the desktop 130 nm 754, 939, FX in 130 nm are specified by T_CASE (case temperature). T_CASE_max = 70 C.

The max temperature for the 90 nm 939 are specified by T_CASE (case temperature). T_CASE_max = 65 C.


The max temperature for the mobile DTR, 1.4 V, 1.2 V are specified by T_DIE (die temperature). T_DIE_max = 95 C.

(Revised 3.41 Oct 2004)
TCASE max is the maximum case temperature specification which is a physical value in degrees Celsius. This
value is programmed into Rev D and later processors. Refer to the appropriate functional data sheet, and the
THERMTRIP Status Register in the BIOS and Kernel Developer’s Guide for AMD Athlon™ 64 and
AMD Opteron™ Processors, order# 26094.

TCONTROL max (maximum control temperature) is a non physical temperature on an arbitrary scale that can be
used for system thermal management policies. TCONTROL max represents the value at which the processor has
reached TCASE max when measuring the thermal diode with a dual sourcing current temperature sensor. Refer to
the appropriate functional data sheet, and the THERMTRIP Status Register in the BIOS and Kernel Developer’s
Guide for AMD Athlon™ 64 and AMD Opteron™ Processors, order# 26094. Temperature is in degrees Celsius
on the TCONTROL scale.

Estimation of max die temperature (lower bound)

As the AMD tech document does not specify directly the max die temperature for the desktop processor, I interpret/extend these specification by adding a temperature different between the max case temperature to estimate the max die temperature.

Example, for 130 nm, assuming rated power is 89 W, and assuming using cooling with thermal resistance of 0.2 C/W (use top end air/water cooling to estimate lower bound),
the die temperature rise over case = 89 x 0.2 = 17.8 C

So assuming an allowance of about 18 C is added for further temperature rise at rated over the specified T_CASE_max, so I would estimate a lower bound for the

max die temperature (estimate) = 70 + 18 = 88 C (130 nm A64)

This number can be used in conjuction with the max die temparature of 95 C of the 130 nm A64 mobile for reference.

Using the same way to estimate for the 90 nm A64, rated power is 67 W,
the die temperature rise over case = 67 x 0.2 = 13.4 C. So

the max die temperature (estimate) = 65 + 13 = 78 C (90 nm A64)

 

[c627627]

The following should be featured more prominently and updated regularly.

Very few people realize just how high the performance difference between XP and 64 is at same frequencies:



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).


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).


A one sentence quick description right after or before the percentage figures, describing how they were obtained, would reinforce their credibility, which is increasingly (wrongly, if I may say) questioned all over the place. Thanks.

 

[hitechjb1]

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.

1. The price has been updated (for comparison) in this post
On performance between various CPU variants and platforms and the thread,
right after the July 04 price drop, if you have looked at the posts in the thread.

2. I suggest don't copy and paste without looking at the original, since it may be out of date and may not reflect the latest situation and the overall picture. If you choose to cut and paste, please include a link to the original post since less confusion would arise. Quoting without a link or source, can be misleading or incomplete.

3. Technical correctness is more important than price which will always change over time, over different sellers. Price is for illustaration only, not for absolute.


The details and analysis for obtaining these numbers are detailed in the posts such as

On performance between various CPU variants and platforms

Difference between a ClawHammer and a NewCastle(post 53)

Without reading these posts and the associated links and just looking at and quoting the numbers, it can be very misleading and controversial.

 

[hitechjb1]

Why native device support from chipset is better

For KT133, KT266, KT333, KT400, and some old Nforce2 motherboards, many devices are supported by the PCI bus. PCI bus has max bandwidth of only 132 MB/s (32 bit x 33 MHz).

Due to the use of legacy motherboards, many high speed devices such as IEEE 1394 (firewire) devices (HD, optical drives, video devices, camcorder, ...) are still running on firewire ports supported by PCI bus.

IEEE 1394 is a high speed serial bus standard that supports data transfer rates of up to 400 Mb/s (in 1394a) and 800 Mb/s (in 1394b). A single 1394 port can connect up to 63 external devices. As such, a firewire port is reaching the PCI bus limit. If there are more than one firewire port and more firewire devices are used, the PCI bandwidth and PCI bus contention would limit the performance of the connected firewire devices.

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.

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

PCI-express and devices (post 17)

 

[wannaoc]

Excellent post. It helped me a ton with determining what to do and expect with my new A64 setup. Well done!

 

[hitechjb1]

DOOM3 and performance comparison of DC/SC, cache size

Data from: When does Doom 3 Need a Fast CPU? (from Anandtech)

A study from Anandtech using DOOM3 for the impact of various CPU and its speed on graphic performance - frame rate.

CPU covered are the
- P4 Northwood, Prescott, EE, Celeron D
- A64 FX 939, 939, 754
- Sempron (both A64, Tbred B core), XP
at various speed

with some emphasis on
- 256 KB L2 vs 512 KB L2 vs 1 MB L2
- dual channel vs single channel
- among P4
- among A64
- Sempron vs Celeron
- Intel vs AMD (all CPU's)

Graphics card used - Nvidia GeForce 6800 Ultra, ATI 9800 Pro


Some remarks:

DC vs SC:
Between a FX-51 939 DC and a 754 3200+ 1 MB L2 SC, both clocked at 2.2 GHz, both memory at DDR 400, the FX-51 DOOM3 frame rate is 2.9% faster.

1 MB L2 vs 512 KB L2:
Between a FX-53 939 DC (1 MB L2) and a 939 3800+ DC (512 KB L2), both clocked at 2.4 GHz, the FX-53 DOOM3 frame rate is 3.6% faster.

 

[hitechjb1]

Chipset performance analysis and performance scaling of memory, HT and CPU

Data from: http://www.spodesabode.com/content/article/a64chipsets

This article entitled "Athlon 64 Chipset Performance Analysis" discusses performance analysis of the following chipsets on some benchmarks and games (Sandra memory bandwidth, Sysmark 2002, 3Dmark03, UT2004, Far Cry).
- Nforce3 150, 250
- SiS 755
- VIA K8T800 Pro

It also shows how the following parameters scale with some benchmark performance (UT2004, Aquamark).
- the memory bus frequency, from DDR200 to DDR400 with 15.4% max difference
(not clear what timing and command rate was used)
- HT bus frequency, from 200 to 1000 MHz with 5.8% max difference
- CPU frequency, from 1200 to 2000 MHz with 36.5% max difference

 

[hitechjb1]

Some notes on how safe is to run Vdimm above max absolute of VDDIO

Here attempts to outline what to look for and how to confirm any link between Vdimm and VDDIO:

The AMD specification for the A64 (754, 939) SDRAM IO Ring abs. max voltage is 2.9 V (VDDIO).

Question is whether the Vdimm voltage applied to memory module would affect the SDRAM IO Ring interface which are represented by a few dozen of so pins of the A64 CPU (the exact pins can be found in the AMD tech doc for 754, 939). E.g. to name a few, D5, D7, D9, D11, D13, D15, ....

If one is skill enough to perform some testing, it would be able to see whether changing the Vdimm, say from 2.5 V to 2.8 V, would affect the voltage on those SDRAM IO Ring pins which should be kept below the abs. max voltage 2.9 V of VDDIO (according to spec without risk, of course one can always take certain risk also).

If the voltage of those SDRAM IO Ring pins mentioned earlier indeed changes with Vdimm, I would suggest not to raise Vdimm above 2.9 V (without taking risk).

If the voltage of those SDRAM IO Ring pins mentioned earlier does not vary with Vdimm, then it is a different story.

Further, if it is an expensive CPU like a 939 3500+, 3800+, FX, I strongly suggest not to experiment with. A $150 CPU may be a good candidate for such experiments.

 

[hitechjb1]

Example to setup frequencies for CPU and memory

For A64, the memory_divider is not set explicitly, it is set internally based on the memory to HTT ratio in the bios.

The memory_divider is calculated internally based on CPU_multiplier and memory_HTT_ratio, both are set in bios.

1. memory_HTT_ratio = 1:1 (bios)
memory_divider = CPU_multiplier

2. memory_HTT ratio = 5:6 (bios)
CPU_multiplier = 9 (bios)
memory_divider = 11 (internal)

3. memory_HTT_ratio = 5:6 (bios)
CPU_multiplier = 10 (bios)
memory_divider = 12 (internal)

4. memory_HTT_ratio = 5:6 (bios)
CPU_multiplier = 11 (bios)
memory_divider = 13.5 (internal)

memory_bus_frequency = CPU_frequency / memory_divider

An approximation for memory bus frequency is
memory_bus_frequency ~ HTT x memory_HTT_ratio


Examples

Stock CPU_multiplier = 10 (e.g. 3200+ CH, 3000+ NC)
Assume CPU can do 2.6 GHz.
Cases with different max memory frequency of 260, 217, 240 MHz.

HTT = 250 MHz - 260 MHz
with x10 will give CPU at 2.5 GHz - 2.6 GHz

Case 1. Fast memory (to 260 MHz)

Set in bios memory_HTT_ratio at 1:1, so memory would run at 250 - 260 MHz.
Internally, the memory is set to run at CPU_frequency / 10 = 250 - 260 MHz.

Case 2. Slow memory (to 217 MHz)

Set in bios memory_HTT_ratio at 5:6, so memory would run at around 208 - 217 MHz.
There is no or minimal penalty running non 1:1 memory_HTT_ratio in A64.
Internally, the memory is set to run at CPU_frequency / 12 = 208.3 - 216.7 MHz (CPU = 2500 - 2600 MHz).
(CPU_multiplier = 10, memory_HTT_ratio = 5:6 gives 12 memory_divider, see table in link below).

Case 3. Medium memory (to 240 MHz)

To fill the gap between 216 and 250 MHz to max out memory at 240 MHz,
the CPU multiplier can be lowered to do more "advanced" setup.

If HTT at 260 MHz, CPU would be 2600 MHz.
At 5:6 memory_HTT_ratio, memory would be at around 217 MHz,
but this would be too low for memory !!!.

How to setup so memory runs at around 240 MHz while CPU is around 2600 MHz?

Lower the CPU multiplier to x9.
Set HTT at 288 MHz, so CPU would run at 288 x 9 = 2592 MHz.

Set memory:HTT ratio at 5:6 (166:200) in bios.
Then memory would run at around 288 x (5/6) = 240 MHz.
Internally, the memory is set to run at CPU_frequency / 11 = 2592 / 11 = 235.6 MHz
(CPU_multiplier = 9, memory_HTT_ratio = 5:6 gives 11 memory_divider, see table in link below).


Relationship between CPU_memory_divider and CPU_multiplier, memory_HTT_ratio

Overclocking setting for various bus frequencies

 

[hitechjb1]

How to choose memory divider and memory_HTT_ratio

In theory, due to the discrete nature of CPU multiplier and memory divider, it is very hard to maximize both CPU and memory at the same time, unless the ratio of the max CPU frequency to max memory frequency is an exact integer or half integer.
So in general, either one is max out while trying to make the other one as close to max as possible.

For A64, the memory_divider is not set explicitly, it is set internally based on the memory_HTT_ratio in the bios.
The memory_divider is calculated internally based on CPU_multiplier and memory_HTT_ratio, the latter two are set in bios.
So first find memory_divider, then determine the memory_HTT_ratio.

Let
max_CPU_frequency be the max CPU frequency that can be obtained (including vmod, cooling)
max_memory_frequency be the max memory frequency that can be obtained (including vdimm mod)
CPU_memory_ratio = max_CPU_frequency / max_memory_frequency

Example 1:
max_CPU_freuency = 2500 MHz
max_memory_frequency = 245 MHz
CPU_memory_ratio = 2500 / 245 = 10.2

So a memory_divider of 10 should be used,
but cannot max out both CPU and memory since 10.2 is not an integer or 1/2 integer.
Memory will be maxed out while CPU will not.

From the table in first link below,
CPU_multiplier = 10,
memory_HTT_rato 1:1 would give memory_divider of 10.

memory_frequency = 245 MHz
CPU_frequency = 245 x 10 = 2450 MHz (memory max out, CPU is not)
HTT = 2450 / 10 = 245 MHz

Example 2:
max_CPU_freuency = 2500 MHz
max_memory_frequency = 230 MHz
CPU_memory_ratio = 2500 / 230 = 10.87

So a memory_divider of 11 should be used,
but cannot max out both CPU and memory since 10.87 is not an integer or 1/2 integer.
CPU will be max out while memory will not.

From the table in first link below,
CPU_multiplier = 9,
memory_HTT_rato 5:6 would give memory_divider of 11.

Set HTT = 2500 / 9 = 277 MHz (CPU maxed out, memory is not)
memory = 2500 / 11 = 227 MHz


For other values of max_CPU_frequency, max_memory_frequency, it can be calculated accordingly.



Relationship between CPU_memory_divider and CPU_multiplier, memory_HTT_ratio

Overclocking setting for various bus frequencies

 

[hitechjb1]

Comments on Integrated Heat Spreader (IHS)

IHS is the metallic lid, which is slightly smaller than the size of the CPU package, covers the much smaller CPU chip (die). The CPU die is bonded to the chip carrier (CPU package). Heat sink is installed to contact the IHS directly. The space between the IHS metallic lid and the CPU chip, is filled with thermal compound so heat from the chip can be passed to the IHS metallic lid and then to the heat sink. IHS has been used in P4 and recently in most A64.

Desktop A64 has IHS. Mobile DTR, 1.4V, 1.2V do not come with IHS.

Thermalright SLK-948U can make direct contact with A64 die without IHS.

People are saying the IHS may add 3-5 C to die temperature, which may correspond to about 1.6% or 40 MHz at 2500 MHz level. This is just my simple estimate for A64, so don't quote it further.

IHS has been generally used so apparently it is not a widespread problem (as people are getting >= 2.5 GHz with 1.6-1.7V on air for A64), but it might and there is just one more potential chance (no statistics to support) for getting a below average CPU due to thermal contact imperfection, in additional to a below average CPU due to the silicon die itself.

The average behaviour of IHS for P4 and A64, is not extraordinarily bad, otherwise we would have known about it over the years from P4.

The point is, there is one more uncertainty in the manufacturing process that a below average overclocking chip (low chance but possible) was packaged with good die but due to IHS thermal contact imperfection.

http://www.ocforums.com/showthread.php?t=322427
http://www.burn-it.dk/index.php?state=artikler_read&artikler_nr=38
http://www.xtremesystems.org/forums/showthread.php?s=&threadid=36888