[Coz]

For any of you guys wondering if it's worth upgrading to a Granny Bay board here are some benchies from mine to help you decide!!

For me, upgrading to GB was well worth it. I didn't get a massive speed improvement from my old i845E board but it's a much nicer board than my old BD7II IMO. For you guys already running DDR400+ on your i845X boards though it may not be worth the money. The point of having the benchmark numbers here is not for some kind of childish contest (naaahh, mine is faster than yours) but to help anyone decide if they want to bother with GB. Personally I'm very pleased with it but I do have some reservations about it's performance. Given the uber-bandwidth I can achieve it is extremely puzzling that the performance isn't higher. I don't know if this is because the Plumas (E7500) workstation memory controller it uses isn't designed for performance but rather for stability maybe? I mean it's not slow but it should be a lot faster IMO. Maybe some BIOS updates will get things moving a little faster but I doubt it.

So here are the numbers you're dying to see, I hope they help you make your minds-up about GB and if you're all going to splash your cash on one. The raw numbers are followed by the screenies. Enjoy friends..........



GB180 Benchmark Results

Asus P4G8X Deluxe, 180MHz FSB (720FSB), Pentium 4 2.26B @ 3.06GHz (1.65v), Kingmax PC2700 DDR SDRAM @ DCDDR360 (CAS2/2/2/5 Turbo, 2.7v).
Leadtek GeForce4 Ti4600. NVIDIA 41.09 Det. Default Speed - 300/650. Overclocked Speed - 315/735.
Windows XP Pro SP1

Benchmark Programs
ScienceMark beta 2.0 - Run standard memory benchmark. DL - http://www.sciencemark.org
3DMark2001SE Build.300 - Run with standard settings. DL - http://www.madonion.com
PCMark2002 - Run with standard settings. DL - http://www.madonion.com
POV-Ray v3.5 - Run built-in benchmark program POVBENCH. DL - http://www.povray.org
Prime95 v22.12.1 - Run built-in benchmark. DL - http://www.mersenne.org
Super PI v1.1 - Calculate 1M places. DL - http://www.pcekspert.com/download/get-32.html
Unreal Tournament 2003 Demo - Run benchmark.exe with Anand's MaxDetail settings applied to all batch files.


Results


ScienceMark

Bandwidth
L1 Cache - 15,662.28 MB/s
L2 Cache - 18,972.11MB/s
Memory - 3777.63MB/s

Latency
L1 Cache - 2 cycles
L2 Cache - 2/10/20/24/27 cycles
Memory - 4/15/56/217/224 cycles


3DMark2001SE

GF Default Speed - 14,619 3DMarks
GF @ 315/735 OC - 15,028 3DMarks
GF @ 315/735 OC - 18,755 3DMarks (640x480x16, Z16, low D3D detail)


PCMark2002

CPU Score - 7528
Mem Score - 8522
HDD Score - 1637


POV-Ray v3.5

Render Averaged 80.96 PPS over 147456 pixels
Total Time: 0 hours 30 minutes 29.0 seconds (1829 seconds)


Prime95

Intel(R) Pentium(R) 4 CPU 2.26GHz
CPU speed: 3060.14 MHz
CPU features: RDTSC, CMOV, PREFETCH, MMX, SSE, SSE2
L1 cache size: 8 KB
L2 cache size: 512 KB
L1 cache line size: 64 bytes
L2 cache line size: 64 bytes
TLBS: 64
Prime95 version 22.12, RdtscTiming=1
Best time for 256K FFT length: 7.835 ms.
Best time for 320K FFT length: 10.400 ms.
Best time for 384K FFT length: 12.573 ms.
Best time for 448K FFT length: 15.107 ms.
Best time for 512K FFT length: 17.059 ms.
Best time for 640K FFT length: 22.111 ms.
Best time for 768K FFT length: 26.859 ms.
Best time for 896K FFT length: 33.161 ms.
Best time for 1024K FFT length: 35.822 ms.
Best time for 1280K FFT length: 50.975 ms.
Best time for 1536K FFT length: 62.751 ms.
Best time for 1792K FFT length: 77.967 ms.


Super PI

1 Million places in 57 seconds


UT2K3 Demo

GF @ Stock
640x480 - Flyby = 233.40, Botmatch = 79.95
800x600 - Flyby = 225.55, Botmatch = 80.07
1024x768- Flyby = 181.05, Botmatch = 78.11

GF @ 315/735
640x480 - Flyby = 1854.17*, Botmatch = 80.26
800x600 - Flyby = 229.90, Botmatch = 80.24
1024x768- Flyby = 191.67, Botmatch = 79.11

* WTF??

Now the screenies....................

ScienceMark v2.0 beta



3DMark 2001SE Build.330
GF4 Stock



GF4 OC

[Coz]

Stupid maximum-number-of-images-per-post rule........


GF4 OC, 640x480x16, Z16, D3D LQ Settings (I wanted a high score )




PCMark2002



POV-Ray v3.5



Prime95 Benchmark

[Coz]

Super PI



Unreal Tournament 2003 Demo (Anand's Max Detail Settings)

GF4 Stock

[Coz]

Unreal Tournament 2003 Demo (Anand's Max Detail Settings)

GF4 OC



Phew!! Sorry for all the multiple posts guys.....

[Moog]

Originally posted by Coz
GF @ 315/735
640x480 - Flyby = 1854.17*, Botmatch = 80.26
800x600 - Flyby = 229.90, Botmatch = 80.24
1024x768- Flyby = 191.67, Botmatch = 79.11

WTF!? indeed.

All in all it looks to be a pretty nice board. Those mem benchies are giving me wood.

I'll assume the POS 9000 is no more?

[Coz]

Originally posted by Moog
I'll assume the POS 9000 is no more?

Hey Moogy-baby, sup buddy?

The BD7II is awaiting trial for crimes against the Nation of Coz. If convicted it could face a sentence of 4 months on top of my cupboard and then death by 50,000 volts.

Yeah, the bandwidth is nice - Sandra memory benchies hit around 4440MB/s using the 180MHz FSB setting. If only all the benchies showed an improvement like that.....

[Terry]

Gotta...resist...gotta resist...you are tempting me...gotta resist...Springdale coming...gotta resist.

[Coz]

Originally posted by Terry
Gotta...resist...gotta resist...you are tempting me...gotta resist...Springdale coming...gotta resist.

With an IT7+DDR453 you'd probably not notice an improvement by getting a GB board. You'd be much better off waiting for Canterwood/Springdale, that's only about a 4 month wait (but it'll seem like 4 years!!).

[larva]

Your summation that GB is not an earth-shattering developement for 845x users with memory above 400MHz is spot on. 845x is smoking fast under these conditions, and with some skill and luck can well approximate the perfromance of GB and 850e.

The reason that the huge bandwidth numbers don't translate into more performance wise is that memory bandwidth is but one aspect of a complex system. A 2 to 1 difference in memory bandwidth does not translate directly into a 2 to 1 system performance difference in nearly all cases. Not only is memory bandwidth but one aspect of the system as a whole, but the architecture of modern cpu's and chipsets is driven by technologies that reduce the impact of memory performance, as we typically have less than we might wish for.

CPU's have L1, and nowdays L2 caches built in for the simple reason that these are amazingly effective at minimizing the impact of the relative slowness of the memory subsytem. 4-8 way set associative caches running syncronously with the cpu core allows hit rates of 90% or greater. Once we cascade two of these caches the effective hit rate approaches 100% more closely than one might suspect. As the data the cpu needs is nearly always already in the cache's, altering the memory subsystem's performance has a limited effect.

More memory performance is always a good thing, and if the differences are drastic enough can have a significant performance ramification. If we compare 845x with the ram running at a mere 266MHz to GB and 850e alternatives, the lack of bandwidth reaches crippling porportions. But given an extra 150-200MHz on the memory speed, 845x's bandwidth approximates the level generated by GB or 850e closely enough that impact on application performance is very minimal. 845x is a very efficient chipset, and gets more out of what it has than any chipset since the mighty 440BX. As such when it doesn't give up a huge amount of bandwidth the effectiveness of the caching along with the general excellence of its design allows it to in most disciplines match its theoretically superior bretheren, and in some isolated instances even outperform them.

Do you have BD7-II numbers at similar cpu clock rate to compare to all the GB figures? They are really required to paint as complete a picture as it takes to make the points you have raised clear. I do like what I see of GB in terms of overlcocking ability and stability, but your point that 845 can be coaxed into comptetive performance with more pricey alternatives is 100% correct. For those that already have a developed 845 that they are happy with, there is little impetus to upgrade to GB. Memory bandwidth only does so much for us, and we can inject 845x with enough of it to keep performance at competitive levels.

[Coz]

Originally posted by larva
Do you have BD7-II numbers at similar cpu clock rate to compare to all the GB figures?

Thanks for your interesting and informative response.

I do have some BD7II vs GB numbers at 166FSB.......

Test Settings.....
Clean WinXP install, no updates or service packs, NVIDIA Ref Drivers 41.09, Intel chipset drivers 4.10.1012, P4G8X BIOS 1003, GF4

Ti4600 @ 300/700. Sandra ver 2003.1.9.31, Quake2 ver3.20, Quake3 original ver, Windows Media Encoder v9.0, PCMark2002,

3DMark2001SE build 330.

BD7II, 166MHz FSB, 2.83GHz, DDR333
P4G8X, 166MHz FSB, 2.83GHz, DCDDR333
P4G8X, 177MHz FSB, 3.01GHz, DCDDR354, CPU vcore 1.65v

Sandra Memory Bandwidth Test

BD7II 166 - 2550MB/s
P4G8X 166 - 4075MB/s
P4G8X 177 - 4375MB/s


3DMark2001SE build. 330, Default Settings

BD7II 166 - 13,649
P4G8X 166 - 14,117
P4G8X 177 - 14,562
P4G8X 177 - 15,006 (GF4 315/735)


3DMark2001SE build. 330, 640x480x32

BD7II 166 - 15,933
P4G8X 166 - 17,078
P4G8X 177 - 18,060


Quake2 v3.20, 800x600x32 (Q2 for a laugh!)

BD7II 166 - 516.1 fps
P4G8X 166 - 563.4 fps
P4G8X 177 - 590.9 fps


Quake3A, Demo1.dm3, original release, Fastest Setting

BD7II 166 - 379.4 fps
P4G8X 166 - 411.1 fps
P4G8X 177 - 451.1 fps


Windows Media Encoder v9.0, 424MB MPEG movie encode

(time in secs - lower is better)
BD7II 166 - 993 secs
P4G8X 166 - 953 secs
P4G8X 177 - 887 secs


PCMark2002 Memory Test

BD7II 166 - 6605
P4G8X 166 - 7758
P4G8X 177 - 8179



As you can see, even with only half the memory bandwidth, the i845E platform is still very close to the performance of GB. Users running the 3:4 ratio at that FSB speed would probably have near-identical scores to GB.

Conclusion - save your money for Canterwood/Springdale guys!!

[larva]

Thanks for the additional figures. I take it your ram won't tolerate the 3:4. My Kingston won't tolerate it at 166 either, but will at 160-162 (depending on the stick). When I tested a 2.4b that had 166fsb in it the numbers were indeed better to back it down to 162 and enable the 3:4 than to run 1:1 at 166. 845e's results are very much influenced by getting on the 3:4 ratio, as it is rather memory bandwidth lmited at 1:1.