Parhelia Fill rate or CPU limited??

[DegreeC]

I 've been running some limited benchmarks with my new Parhelia and have run into some puzzling results. I've seen repeatedly in this forum and elsewere how CPU bound the performance of the Parhelia is. So I ran my first set of benchmarks with an Athlon Xp 2100 and then re-ran them with an Xp 2400 with the expectation of a noticable performance increase. Instead, I got NO significant increase at all. This screams fill-rate limitation to me. To test this theory I used Powerstrip to overclock the GPU core only. I was only able to get a stable 8 mhz overclock but even that small increase showed up as an increase in the benchmark scores. This is typical of a fill-rate limited card. Here are my results, they were all run with 16X FSAA, Ansio on:

3Dmark 2001SE(default): XP2100-5081, XP2400-5095, XP2400w/8mhz Parhelia oc-5165

Commanche4 demo(1024X768X32): XP2100- 17.46, XP2400-17.35, XP2400w/8mhz Parhelia oc-18.72

Quake3 Demo1(1024X768X32, Max settings): XP2100-85.2, XP2400-84.7, XP2400w/8mhz Parhelia oc-86.2

Quake3 Demo1(1600X1200X32, max settings): XP2100-40.1. XP2400-40.3, XP2400w/8mhz Parhelia oc-40.6

These suprising results made me verify that my XP2400 was setup properly. I used Sandra 2002 and the CPU benchmarks were on target. I also re-ran 3Dmark 2001 without FSAA and Ansio and got a similar pattern of results.

I may not have an adequate spectrum of benchmarks here to see the full picture but these results are very puzzling to me. I love my Parhelia and have become addicted to surround gaming but I seem to have hit the wall as far as squeezing any more performance out of the card.

__________________
Athlon XP 2100/2400
Retail Matrox Parhelia
MSI KT3 ultra
512 MB CorsairXMS PC2700
Maxtor D740X 60 GB
WinXP Pro SP1

[Novdid]

This is already known, it's the low clocks of the chip that has a part in this behaviour.

This really should be in the benchmark section no?

And BTW, those scores look a bit low to me.

[Greebe]

Off hand I can say your 3Dmark scores are low by ~3000 points

[DegreeC]

Off by 3000 points with 16X FSAA and Ansio enabled? The highest scores I've ever seen with 2.8 P4s were around 6000. My 3D mark scores without FSAA and ansio are around 7800-7900. This may be 300-400 points off the average scores, but not 3000.
I have my BIOS settings and memory times set pretty conservatively.

I guess my question can be better stated: Has anyone else noticed a flat benchmark performance curve as CPU power is increased?

[Greebe]

Turn them off (sorry missed that last little tidbit)

[Indiana]

Go look in the benchmark section, there is a thread about benchmarks with the Radeon9700PRO with it's fancy ColorCompression turned off.
At the end of the thread you can see that at least in 3DMark the Parhelia does indeed get bandwidth limited when using FAA in spite of it's 256 bit DDR memory bus - most likely due to the lacking bandwidth-saving techniques.

[WyWyWyWy]

DegreeC, any comment on non-FSAA?

[DegreeC]

Thanks for the point to your 9700Pro post in Benchmarking Indiana. I can see how the memory bandwidth limitation might come into play with FSAA and Ansio enabled, especially at higher resolutions. Unfortunately, I don't have have a complete set of benchmarks without FSAA and Ansio. However, what I do have points to bandwidth NOT being the issue in this case:

3Dmark2001SE(default, FSAA off, Ansio off): XP2100-7602, XP2400-7628, XP2400w/8mhz Parhelia oc-7721

Commanche4 demo (1024X768X32,FSAA off, Ansio off): XP2100-32.5, XP2400-35.3, XP2400w/8mhz Parhelia oc-35.5

The 3Dmark2001 scores are the really puzzling ones; I think I will have to spend some time at Madonion and break down my individual scores to try to see where the bottleneck is. Interestingly,the Commanche4 demo shows what might be considered a reasonable bump up from the 2100 to the 2400.

What I would love to see is a set of benchmarks that fully explores CPU scaling for AMD and Intel systems.

BTW, I posted this thread here rather than in Benchmarking because there seemed to be a lot more activity here. Thanks!

[Roark]

Try doing these tests again using an OpenGL app for benchmarking. Personally I have found that the OpenGL ICD for the Parhelia is extremely CPU bound. of course, the difference between a 2100+ and a 2400+ shouldn't be too extreme anyways.

Also, i would check to see how the card scales from low resolutions to high resolutions as if the card was the limiting factor it should experience a fairly decent drop off for every increase in resolution. I have found the Parhelia to be quite opposite in this fashion...

ie, going from 640x480 in GLQuake w/o FAA and Ansio to 1024x768 w/ FAA and Ansio causes about a 20fps drop on my box... and the FAA/Ansio hit is about 15fps of it...

[leech]

I was getting 5100 on 3dmark with my P3-850mhz (overclocked to 1011mhz) and the parhelia, though it was at 1024x768 with no anisotropic and FAA.

My friend who just got his P4@2.4ghz and it ran at 8900. So I'd guess that something is messed up there.

Leech

[Indiana]

Originally posted by DegreeC

3Dmark2001SE(default, FSAA off, Ansio off): XP2100-7602, XP2400-7628, XP2400w/8mhz Parhelia oc-7721

Those scores do seem strange. Others with the Parhelia have shown that it does in fact scale with the CPU in 3DMark when no FAA or aniso are used, reaching nearly 10.000 points with faster CPUs. Those mostly were Intel systems, but I don't see any reason why this should be different on AMD systems.
You're sure you have a clean install?

(Another point: as long as it's only 3DMark and the actual games do scale, why do you care at all?)

[DegreeC]

The multiple responses here this evening have convinced me that I have to do some more benchmarking with actual games to see if the 3DMark2001 puzzle is unique to that benchmark. I think I need to concentrate on benchmarking without FSAA or ansio and start at lower resolutions to avoid bringing any memory bandwidth issues into play.

The limited or non existent improvements in my benchmark scores when upgrading to the XP2400 may be exposing some inherent bottlenecks in this CPU/motherboard combination. I think I will attempt to unlock my XP2400 and run the FSB at 133/166 to try and coax more performance out of my system.

Thanks to eveyone for thier input; I'll post my results when complete.

[WyWyWyWy]

I got over 600 points going from 1.9g to 2.2g, of course the main memory bandwidth is increased too.

[chhfchhf]

my Parhelia 3dMark2001pro point is 6600-6700
P4 1.7G williamette
845D
DDR266 512MB
Parhelia-512 128MB OEM
driver:1.03
winXP without SP1
not OC
fillrate is 2400MP/s
i consider my CPU can not fit on my Parhelia!

[Roark]

problem is if you introduce too many variables into the equation you will not really be showing of the P is CPU bound or not...

ie, FSB and memory speed... this will make a huge difference with 3d Mark, as will processor speed... this benchmark in particular has been shown to be horribly processor/memory bound...