Relative performance: ATI -- Nvidia

Moderators: Site Moderators, PandeGroup

Relative performance: ATI -- Nvidia

Postby SKeptical_Thinker » Fri Aug 15, 2008 4:37 pm

It may well be too early to start drawing conclusions from the GPU figures on the stats page, but why should that stop me?

Right now the teraflops and GPU figures on the status page are:
Code: Select all
Brand     TF   GPUS      GF/GPU
ATI      257   2340      109.83
Nvidia   954   8670      110.03


Both brands are averaging 110GF/GPU to within a percent. That leads one to wonder if GF have any relation to science since Nvidia GPUs seem to get over twice as many PPD as ATI GPUs while providing the same computational performance as measured in GF.

Perhaps the issue is the way Stanford compute teraflops for each platform?
Image
SKeptical_Thinker
 
Posts: 254
Joined: Wed Apr 30, 2008 12:02 am

Re: Relative performance: ATI -- Nvidia

Postby 7im » Fri Aug 15, 2008 5:07 pm

SKeptical_Thinker wrote:Perhaps the issue is the way Stanford compute teraflops for each platform?


It seems the opposite may be more accurate. Since points are supposed to be based on the science, maybe the NV points are set too high?

Then again, we've seen Vijay's post stating the FLOPS numbers are deceiving because some clients have to do twice the math to get the same amount of science completed. Maybe ATI clients have to spin the wheels only half as fast to get the same science done, but which makes the FLOPS count look different.

With out more info, no accurate estimations can be made from that chart.
How to provide enough information to get helpful support
Tell me and I forget. Teach me and I remember. Involve me and I learn.
User avatar
7im
 
Posts: 14648
Joined: Thu Nov 29, 2007 5:30 pm
Location: Arizona

Re: Relative performance: ATI -- Nvidia

Postby Foxery » Fri Aug 15, 2008 5:15 pm

Note that all GPU projects are benchmarked on ATI hardware. There is no seperate test done on nVidia hardware to determine point values. Something is getting lost in translation between FLOPS and Points when moving from one architecture to the other.
Core2 Quad/Q9300, Radeon 3850/512MB (WinXP SP2)
User avatar
Foxery
 
Posts: 322
Joined: Mon Mar 03, 2008 4:11 am
Location: Syracuse, NY

Re: Relative performance: ATI -- Nvidia

Postby 7im » Fri Aug 15, 2008 5:32 pm

Foxery wrote:Note that all GPU projects are benchmarked on ATI hardware. There is no seperate test done on nVidia hardware to determine point values. Something is getting lost in translation between FLOPS and Points when moving from one architecture to the other.


I doubt that Pande Group would do something so inept as to set the points for NV work units without actually testing their performance. The translation may be slightly off because of the many different types of NV hardware, but nothing got lost. Look at the difference in the scaling of points between the Intel and AMD CPU benchmarks from a few years ago. It's never going to be perfect.
User avatar
7im
 
Posts: 14648
Joined: Thu Nov 29, 2007 5:30 pm
Location: Arizona

Re: Relative performance: ATI -- Nvidia

Postby SKeptical_Thinker » Sat Aug 16, 2008 12:20 pm

On the face of it, it would appear that the GF/GPU figures support the statements that have been made about optimization. Once the cores and clients have be optimized (I'm talking about both ATI and Nvidia) to the maximum practical extent, the PPD on the two platforms should be similar
SKeptical_Thinker
 
Posts: 254
Joined: Wed Apr 30, 2008 12:02 am

Re: Relative performance: ATI -- Nvidia

Postby DjSoulshot » Sat Aug 16, 2008 6:15 pm

SKeptical_Thinker wrote:On the face of it, it would appear that the GF/GPU figures support the statements that have been made about optimization. Once the cores and clients have be optimized (I'm talking about both ATI and Nvidia) to the maximum practical extent, the PPD on the two platforms should be similar

Yes that really does sound like the correct way of thinking about this - very good to have numbers for all this :)
DjSoulshot
 
Posts: 41
Joined: Sat Aug 09, 2008 11:49 pm

Re: Relative performance: ATI -- Nvidia

Postby yadge » Mon Aug 18, 2008 1:47 pm

So it's basically just because it's optimized more for Nvidia than for ATI?

I'm just curious, because I know my 4870 is much more powerful than an 8800gt, but gets less than half the points per day. Does this mean that the 8800gt is actually doing the work twice as fast as my 4870?
yadge
 
Posts: 6
Joined: Sat Aug 16, 2008 7:18 pm

Re: Relative performance: ATI -- Nvidia

Postby Xilikon » Mon Aug 18, 2008 2:02 pm

yadge wrote:So it's basically just because it's optimized more for Nvidia than for ATI?

I'm just curious, because I know my 4870 is much more powerful than an 8800gt, but gets less than half the points per day. Does this mean that the 8800gt is actually doing the work twice as fast as my 4870?


Not exactly, it's the same WU which can run on either card so we cannot say it's optimised for NVIDIA. What differ is the core used to crunch the unit, in which case NVIDIA is more optimized than ATI.
Image
User avatar
Xilikon
 
Posts: 539
Joined: Sun Dec 02, 2007 2:34 pm

Re: Relative performance: ATI -- Nvidia

Postby yadge » Mon Aug 18, 2008 2:23 pm

Oh ok, that makes sense. I'm aware of the differences in architecture, and I know that the nvidia architecture is kind of more like "raw power"' and ati is more sophisticated but harder to use and only better in certain circumstances.

Ive also heard that currently it only uses 320 of the 800 shaders, because it was just adapted from the 3000 series. Is this true, and if so, will they develop a new system that can use all 800?
yadge
 
Posts: 6
Joined: Sat Aug 16, 2008 7:18 pm

Re: Relative performance: ATI -- Nvidia

Postby Xilikon » Mon Aug 18, 2008 2:32 pm

About using only 320 out of 800, I cannot give a definitive answer but we know that it seems to not use all the shaders, maybe it's not optimized, maybe the protein size is not big enough to tap into all shaders. Only Mike Houston could answer this with certitude.
User avatar
Xilikon
 
Posts: 539
Joined: Sun Dec 02, 2007 2:34 pm

Re: Relative performance: ATI -- Nvidia

Postby Foxery » Mon Aug 18, 2008 2:44 pm

Essentially correct - right now, 4000 series cards perform the same as a 3000 series card at the same clock rate. Future FAH_Cores and larger work units will take advantage of their extra power. In theory, the PPD gap should shrink a bit.
User avatar
Foxery
 
Posts: 322
Joined: Mon Mar 03, 2008 4:11 am
Location: Syracuse, NY

Re: Relative performance: ATI -- Nvidia

Postby Ivoshiee » Mon Aug 18, 2008 3:02 pm

It is not any secret that the ATI code under utilizes the GPU resources. How or when it will get improved is pointless to speculate. Mike Houston is one you can turn to with your questions.
Ivoshiee
Site Moderator
 
Posts: 1286
Joined: Sun Dec 02, 2007 1:05 am
Location: Estonia

Re: Relative performance: ATI -- Nvidia

Postby skymera » Tue Aug 19, 2008 2:13 am

I think ATi is pointless. Just look at AMD.

Clearly they suck at making cards since they need GDDR5 and 800 shaders to even compete with Nvidia.

I wouldn't get your hopes too high about 800 shaders since the 320 used for the 3xxx dont get much PPD.

Just meh 0.02
Machine 1: 8800GT - 5000PPD Average
Machine 1: SMP Client - 1500PPD
Machine 2: 800MHz AMD Sempron - Very low PPD xD
skymera
 
Posts: 59
Joined: Tue Jun 24, 2008 5:28 pm

Re: Relative performance: ATI -- Nvidia

Postby Evil Penguin » Tue Aug 19, 2008 4:10 am

skymera wrote:I think ATi is pointless. Just look at AMD.

Clearly they suck at making cards since they need GDDR5 and 800 shaders to even compete with Nvidia.

I wouldn't get your hopes too high about 800 shaders since the 320 used for the 3xxx dont get much PPD.

Just meh 0.02

What's pointless about AMD/ATi?

They use different techniques to compete with nVidia which by the way were apparently effective.
They used GDDR5 so they wouldn't have to resort to a larger memory interface and increase the GPU's over all size.
ATi's shader architecture is different than that of nVidia's so your count comparison is pointless.

The R670 isn't being utilized to it's fullest either.

Your post is full of ignorance.
User avatar
Evil Penguin
 
Posts: 380
Joined: Sun Apr 13, 2008 5:34 am
Location: Texas, United States

Re: Relative performance: ATI -- Nvidia

Postby Trivolve » Tue Aug 19, 2008 1:40 pm

Maybe we need to start running SMP-like GPU-clients on ATI x| (i.e. 2 GPU clients on 1 GPU core, unless you want to study proteins with 800 atoms)
Team 134888 - Team Trivolve.
User avatar
Trivolve
 
Posts: 149
Joined: Sat Mar 22, 2008 1:08 pm
Location: Singapore

Next

Return to General GPU client issues

Who is online

Users browsing this forum: No registered users and 1 guest

cron