NVIDIA Tesla vs. AMD Radeon HD

[jwgolf]

My company is in the process of transitioning into GPGPU computing to accelerate our finite element analysis. We recently purchased an NVIDIA Tesla M2090 and are very pleased with the results so far. We're predominantly interested in the card's capability to perform double precision FLOPS as fast as possible since all we need the card to do is essentially performing matrix solving operations. The Tesla card we have now has 6GB of memory but we're not sure we need quite that much. We could probably get away with a 3GB capacity for now. On that note, what would the benefit be for going with, say, the Tesla C2050 as opposed to the much, much cheaper AMD Radeon HD 7970? In terms of FLOPS, the Radeon is almost twice as fast as Tesla C2050. What am I missing here?

For more background on what we're doing with the GPUs, see here for a summary of how we're utilizing these cards. And here for an example of the type of solver that is being used.

Thanks!

 

[biscuitasylum]

"Solving a system of linear equations is a common numerical technique applied in many fields including fluid dynamics, thermal analysis, mechanical simulations, and economics."

Economics? You should send some solutions to the White House. lol

 

[renz496]

i'm no expert but i do know that tesla was geared towards pure GPGPU application and 7970 was geared towards gaming application.

 

[jwgolf]

i'm no expert but i do know that tesla was geared towards pure GPGPU application and 7970 was geared towards gaming application.

That's the extent of my understanding as well. Although the numbers would indicate the 7970 is faster, there must be something I'm missing.

 

[janiashvili]

tesla is tesla, don't compare GPGPU to gaming card

 

[Pinhedd]

That's the extent of my understanding as well. Although the numbers would indicate the 7970 is faster, there must be something I'm missing.

The architectures used by AMD for the previous HD releases (6 and prior) were not very friendly when it came to computations which had dependencies in them. The new HD7 cards use a different architecture which works around some of the problems and this has resulted in a much improved number crunching suite.

Here's a nice round up showing some of the differences and comparison with other cards

http://www.tomshardware.com/reviews/radeon-hd-7970-benchmark-tahiti-gcn,3104-14.html

 

[renz496]

for their previous generation flagship the spec indicated it will be vastly superior compared to nvdia's flagship in terms of FLOPS. but in real world most top super computer available still opting for nvidia's tesla solution instead of amd's fire stream. honestly i can't talk about GCN yet since it still new though AMD themselves also saying the new architecture will be better in GPGPU application. time will tell if any card based on GCN will be able to offer excellent alternative to nvidia tesla

 

[jwgolf]

Thanks for all the great info. Do you guys think it's worth a shot at trying one of the new AMD HD cards for GPGPU? Right now our $4k Tesla card peaks out at ~660 GFlops DP. The AMD HD 7970 promises upwards of 1k GFlops DP, even if we got half of the promised performance for our calculations, it would be far cheaper than an equivalent Tesla card that would get around 500 GFlops.

 

[esrever]

The 7970 will be faster but might not be as reliable because it is still just a gaming card. If your software can use the 7970 effectively then it should deliver about 800GFLOPs DP.

If you need reliability and support it would be much better to go with a tesla or even a firepro but for raw performance the 7970 is the most powerful card right now.

 

[yeochins]

They do not compare in the same segments. None of the benchmarks you see directly compare Tesla to the 7970 for the reason that performance wont compare against a card designed specifically for compute-jobs.

The big difference is in data-reliability. Tesla has Error Checking and Correction (ECC), where as typical consumer graphics cards don't ship with ECC because it would substantially LOWER graphical performance. The difference between the two is whether or not you want the RIGHT result as opposed to a FAST result.

 

[ashinms]

If its still not too late to chime in (I'm sure by now that you've already made your purchase), one thing to consider is that Teslas are made to be used en masse in supercomputers. Yeah, sure, you'll see a workstation or two using one due to drivers being optimized for GPGPU, but the real selling point isn't total compute performance per card, but how efficiently it does its work. It doesn't matter if the card is 2x as powerful so long as it sucks up three times the electricity when building a supercomputer. In that scenario, it is a much better option to just buy a large number of less power hungry teslas than a few gaming cards. For your situation, I'd say go with the radeon.

 

[ashinms]

If its still not too late to chime in (I'm sure by now that you've already made your purchase), one thing to consider is that Teslas are made to be used en masse in supercomputers. Yeah, sure, you'll see a workstation or two using one due to drivers being optimized for GPGPU, but the real selling point isn't total compute performance per card, but how efficiently it does its work. It doesn't matter if the card is 2x as powerful so long as it sucks up three times the electricity when building a supercomputer. In that scenario, it is a much better option to just buy a large number of less power hungry teslas than a few gaming cards. For your situation, I'd say go with the radeon.

 

[ashinms]

If its still not too late to chime in (I'm sure by now that you've already made your purchase), one thing to consider is that Teslas are made to be used en masse in supercomputers. Yeah, sure, you'll see a workstation or two using one due to drivers being optimized for GPGPU, but the real selling point isn't total compute performance per card, but how efficiently it does its work. It doesn't matter if the card is 2x as powerful so long as it sucks up three times the electricity when building a supercomputer. In that scenario, it is a much better option to just buy a large number of less power hungry teslas than a few gaming cards. For your situation, I'd say go with the radeon.

 

[the-devo]

That's the extent of my understanding as well. Although the numbers would indicate the 7970 is faster, there must be something I'm missing.

I haven't read the entire thread, but I know I had this question in mind too except with the Tesla C2075 vs GTX590.
ALTHOUGH THE ARCHITECTURE IS PRACTICALLY THE SAME, THE TESLA DRIVER ENABLES THE FULL POTENTIAL OF THE GPU
AND THE DOUBLE PRECISION COMPUTING ABILITY. THIS MEANS YOU COULD RACK UP 4 GTX 680'S IN SLI OR DIFFERENT PC'S
AND STILL NOT BE ABLE TO OUTPERFORM A SINGLE M2090 TESLA CARD!

It is the software and the ass-kissing support they provide that makes the Teslas cost as much as they do.
If you need it, BUY IT! The 7970 will never meet your needs if you've even considered buying a workstation card.

 

[rishiswaz]

DUDE I CAN TYPE WITH CAPSLOCK TOO. And check benchmarks in raw compute 7970 price/performance blows away the new Kepler GPUs and with the new Kepler Tesla cards it is not looking too bright for Nvidia on compute, when the new FirePro cards it will be pretty clear on how the GPGPU market is going now.
http://images.anandtech.com/graphs/graph6025/47487.png
http://media.bestofmicro.com/A/W/342536/original/musemage.png

 

[renz496]

DUDE I CAN TYPE WITH CAPSLOCK TOO. And check benchmarks in raw compute 7970 price/performance blows away the new Kepler GPUs and with the new Kepler Tesla cards it is not looking too bright for Nvidia on compute, when the new FirePro cards it will be pretty clear on how the GPGPU market is going now.
http://images.anandtech.com/graphs/graph6025/47487.png
http://media.bestofmicro.com/A/W/342536/original/musemage.png

it was for GK104. just because the radeon beats GTX680 doesn't mean the entire line up for Kepler gpu will be bad for GPGPU application. with GK104 nvidia main focus is gaming performance because that what matters in Geforce cards. the true GPGPU chip from Kepler architecture will be GK110 which according to news will come out around december this year.

 

[rishiswaz]

Yeah and those are benches with non GPGPU Radeons as well so it is apples to apples. There will be the enhanced FireStream/FirePro cards for GPGPU and HPC applications, those will compete with GK110. Seeing as none of these are released yet we have to base the GK110 performance based on the rest of Kepler because there is not going to be an architecture change between GK104 and GK110, just optimizations for GPGPU. Whereas what GCN in Southern Islands has shown that optimizations will make what they come back with a beast.

Big Edit: There will be no FireStream card but rather dedicated drivers and a special card in the FirePro W series. They have the card with several benches putting it pretty damn high. We will see what GK110 has. My bad have not been in pro card news lol