generally which is faster: nVidia or ATi?

[alpha754293]

Regardless of the card/device specifics, which is usually faster, nVidia or ATi?

[shdbcamping]

At present... NVidea... by a lot in points per dollar spent on equipment.

[alpha754293]

At present... NVidea... by a lot in points per dollar spent on equipment.

So...CUDA is faster than Stream?

[shdbcamping]

So...CUDA is faster than Stream?

Not neccessarily as it depends on the apps they are running. F@H is faster on the current drivers for CUDA than those for ATI/AMD.

[Beberg]

They are by any practical benchmark about the same.

For the small proteins we often use, nVidia is faster. For large proteins they should be about the same again.

Not a very satisfying answer I know. Intel and AMD are just about the same too.

[alpha754293]

Well...I'm wondering because one of the profs that I've worked with is looking into building a 64-processor cluster and I've showed him the CUDA stuff (which he hasn't seen before) and then I started digging up the stuff about ATI Stream and it's like..."ok...if I were to present this to him for his research, which should I recommend?" And that's how this question came about.

He writes his own CFD code in Fortran90 for his high speed air inlet research and he's trying to do it for under $15k and sub-$100/GFLOP BUT I think that neither CUDA NOR Stream supports FORTRAN right now (although I did read last night that you can call the cuBLAS and cuFFT routines from within FORTRAN), but it also mentioned that it has a hard time doing some of the hardware level parallelization which kind of negates the performance advantages when dealing with some of the matrix-matrix multiply operations.

[mhouston]

Soon you will be able use OpenCL on multiple platforms including GPUs, CPUs, and Cell with one code base. So, if you are getting started just now, you might want to hold out a little bit to develop a single code base that will work on multiple devices and vendors.

[shiryunaga]

interesting that just from F@H perceptive, we can see the difference performance between nvidia and ati, we know that RV770 so popular coz the performance they get for gaming, and ati has get good reputation from that, but if we look from F@H perseptive, nvidia still have better choice, maybe coz the PPD CMIIW

Soon you will be able use OpenCL on multiple platforms including GPUs, CPUs, and Cell with one code base. So, if you are getting started just now, you might want to hold out a little bit to develop a single code base that will work on multiple devices and vendors.

i hope that implementation can make something different, oh yeah can ati have new spesific project just like 59xx for nvidia, just to balance IMO

[alpha754293]

well...supposedly the Stream 9270 is faster than the Tesla C1070, but according to what you guys tell me, that Stream is slower than CUDA which thens makes me wonder "how accurate is the marketing crap that these companies put out?"

[Grandpa_01]

Actually I would ask Stanford why and how they assign the PPD value. For some reason in Stanford's judgment the Wu's that ATI work on are not as valuable to them as the ones Nvidia work on. From what I understand both ATI and Nvidia GPU's fold at about the same rate.

[mhouston]

Nvidia is faster with GT200, especially on the smaller proteins. WUs are benchmarked on ATI, but it's been awhile since WUs have been rebenchmarked, especially after the science updates. Originally, PPD was much more consistent on ATI across different proteins. Now there is swing but not quite as amplified as some of the Nvidia PPD swings with different proteins. There are not shenanigans in the PPD assigned, but the distribution of WUs and some of the projects that are Nvidia only can have a large effect on the PPD people see. When comparing a 770 to a GT200, on the really small proteins, the GT200 is is almost 2X a 770 currently. On the large proteins, that gap can narrow into a much smaller delta, approaching <20% differential. We continue to improve the performance of small proteins, but we are trying to avoid implementing 7XX only code paths, thus making 6XX paths difficult to maintain. Right now, ATI and Nvidia take very different algorithm approaches to folding@home.

[alpha754293]

Nvidia is faster with GT200, especially on the smaller proteins. WUs are benchmarked on ATI, but it's been awhile since WUs have been rebenchmarked, especially after the science updates. Originally, PPD was much more consistent on ATI across different proteins. Now there is swing but not quite as amplified as some of the Nvidia PPD swings with different proteins. There are not shenanigans in the PPD assigned, but the distribution of WUs and some of the projects that are Nvidia only can have a large effect on the PPD people see. When comparing a 770 to a GT200, on the really small proteins, the GT200 is is almost 2X a 770 currently. On the large proteins, that gap can narrow into a much smaller delta, approaching <20% differential. We continue to improve the performance of small proteins, but we are trying to avoid implementing 7XX only code paths, thus making 6XX paths difficult to maintain. Right now, ATI and Nvidia take very different algorithm approaches to folding@home.

In your expert and professional opinion (and I know that your thingy says that you're an AMD developer), and someone has about a 2.5 month deadline to get results out the door; which would be a faster approach?

[7im]

Actually I would ask Stanford why and how they assign the PPD value. For some reason in Stanford's judgment the Wu's that ATI work on are not as valuable to them as the ones Nvidia work on. From what I understand both ATI and Nvidia GPU's fold at about the same rate.

Let's not get off topic on to another points debate. Read the GPU FAQ and the Points FAQ for info on how they assign the PPD value. (Hint: It's primarily based on the value of the science that client can produce on the specific benchmark hardware for that client.)