Nvidia's next-generation graphics cards

[ixfd64]

A lot of people were disappointed with Nvidia's GTX 480 video card, one of the reasons being that it had a much lower FLOPS number than the HD 5870. Of course, the FLOPS number alone does not determine the performance of a graphics card. However, the GeForce 400 series only yields 12.5% double precision, while the Fermi architecture was designed to allow 50%. For the purpose of finding prime numbers, this isn't very useful.

Anyways, Nvidia recently released a rough roadmap that mentions two future architectures: Kepler and Maxwell. More information here: http://www.tomshardware.com/news/fer...att,11339.html

Also, the GeForce 500 series was announced about two weeks ago: http://www.tomshardware.com/news/gpu...gtx,11529.html

Some people say that the GeForce 500 series is just a "tweaked" version of GeForce 400, while others are saying that it will be a brand new design that will offer as many as 768 CUDA cores.

What does everyone make of this?

[msft]

Hi ,ixfd64

Frmky report GTX480 2M FFT 4.43 ms/iter.
TheJudger report Tesla C2050 2M FFT 4.3 ms/iter.

I think LLR performance depend memory bandwidth.

[fivemack]

I expect the graphics cards sold to the gaming market to continue to have double-precision crippled: the cost of getting decent CUDA code written is still enough that people only do it when they have a serious need, and organisations with a serious need are reasonably budget insensitive. AMD has no double-precision in anything but its highest-end graphics cards (the 58xx series), but AMD's attempts to sell something like the Tesla boards has had no perceptible success.

So they'll continue to get faster, but you'll always have to buy a $300 card to get the double precision, even if the $120 card outperforms in most other metrics the last generation's $300 card.

[TheJudger]

Originally Posted by msft

Hi ,ixfd64

Frmky report GTX480 2M FFT 4.43 ms/iter.
TheJudger report Tesla C2050 2M FFT 4.3 ms/iter.

I think LLR performance depend memory bandwidth.

Yepp, memory bandwidth is a problem in many cases, not just LLR.

A simplified calculation:
A GTX 480 has 480 cores (enabled), all running at 1401MHz: It is capable of 168GFLOPS in double precission (DP): 480(cores) * 1401MHz * 2 / 8 = 168.12GFLOPS.
Those FLOPS are multiply-adds. For each independend multiply-add we need 3 inputs and have one output. So each DP float has 64bits / 8bytes. So for each multiply-add we need to read 3x 8byte and write 1x 8byte => 32bytes of bandwidth for a single multiply-add.
168.12GFLOPS * 32bytes = ~5.4TB/sec bandwidth needed. A GTX 480 has 177.4GB/sec on device memory...
A paper from Vasily Volkov mentions that a GTX 480 has ~1.3TB/sec bandwitdh in onchip shared memory. The size of the shared memory is ~1MB in total (which is splitt into smaller pieces of L1-cache and shared memory for each multiprocessor). Register bandwidth seems to be at least 8TB/sec but all registers together are ~2MB.
None of the onchip registers, shared memory and L2-cache is big enough to hold the hole dataset for an LLR so we need really the bandwidth of the device memory.
Of course this is a very simplified reflection. In reality you can hide some traffic to the slow device memory... but device memory bandwidth is a limitation.
So now imaging why a Tesla 20x0 doesn't perform much better than a GTX 480: C2050 has 515GFLOPS DP and only 144GB/sec on device memory...


Oliver

[ET_]

I will repost here an interesting link coming from this thread...

http://www.gpgpgpu.com/gecco2009/6.pdf

Luigi

[msft]

Hi, TheJudger

Originally Posted by TheJudger

Yepp, memory bandwidth is a problem in many cases, not just LLR.

Linpack(TOP500) is gift.
Matrix multiplication is not depend mamory bandwidth.

[Ken_g6]

Originally Posted by ET_

I will repost here an interesting link coming from this thread...

http://www.gpgpgpu.com/gecco2009/6.pdf

Luigi

FYI, that looks like it's doing tests of primes < 2^32! There's no FFT involved there, and the FFT is what makes things much harder for LLR.

[ET_]

Originally Posted by Ken_g6

FYI, that looks like it's doing tests of primes < 2^32! There's no FFT involved there, and the FFT is what makes things much harder for LLR.

You are right, but I'm not wrong!

I was considering uses of GPU for other taks, apart from LLR on a thread dedicated to nVidia GPUs...

Luigi

[ewmayer]

I am preparing to port some Mersenne TF code to a Tesla GPU over the coming months ... basic question about the FP arithmetic coding and optimization: Should one target HLL code to the GPU's FMADD-based ISA, e.g. by refactoring one's C code to provide "hints" to the compiler, or are there more-direct way of doing so, e.g. by way of compiler intrinsics or direct GPU-specific inline ASM?
(I have an SSE2 inline-ASM version of the key factoring modpow loop, but I expect the GPU analog of same is going to look rather different.)

Sorry if this has been answered elsewhere - I'm just at the point where I'm starting to gather the various GPU-specific documentation and figuring out which is likely to be the most relevant to my work.

Thanks,
-Ernst

[delta_t]

Saw this up on AnandTech:
http://www.anandtech.com/show/4008/n...eforce-gtx-580

[Karl M Johnson]

I was very worried if GTX 580 would become sm_21 sh!t, but, luckly, it aint!
Woohoo!