🔒 Locked

NVIDIA Tesla

Started by ToohrVyk Jun 21, 2007 at 11:47 AM 11 replies 1.8k views

Original Post

Jun 21, 2007 at 11:47 AM

Tesla is NVIDIA's new suite of GPGPU-oriented deskside supercomputers. The current version is a GeForce 8800 GTX without the video outputs—16 multiprocessors running at 1.35GHz, each running SIMD over 32 values at a time. The result is powerful—I manage to generate 3 billion random numbers per second (the fundamental building block in most statistical simulations) with little effort. And the development language (CUDA) is really lovely, even if its current beta version compiler still manages to confuse itself quite often. What do you think? Will we be using the "parallel coprocessor" for graphics in the future, instead of using the GPU for parallel computing as we do now?

Ravuya

Jun 21, 2007 at 11:57 AM

Me wanty!

I bet it howls like a banshee when you get the GPU load up there, though.

ToohrVyk

Jun 21, 2007 at 12:03 PM

Quote:
Original post by Ravuya
I bet it howls like a banshee when you get the GPU load up there, though.

Not really. The power supply screams a lot though, and I've got my headphones with Rammstein and System of a Down on anyway [wink]

The problem is that accessing memory you don't own on the GPU results in a freeze-and-reboot. That, and GDI will kill your GPU-running program after five seconds if the GPU is also used for display, but that problem would disappear for GPGPU-only Tesla cards.

Ravuya

Jun 21, 2007 at 12:07 PM

How much did it run you? I could totally see a use for this baby.

ToohrVyk

Jun 21, 2007 at 12:13 PM

Quote:
Original post by Ravuya
How much did it run you? I could totally see a use for this baby.

Programming that thing is my job, so I paid nothing [smile] The card itself is around $600 (for the GeForce 8800 GTX version I use, the GTS one might be cheaper but it's smaller and slower).

Ravuya

Jun 21, 2007 at 12:20 PM

I know the 8800GTX is about $600, but are you actually using the "Tesla" card (without any video outputs)? Because I read this article which says it'll be like a grand and a half.

ToohrVyk

Jun 21, 2007 at 12:39 PM

I'm using the GTX one, which has the same specs and architecture as the C870, except for having less memory—which isn't a problem for me—and video output. I expect it to be on the exact same level of raw performance as the 8800 GTX, though, but overclocking (and NVCC optimization) may improve it even more.

As mentioned by NVIDIA, Tesla will only be available in August (along with CUDA 1.0, I assume, as only the 0.9 is ready right now).

Ravuya

Jun 21, 2007 at 12:41 PM

That's good to know; we were looking at it and wondering how the hell you got your hands on it. [wink]

The 1.5GB of VRAM is a big motivator, so we'll stick it out for the C870.

jonnyfish

Jun 21, 2007 at 6:06 PM

So how long before we have a full GPU-side LAPACK implementation?

Funkymunky

Jun 21, 2007 at 6:58 PM

Holy crap I want one and I want your job

alnite

Jun 21, 2007 at 10:43 PM

OH NOES, an Admin who double post! I postulate that either Ravuya is excited or too excited. btw I am excited too.

Interesting piece of hardwares, nevertheless. ToohrVyk, what exactly do you do with it? Do you just get a job to play around and find the beast's true power, or are you supposed to make a killer demo that features rendering 1 billion polygons on a 6400x4800 resolution?

ToohrVyk

Jun 22, 2007 at 1:05 AM

Quote:
Original post by alnite
Interesting piece of hardwares, nevertheless. ToohrVyk, what exactly do you do with it? Do you just get a job to play around and find the beast's true power, or are you supposed to make a killer demo that features rendering 1 billion polygons on a 6400x4800 resolution?

My job has two components:

Determine if CUDA is worth it: is it simpler than CG to work with, and does it provide an increase in performance over CG (the answer is yes: a naive LCG random number generator without any hand optimization results in 8.4 billion numbers per second, versus 3.2 billion numbers per second for an equivalent generator in CG). I also have developed a library to compile CUDA code on-the-fly (useful if you want to alter the code that you want to run, instead of precompiling it, as in CG).
Build an application which performs sound Monte-Carlo pricing of options and baskets in GPGPU. For those who don't know, this type of pricing consists in simulating a few million trajectories for the underlying assets (which represents generating about 10¹⁵ gaussian white noise values and adding them up) and then computing the average return of the option or basket over all these trajectories. In theory, it should be 768 times faster than the equivalent CPU program, but memory access considerations will probably reduce this value.

Ravuya

Jun 22, 2007 at 9:00 AM

Quote:
Original post by alnite
OH NOES, an Admin who double post! I postulate that either Ravuya is excited or too excited.

Damn server. I am extremely excited. Take that as you may.

Topic Locked

This topic has been locked by a moderator. New replies are not allowed.