Acceleware's CTO Ryan Schneider introducing our Oil and Gas solutions and it's impact on seismic imaging and reservoir simulations.

Comments

Hello AXE,

Now that the Nvidia Fermi is out on the market, can you please give us your overview of what you expect in performance gain from this GPGPU dedicated processor?

Thanks

# Posted By david | 4/4/10 6:06 AM

At least on the surface, there are 4 benefits that our customers and prospective customers will see once they are able to buy Fermi-class GPUs, ie: Tesla 2xxx (preferred) or Quadro xxxx:

1) Increased memory bandwidth (we estimate 1.5X) for applications where memory bandwidth is the limiting factor

2) Increased computational capacity (we estimate 2X, given 240 vs. 512 cores), for applications where raw GFLOPs is the limiting factor,

3) Double-precision – the ability to target and accelerate double-precision applications; former hardware generations supported double-precision but not significantly enough to be commercially compelling. Fermi should change that equation.

4) ECC Memory - should alleviate any memory glitches and at least reassure IT managers and end users that their results will always be correct, in a production setting. (As one gets closers to installing tens of thousands of GPUs in a single installation, then the statistics of your compute hardware or memory cells being hit by errant cosmic radiation goes up and then ECC helps one to recover from this… but I digress.) Rumor has it that ECC can be turned off, with potential performance benefit for doing so. Further exploration on our part is necessary.

Long story short, the raw hardware specs will probably give us between 1.5-2X speedup for the applications that are material to our customers. There will be an additional lift for double-precision, GT200 (today) vs. Fermi (soon). Finally, there are a whole bunch of other enhancements (the GigaThread Scheduler, better context switching, etc.) that should also lend some benefit to the kernels that we write. These will need to be explored on a case-by-case basis to see what benefit and how much uplift they provide for different kernels/applications.

Although not explicitly part of your question, another way to look at this is from the alternate technologies landscape. Once Fermi is generally available, this will help to widen the gap that Intel’s Nehalem release ate into. (That gap caused us some pain with our GPU-enabled solutions, because it reduced the price-performance benefit of GPUs.) From our perspective, it isn’t that new releases from the GPU vendors continue to widen the performance gap between GPUs and CPUs, rather they maintain the performance differential, whether that be 3X, 5X, 10X or whatever that may be. In a similar vein of commentary, the coming release of Westmere-EP and Nehalem-EX are also good for a lot of the applications that are materially interesting to our customers, for two reasons: 1) All of Acceleware’s library products support multi-CPU and/or multi-GPU, interchangeably. Increases in performance from any of the hardware vendors in that regard are good for our customers. 2) For many applications, both libraries and our professional services work, the host (CPU) performance can actually be the limiting factor (or one of many bottlenecks) so increases on this side can benefit the overall performance of the whole system.

I hope that helps…

# Posted By Ryan Schneider | 4/6/10 8:43 AM