vetta project

vetta project header image 2
← SIAI Canada Academic Prize for 2008 Kolmogorov, Solomonoff, AIXI etc. questions →

Super computer on a chip

June 24th, 2008 · 6 Comments

In 2006, a computer capable of 3 T FLOPS was enough to get onto the list of the top 500 super computers in the world.

Two years later…

This PC has 3 nVidia GTX 280 graphics cards, costs about $3,000 and is rated at 3 T FLOPS.

Tags: AGI · Computer Power · Hardware · Singularity

6 responses so far ↓

  • 1 Stephen Donnelly // Jun 24, 2008 at 9:15 pm

    The cards are about $NZ1000 each retail, pretty good.

    It will be interesting to see the price/performance of the new Nvidia Tesla machines.

  • 2 Shane Legg // Jun 24, 2008 at 9:25 pm

    So far the Tesla boards have cost a lot more. The have more RAM and apparently use very high quality components… but still, I don’t really understand the price difference.

  • 3 Ivo // Jun 26, 2008 at 7:45 pm

    That was somewhat misleading. I am guessing that those supercomputers from the top 500 list had CPU performance in TFLOP range. I thought you can by a PC capable of 3 TFLOPS for $3000, and that would be great if you compare it with a typical $600 PC cable of few GFLOPS.

    Instead, the 3 TFLOPS number is about graphics card performance, not the CPU performance. That cannot speed up my simulations.

  • 4 Shane Legg // Jun 27, 2008 at 8:11 am

    A modern GPU is basically just a large collection of mini-CPUs that have good floating point performance. Thus, if you have a problem what is numerical intensive and parallel in nature then you can get huge speed ups with a GPU. For example, options pricing in finance, n-body simulations in physics, neural networks in artificial intelligence.

    On the other hand, if your problem can’t be made parallel then most super computers can’t help you much either because they also get their power from using large numbers of CPUs in parallel.

  • 5 Ivo // Jun 27, 2008 at 2:00 pm

    My comment was based on my ignorance. I had no idea there were GPU based PCs. I thought these were just some high end graphic cards used for computationally intensive 3D gaming.

    I am reading more about it now. So, when the prices come down, pretty soon we will all be running TFLOP machines. I did not expect the future to come so soon. Have we (or maybe just I) already entered singularity?

  • 6 Shane Legg // Jun 27, 2008 at 2:40 pm

    Ivo:

    Yeah, they are pretty cool. However, there are a few if’s and but’s to take into account. Firstly, the numbers you usually see are for single precision floating point performance. Real super computers quote their double precision performance. With double precision, graphics cards are quite a bit slower. 10x I heard one person say. Most likely this will significantly improve in further iterations. On the other hand, some applications are fine with single precision anyway…

    Another point is that the quoted performance isn’t on a standard LAPACK benchmark set of problems. It’s more of a best case kind of performance. Thus, while the card might say 1 TFLOP, in applications where you need to multiply matrices etc. you might only get 200 GFLOPS at best.

    Another point is that graphics cards don’t have much memory compared to even relatively small super computers from two years ago. A PC with 4 maxed out Tesla cards has 16 GB of graphics RAM, which is very small by super computer standards.

    Nevertheless, for certain applications you can get a pretty impressive speed up over using the main CPU. Often 10x, sometimes 100x. With all the interest this is generating I expect that over the next couple of years the hardware and software for general computation on GPUs will significantly advance.

Leave a Comment