Well, I didn’t pay attention for a bit and missed the release of CUDA 1.0. For those of you who aren’t infected with the geek gene, stop reading now. (-:

 CUDA is a programming environment for the new G8 series nVidia video cards. Unfortunately the beta releases of the code only worked with the $400 plus 8800 series cards. The final release will work with my “cheaper” (read $200+) 8600 series video card.

So, what is CUDA and why should anyone care. Well, CUDA stands for Compute Unified Device Architecture. It’s a bull name to fit an acronym. What it basically allows a person to do is to offload compute cycles to the Graphical Processing Unit or GPU. What’s the big deal with this? Well, my cheap card has 32 shaders in it that I can use as 32 individual compute units. IE… A 32 way (very very primative) processing systems. The 8800 cards have 128 Shaders.

These cards give performance in the range of 384 Billion mathematical operations per second. Yes, 384 Billion. My first supercomputer 7 years ago only did 52 Billion and consisted of 4 cabinets each bigger than a refrigerator.  In other words I can now do the same theoretical work on my PC at home for about $400.  Seriously. All this in less than 7 years, from over a $million dollars to desk side. Amazing.

 Why would anyone care? Well, certain codes run embarrisingly parallel. Things like bioinformatics, high energy physics, crash analysis, etc etc etc.. There are a ton of codes that can run really well under this type of environment. Others would suck big time. But, oh, some will just scream. Imagine, instead of fighting to get time on a big cluster at a DOE lab like Oak Ridge you could just spend a couple of grand and run all the codes on your PC sitting next to your desk.

Very High end PC. $3500. Two of the fastest GeForce 8 Ultra cards to be found and SLI’d together. $700 each. CUDA software, free. Now you have almost 750 GFlops of compute power at your desk to do with as you will for about $5000. Joe Shmoe scientist trying to get things done on a budget should be jumping for Joy. 

That’s not to say my $100 million dollar toys are outdated. Remember, only certain codes will run well on a system like I described. It’ll also take someone who can program the darn thing, which isn’t an easy thing. That, and you can’t build a 250 Teraflop system at home, yet. Maybe in 7 more years you can. Till then, I think my job is still safe.

Advertisements