There is a product in Nvidia’s catalogue that is a graphics card in every respect except that you cannot plug a monitor into it. No DVI, no VGA, nothing. It goes in a workstation or a rack and it computes.
I find that more interesting than anything in the gaming line this year.
The reasoning is not complicated once you look at the silicon. A modern GPU is a few hundred small processors arranged to do the same operation to a great many values at once, because that is what shading pixels is. It turns out that a lot of other work has the same shape. Molecular dynamics has that shape. Seismic processing has that shape. Options pricing has that shape. Anything where you apply identical arithmetic across a large array and the values do not need to talk to each other much.
For that shape of problem a graphics card is not a bit faster than a CPU. It is one or two orders of magnitude faster for the money, and that is a large enough gap that it changes which problems are worth attempting.
People were already doing this before there was a proper way to. The technique was to disguise your data as a texture, express your calculation as a shader, and read the answer back as an image. It worked, it was miserable, and it needed someone who understood graphics programming to solve a problem that had nothing to do with graphics. Nvidia’s toolkit removes the disguise. You write something close enough to C, you get the parallelism, and you do not have to pretend your matrix is a picture.
The distributed folding project moved to GPUs and the throughput numbers are not subtle. A few thousand volunteers with the right cards contribute more than a very large number of volunteers with ordinary processors.
Two things stop me getting carried away.
The first is that this is genuinely difficult to write. Not tedious, difficult. You have to think about how memory is laid out, when threads diverge, and what happens at the boundaries, and if you get any of that wrong you can easily produce something slower than the CPU version you were replacing. The demos are written by people who do this for a living, and demos are always the best case.
The second is that it is one vendor’s toolkit on one vendor’s hardware. AMD has its own approach and it is behind. Writing a serious piece of software against a proprietary interface for parts from a single supplier is a decision you will be living with for years, and everyone who has made that bet before has eventually spent a painful period getting out of it.
My guess is that the fix is a standard that works across both vendors, that it arrives within a couple of years because the customers will insist, and that Nvidia will support it while making sure its own tools remain slightly better.
Meanwhile the thing to notice is the card without a monitor port. A company does not remove the display hardware from a display product unless it has decided the display was never the point.