Nvidia pitches Tesla GPU-as-CPU tech 'personal supercomputer'
This time it's real, apparently
Nvidia has introduced a desktop computer architecture based on its Tesla graphics chip and it's calling the system the first "real" desktop supercomputer.
The typical spec comprises an AMD four-core Phenom processor plus a trio of Tesla C1060 two-slot cards and a Quadro FX card to take the GPU total to four. The Quadro handles the graphcs while the three Teslas combine with Nvidia's CUDA GPU-as-CPU software to crunch numbers.
Look for 4Tflop computing wherever you see this sign
The upshot, the chip maker claimed, are machines the provide 250 times the processing power of a conventional PC.
Yes, and many, many more times the price too. One of Nvidia's system builder partners is US company Colfax, and its base-spec CXT3000 PSC system with Windows XP 64-bit, three Teslas, a Quadro FX 370, a 2.2GHz Phenom 9550, 2GB of 800MHz DDR 2 memory, a Western Digital 160GB 7200rpm 3Gb/s SATA hard drive and a multi-format DVD writer will set you back $6574 (£4397/€5206).
That said, it's not a rig for gamers - well not many of 'em, anyway - but for boffins who need to process highly complex data models. Think Folding@Home running many times more rapidly than it can manage on your own PC CPU's downtime.
Up the CXT3000 PSC's graphics card to a Quadro FX 5800, and the thing's capable of a 4Tflops of floating-point maths.
In addition to Colfax, other specialist workstation suppliers offering Tesla boxes include UK firms Armari, CAD2 and Viglen; in the US, the list includes Amax, Boxx, Microway and Western Scientific. Mainstream suppliers set to offer Tesla systems globally include Asus, Dell and Lenovo.
It is interesting to read the perceptions of price. I seem to remember paying thousands for my first computer which was a pretty good deal for a 20 Mghz 386SX. I also remember reading about the 386DX and the hot new 486 but they were up over $10,000.
For you younger people, have fun and go out and look for old computer advertisements.
Worth pointing out
No need to waste money on this. You can build a computer with four GTX280s that will in fact have better performance. (It's a complete misconception that you need Tesla to use NVIDIA's tech.)
Yes, but will it keep her warm on a cold winter night without dimming the city lights when you turn it on? I would keep her warm for free if she turned me on.
But gamers' rigs can play too
No, the example system discussed in the article isn't a rig for gamers.
But one can get, not 4 Tflops, but 1 Tflop or thereabouts, on systems with a GeForce GTX280, or an ATI 4850 or 4870 video card, which would be found in rigs for gamers.
@@@oh come on..
Most research environments involve collaboration - so your data has to go on the SAN just so everyone can see it. On top of that, sometimes data can be expensive to obtain, so you want to get it on the most resilient hardware ASAP. On top of that, even if you've got a beefy desktop, chances are good that will only do for preliminary work, and eventually your data has got to go on the SAN anyway for cluster access.
I agree with some of the other posts - our researchers complain that we don't have 64+GB blades for them yet and they're stuck with a "paltry" 32...