Nvidia's Tesla is desktop supercomputer
SUPERCOMPUTERS are especially trendy this week after SC08, but it seems even regular desktops could soon be getting the super treatment with computer makers like Dell seriously mulling the possibility of a PC built around Nvidia’s Tesla floating point accelerators.
The idea isn’t one that’s just been pulled out of a hat, but previous attempts at desktop supercomputery have all been a bit of a flop (as opposed to FLOPS). But Nvidia seems to think a Tesla desktop might just cut it, and has set about convincing the likes of Dell, Asus, Lenovo, Scan and Boxx to start configuring one up.
The idea seems to be that the systems would be fully functional PCs based on four Tesla C1060 cards with a quad-core CPU and 16 GB of memory bunged in for good measure. The desktop hulk would boast some 960 graphics cores and should be able to pack a theoretical four TFlops of single-precision processing punch.
Of course, that’s not even mentioning the 400 GFlops in double-precision applications and the C1060 GPUs (T10P processor) clocked at 1.33 GHz putting out a performance of around 900 GFlops in a single-unit configuration. All for just $10,000.
Nvidia reckons a super Tesla puter would weigh in at 250 times faster than a normal desktop PC, but that’s not taking into account a super ATI competitor. A computer boasting an Nvidia SLI system or two of ATI’s Radeon HD 4870 X2 cards could theoretically give the Tesla desktop a run for its money. Also, ATI reckons its tow 4870 X2 cards pack some 2.4 TFlops in single-precision, blowing the four Tesla cards out of the water in double-precision apps with 480 GFlops of performance.
But again, there’s more to life than performance and the Tesla desktop could well outdo other graphics cards when it comes to memory availability. With 4GB of GDDR3 800 memory in each CPU, the Tesla monster packs 16GB of graphics memory overall, coupled with a combined memory bandwidth of 408 GB/s. The downside, however, is that all this performance and memory turns the Tesla desktop into a particularly hungry power-guzzling machine, with the four cards and a processor easily eating their way through 750 watts.
Still, no one is kidding themselves that this will be a desktop to have at home. Instead, the machine will probably come in most useful for universities and firms for which a $10,000 small sized supercomputer doesn’t seem too steep. µ
L'Inq
TG Daily
Share this:
Business Intelligence Software
Customer Relationship Management (CRM)
Enterprise Accounting Software
Enterprise Resource Planning (ERP)
Enterprise Systems Management
Many if not most scientific PC environments and projects are Linux-based rather than Microsoft/Windows-based.

Two big factors that would influence potential supercomputer customers away from an ATI-based solution are the quality of ATI's Linux drivers (i.e. much worse than nVidia's) and also the lack of a developed API like CUDA.


AMD 4870X2: 230 GFLOP dual precision. And it costs $500, not $5000 or $50,000.

RichW, AMD already has stated it will support OpenCL which is destined to become the standard for GPGPU. OpenCL is still 6 months away, but it should replace NVidia's CUDA and the Brooks+ api AMD is backing with an interoperable standard capable of hooking into OpenGL.
Nvidia's new naming gimmick. Graphics cores? They're called stream processors. So you get 960 "Graphics Cores" from four cards? Why not just buy two 4870's or a 4870x2 like someone stated above. Since ATI's newest offerings have 800 "Graphics Cores" per card or 1600 for the 4870x2. 

I know that it will take custom software to enable what Nvidia's CUDA does but it seems to me that ATI would have a better solution if they decide to get into this HPC market.
...is if "GPGPU" stuff were supported at the OS level, thereby accelerating *all* aspects of the computing experience, rather than just stuff that's compiled specifically for whatever GPU.
Maybe Microsoft should be investing in a "Windows 7/GPU" (Like Windows 2/386) for all the normal people out there that know not to touch Intel video hardware.
A super computer is only a super computer because it's super, and takes more power than the combined energy usage of the world's vibrator supply. On full buzz.

As soon as it fits on a desktop it's not really super, is it? Because somebody could put 1000 into a big cluster, and that would then be the super computer. 

Bzzzzzzzzzzzzzzz