Calgary (AB) – Nvidia’s Tesla cards are getting into gear, as Acceleware becomes one of the first companies to be offering a deskside supercomputer that can approach a performance of 2 TFlops.

It is no secret that general purpose graphics processor (GPGPU) are likely to play a more and more important role in next-generation supercomputing solutions. A first glimpse what these systems may be able to do is offered by Acceleware’s ClusterinABox systems, which are demonstrated at the SuperComputing 2007 conference.  

The company will be offering standalone accelerator cards - slightly modified high-end Nvidia workstation graphics cards – as well as complete ClusterinABox workstation systems that are based on AMD Opteron or Intel Xeon processors and come with either two or four Tesla cards in one or two deskside units.

As reported before, GeForce 8-based graphics cards can take advantage of Nvidia’s GPGPU software platform CUDA and hit a peak performance of 500 GFlops per GPU. In Acceleware’s A20 and A30 (Tesla) cards, this performance translates into a simulation horsepower of 400 – 450 Mcells per second. The company says that a minimum of 100 MB of hard drive space is required for the card drivers and either Intel Pentium 4 or an AMD Athlon FX or Opteron processor are minimum requirements to run the cards, which will sell for an estimated $10,000.

Customers who are looking for a complete solution can opt to purchase an Intel or AMD dual-core workstation that comes with either one (ClusterinABox Dual D20/D30) or two deskside units (ClusterinABox Quad Q30). The product scales from two Opteron processors and 8 -16 GB of memory and 500 GB of HDD space with two to four Tesla cards from a peak of 800 Mcells/s to 1700 Mcells/s, according to Acceleware.

By consumer standards, these systems are well out of reach for most of us, costing between $25,000 and $60,000. By supercomputer standards, however, this price could be a bargain and organizations that do not need the enormous memory that is available in supercomputers, these Tesla systems could become an interesting alternative.

how to unleash the power?

Nov 01, 2007 22:12     

I am just wondering.. how to unleash that collosol power??? I mean.. which language to be used??? JAVA, C, Fortran ????
Alien report abuse vote down vote up Votes: +0

less money for PS3's horsepower

Nov 01, 2007 22:59     

Using CELL instead of those GPUs is far more interesting, cheaper and faster, hence the clusters we see ppl creating for advanced calculations cuz you can gather a supercomputer computational power with a few paralel PS3s.. http://www.tgdaily.com/content/view/34660/113/ http://www.tgdaily.com/content/view/33218/135/
wateva report abuse vote down vote up Votes: +0

CELL doesn't come close to these GPUs

Nov 02, 2007 09:01     

@wateva Follow this link: http://www.tgdaily.com/content/view/34660/113/ Quote: "Sony says that over 670,000 PS3 consoles have registered to join the Folding@Home network, bringing the total amount of computing power to the program at over one petaflop (one million gigaflops)." This means that you need 670 PS3s to get to one TFlop, whereas the Tesla reaches 500 GFlop per GPU. Not even in the same league, I'd say :)

TeslaCoil report abuse vote down vote up Votes: +0

Unleash the power with CUDA

Nov 02, 2007 11:35     

@Alien NVIDIA created a new programming language, CUDA, specifically to unleash the power of these GPUs. It's essentially C for the GPU. See: http://www.nvidia.com/cuda.
ibolyalany report abuse vote down vote up Votes: +0

@teslaCoil

Nov 05, 2007 03:13     

I'm pretty sure the folding@home project doesn't use up even close to all of the cores of the cell processor. So, I don't think your calculations are valid
loujob report abuse vote down vote up Votes: +0

@PS3 fanboys

Nov 21, 2007 14:28     

The PS3 is not a supercomputer. Not even close. Even if you were able to get speeds close to this (linking many together in a cluster f*** that wouldn't be conventional in the first place) no one would write software for it becuase the CELL was the biggest design flop ever, thats why there are 4 good games for it sense release and why developers do anything they can to avoid porting to it.
Berg report abuse vote down vote up Votes: +0