- Saturday, 21 November 2009
- INQ Mobile
- RSS
I think it's inevitable we move to 64-bit architecture - Bob Colwell, former chief architect, Intel
Transtec unveils 4 Teraflops personal supercomputer
CUSTOMISED HARDWARE MAKER, Transtec, has unveiled a couple of rather compact Cuda supercomputers based on Nvidia's Tesla graphics cards, boasting up to four Teraflops of computing power.
Transtec's 1000W Cuda supercomputer - which the firm is touting as "an individual supercomputing workstation" - sports two Tesla C1060 cards with over four Gigabytes of storage each, an Nvidia Quadro NVS 290 graphics card, 8 Gigabytes of RAM, an Intel Core 2 duo and a 160 Gigabyte SATA-2 hard disk for a purported two Teraflops of computing power, all packed into a small(ish) 43 cm high by 53 cm deep tower.
Meanwhile, its big, rack mounted brother, the 1000R, packs four Tesla 1070 cards with one Teraflop of computing power each plus 16 Gigabytes of graphic storage, 960 cores, up to four hot-swappable SATA hard drives and up to four Intel Xeon processors with up to 128 Gigabytes of main memory, all in just a single 1U rack unit.
Transtec boasts that a comparable CPU-based system would require about 60 quad-core servers and quotes the Nvidia mantra that the computing power of contemporary graphics cards is "significantly higher than that of a multi-core processor."
With just a hint of vindicated smugness, an Nvidia spokesgoblin told the INQ that Transtec's announcement means "you don't have to take our word for it any more" and that "manufacturers understand the benefits Tesla can deliver."
She also noted that a GPU-based system offers "a double whammy" when it comes to cost savings and performance boosts in computationally intensive parallel applications. "There's a real groundswell of demand for these products and we expect the design wins to keep on coming," she asserted.
The 1000W and 1000R are both compatible with Linux and Windows and the firm recommends them to customers interested in "particularly intensive computing operations" which include "process simulations or rendering and modelling complex data items."
We recommend them to punters with particularly intensive bank balances, as the 1000R will set you back a minimum of £10,300 and the 1000W sells for at least £5,560. µ
Share this:
ShareBusiness Intelligence Software
Customer Relationship Management (CRM)
Enterprise Accounting Software
Enterprise Resource Planning (ERP)
Enterprise Systems Management
I like CUDA, and I really wish nVidia would let me have a beta of it's openCL implementation to play with. That said...
Why do vendors like to put Quadro cards in Tesla boxes? Many people use Tesla for number crunching without requiring high quality OpenGL video output. The price difference between a Quadro and a GeForce product is not small: surely a Tesla box with a consumer video out card would sell well?
My $0.02.
And I thought CUDA was s**t because the Inq told us so.
Charlie has pointed out that nVidia had serious manufacturing problems and sought to obscure that fact, and also pointed out that nVidia has been indulging lately in rather egregious price upgrades via product renaming.
I don't recall an article on the Inq that claims CUDA is good, bad or indifferent.
FWIW, I own a pricey nVidia card and a pricey AMD card. I like them both. I reject in advance any allegations of bias as being invidious, nothing less.
CUDA isn't crap because it's NV, it's because it's not open... How well do these systems perform as OpenCL processors?
(OpenCL = Apple, ATI, Cell, Larrabee, etc.. Why get stuck on a proprietary piece of tech from a flaky vendor?)
CUDA isn't crap because it's NV, it's because it's not open... How well do these systems perform as OpenCL processors?
(OpenCL = Apple, ATI, Cell, Larrabee, etc.. Why get stuck on a proprietary piece of tech from a flaky vendor?)
You have short memory.
http://www.theinquirer.net/inquirer/news/1026466/acceleware-cuts-half-staff
"While it is sad to see a promising company like Acceleware implode, it does show exactly what we already knew, CUDA and GPGPU is on it's last legs. That Nvidia has been pouring money into it is a rather transparently desperate attempt to make people believe that it is still worthy. It isn't."
Like nvidia or not, every worthwhile university could save hundreds of thousands of dollars on supercomputer equipment and really push research where there wasn't budget and available processing resources.
I'm not sold on consumer level stuff for cuda other than video encoding... but for scientific research, a $10k box can do 'some' projects faster than some computers that cost 10x that.
Now if only professors were as good at programming as they are at thinking....
More like poor eyesight. I'd not seen that article. I think Charlie is incorrect regarding GPGPU.
CUDA may be largely superseded by OpenCL. From what I've seen CUDA looks closer to the metal (heheh) than OpenCL, so I would be surprised if CUDA went away altogether.
How Much does person need supercomputer? One way to check is install cpu/memory gadget that displays amount of use at any particular time.
If cpu gadget is near zero & goes to 50% on cpu, sometimes to 70 or 80 %, probably you need few more cores, NOT supercomputer, per se'.
However, if cpu is at 60-80% & often freezes, well, its NOT that much per terraFlop, compared to few years ago.
Will it play Crysis?
Core 2 CPU and a pro OpenGL card: it will run Crysis OK though not spectacularly. You could be running arithmetically intense kernels on the Teslas while fighting aliens and improbable N. Korean soldiers. Sounds kinda like the perfect computer.
I think the Inq should start a 'Don't tell Charlie' category....
- where they can safely deposit all the positive news articles about nVidia, without fear of upsetting a certain member of their reporting team!
If you were going to pay 10,000 pounds for that poorly lit crappy looking case with photoshopped logos you must be crazy....
At least they should provide a photo of a production model with decent stickers or something. Doesn't help the bland look of the thing but oh well it is a workstation.
While it uses a chip that's designed to be a GPU first a co-processor second, it is not 16GB of graphic storage it's just storage, unless you consider the GPU-centric fragment data as used by CUDA full blow 'graphics'.
Also, it's not going to play Crysis at all, since the 1070 is a rack mount solution and has no video out; and Crysis doesn't support it running as a video co-processor (ala SLi) to push out the tiny onboard graphics chip it would likely pair with such a solution. Running Crysis in software render mode with these is likely going to be much much slower than even and old G80 running Crysis.
As for the fate of CUDA, it's going to be pushed back to niche market status, only slightly more widespread than CTM and CAL. Micro$oft's compute shader implementation in DX11 will likely become more popular, with OpenCL taking top spot by a long shot.
While it uses a chip that's designed to be a GPU first a co-processor second, it is not 16GB of graphic storage it's just storage, unless you consider the GPU-centric fragment data as used by CUDA full blow 'graphics'.
Also, it's not going to play Crysis at all, since the 1070 is a rack mount solution and has no video out; and Crysis doesn't support it running as a video co-processor (ala SLi) to push out the tiny onboard graphics chip it would likely pair with such a solution. Running Crysis in software render mode with these is likely going to be much much slower than even and old G80 running Crysis.
As for the fate of CUDA, it's going to be pushed back to niche market status, only slightly more widespread than CTM and CAL. Micro$oft's compute shader implementation in DX11 will likely become more popular, with OpenCL taking top spot by a long shot.