backtop

Print E-mail del.icio.us 26 comment(s) - last by PlasmaBomb.. on Nov 11 at 4:48 PM

AMD FireStream 9170  (Source: AMD)
ATI FireStream 9170 challenges NVIDIA's Tesla GPGPU line

AMD and NVIDIA both are looking to expand the GPU into uses other than gaming. When ATI announced Folding@home would run on its GPUs end-users began to see the potential in GPUs for things other than video games.

NVIDIA first announced its Tesla line of general purpose graphics processing units or GPGPU in June of 2007 with the launch of the Tesla C870, S870 and D870 products. At the core, most of the products were NVIDIA 8-series graphics cards minus the DVI outputs for sending graphics to the display of the PC.

The Tesla C870 is a high-end, quad-GPU external system. These Tesla products are used for purposes like medical imaging and render farming because GPUs are more efficient with parallel processing requirements.

AMD has now announced a competing GPGPU, FireStream 9170, which is it billing as the industry’s first GPU with double-precision floating point. The FireStream 9170 is powered by an ATI GPU that is capable of parallel processing with 320 stream cores and 2GB of GDDR3 RAM.

AMD guidance states the card can provide up to 500 GFLOPs of single precision performance. The interface for the FireStream 9170 is PCIe 2.0 x 16 and it is compatible with Windows XP, XP64, Linux 32 and Linux 64.

Not to be outdone NVIDIA announced today that it would be showing the world its first glimpse at the new Tesla S870 GPU cluster at the SuperComputing 2007 show in Reno, Nevada next week. NVIDIA says Tesla is delivering enormous speed increases across many compute-heavy fields such as finance, oil and gas.

Comments     Threshold


This article is over a month old, voting and posting comments is disabled
questions regarding the article
By tobrien on 11/9/2007 5:30:32 PM , Rating: 3
This article is really interesting but I'm curious as to what complex compute-heavy operations oil and gas industries do?

also as is quite obvious the card has no physical output port on it, so is there a master card that DOES? Or is a GPGPU like this for a "headless box" of sorts where all the ins and outs are going through the LAN?

lastly, about how much does a whole system (either bare minimum config or fully decked out) go for that would fully utilize what these cards are intended to be used for or go in?




RE: questions regarding the article
By sapiens74 on 11/9/2007 7:26:45 PM , Rating: 1
quote:
This article is really interesting but I'm curious as to what complex compute-heavy operations oil and gas industries do?


Computing how much to steal from us? which requires supercomputing due to the sheer number


By FITCamaro on 11/10/2007 8:04:32 AM , Rating: 2
Personally I'd rate this to a zero. But thats just me.


By PrezWeezy on 11/9/2007 7:59:32 PM , Rating: 2
I believe that a GPGPU is more like a co-processor. It doesn't output, it's simply for general purpose (ie GP) computing of complex calculations. Oil and gas industries have huge spreadsheets that require this so they can predict how much oil will cost in the future, how much they need to buy, how much they are going to loose, how much it's going to expand or contract in the barrells, how long it will take to arrive, process, and be out to market. It's a very complicated business. But to answer your real question, this isn't at all for outputting to your screen. It's simply another processor for specific tasks. Very similar to the add in cards you can buy for financial companies.


RE: questions regarding the article
By FoxFour on 11/9/2007 10:29:26 PM , Rating: 4
quote:
This article is really interesting but I'm curious as to what complex compute-heavy operations oil and gas industries do?


Fluid dynamics and thermodynamics in refinery design come to mind...


RE: questions regarding the article
By spluurfg on 11/10/2007 1:26:30 PM , Rating: 3
As well as modeling potential explorations and challenging drilling operations. You remember those stupendously high resolution lcd displays showed in trade shows last year? They were demo-ing some of Halliburton's exploration software. This exploration software is custom tailored for each company -- my friend at BP tells me that theirs is GBP30,000 per license.


By Geoff Langdale on 11/10/2007 2:09:00 AM , Rating: 4
They set off explosives underground when exploring, then map the shock waves at various locations with a bunch of sensors. They then attempt to, using absolutely enormous sets of calculations (I've heard stories about out-of-core matrix multiplies of terabyte-sized matrixes) work out the structure under the ground and work out where oil/gas might be.

At least, that's what I've heard about way back in grad school. I may have garbled the details, but I suspect I'm a lot closer to the truth than "really big spreadsheets" or "fluid dynamics in refineries".


RE: questions regarding the article
By grath on 11/10/2007 2:11:24 AM , Rating: 4
quote:
I'm curious as to what complex compute-heavy operations oil and gas industries do?


Processing seismic sonar, tomographic, and ground penetrating radar data into usable information and maps of whats actually underground so they can determine where to drill, the quantity and quality of the crude in a particular field, expected yields, how to most efficiently extract it, etc.


Crysis
By BladeVenom on 11/9/07, Rating: 0
RE: Crysis
By Serlant on 11/9/2007 2:43:18 PM , Rating: 5
Yes! but you can't see it..


RE: Crysis
By munky on 11/9/2007 2:49:06 PM , Rating: 2
If you want to play games on it, you're looking at the wrong product. A card like this will run circles around your best quad-core cpu in HPC applications, and now for the first time a card with this capability supports double precision floats, which would make it even more appealing.


RE: Crysis
By BladeVenom on 11/9/2007 3:38:23 PM , Rating: 2
I was joking.


Oh my science.
By therealnickdanger on 11/9/2007 2:24:22 PM , Rating: 4
That card is hot! GPU pr0n!




PCIe 2.0?
By Von Matrices on 11/9/2007 2:49:59 PM , Rating: 2
Is this a different core than R600? The card and heatsink look identical, as do the specifications, but as far as I know, the R600 does not support PCIe 2.0. Is this a new chip?




RE: PCIe 2.0?
By PlasmaBomb on 11/11/2007 4:48:17 PM , Rating: 2
It is based on the R670 core.


Not New, and Tesla wasn't first...
By Darkskypoet on 11/10/2007 3:16:13 AM , Rating: 2
This isn't new, and Tesla wasn't first. AMD / ATI released X1900 series cards as GPGPUs as well. Why do you think there is a client for them for Folding @ Home.

When the X1900's were used for the last model of these, AMD developed and released Close to Metal which was the SDKish /APIish thing for the cards.

As well, not only folding, but other tasks as mundane as running a heuristic anti-virus on a massive raid array can be sped up greatly by these cards...

Sigh, its one of the reasons the r600 was designed the way it was. Because GPGPU cards sell for so much more then video cards do, and the massive programmability of the ATi Cards make it superior in many ways to the NVidia offering.

With CTM, you can do almost anything you want with the r600 because it is such a very general purpose set of Stream processors... rather then being fully specialized for graphics.




By Laitainion on 11/10/2007 5:54:50 AM , Rating: 2
You're getting confused with the concept of GPGPU's (as in this article) and the corresponding product (e.g. Tesla) and the concept of being able to do general purpose calculations on these graphics chips.

You are right, ATi did introduce an SDK to allow folding@home on a few of it's GPUs first. But that was only ever secondary to the main purposes of those cards. If you like, it could be seen as a test run, to see the validity.

Tesla *was* however the first 'finished' GPGPU product to go on sale as a dedicated GPGPU box, for whatever that counts for.

I would also question your statement about ATi offering better programmability as the wikipedia articles on both CTM and CUDA state that both allow the programmer to directly access the native instruction sets. Since both lines of GPU's were essentially designed for directx 10 and shader model 4, it's hard to see where ATi cards could get this extra programmability but feel free to prove me wrong.

Finally, I do not believe that any of the current generation of graphics cards are really specialised for graphics since all graphics involves are lots of floating point calculations. With the introduction and maturing of shaders (specifically unified shaders) the modern graphics card is no longer a purely graphics chip, but a massively parrallel floating point monster.


By Orbs on 11/9/2007 4:00:51 PM , Rating: 3
It would be great to see Folding@Home and other cool, parallel applications, designed for use with main-stream video cards. Currently (to my knowledge) only a handful of ATI cards work with Folding@Home, and I don't know of any other consumer-level applications that take advantage of the largely unused graphics horsepower (other than games).

It would be great if every nVidia DX 9 & 10 card and every AMD/ATI DX 9 & 10 card could run Folding@Home for example.




I don't believe they get it
By Nanobaud on 11/9/2007 4:16:24 PM , Rating: 2
Both ATI and Nvidia seem to be looking to sell these Firestream and Tesla cards in the mid $2K and up range, even though they are essentially the same chips as on their $200 graphics cards, except the graphics cards have the 64-bit precision and a few other gpgpu enhancements disabled. Problem is, any software that works well on these will generally also work well on small clusters of CPUs (usually better, due to higher composite memory bandwidth and size. These algorithms are as much or more about getting the data out of and back into the memory as the math.) As long as they put the cost of one of these processors higher than getting 4 or 5 simple quad(or more)-core boxes, they are not going to create a better value point for the end user, and they really miss the opportunity that could be there. (Probably the same argument for game-physics coprocessing, I don't know the needs there well but I imagine that needs dual-precision as well. With dual-precision disabled for all but the $2K+ cards, who's going to bother). There seem to be nascent system integrators dreaming of selling $40K single-box supercomputers lining up to design around Tesla's / Firestreams, but they are just blowing smoke up each others skirts.

I write the kind of software that could expand the applicability of GPGPU clusters (mine is to a limited audience, but the principles are the same) and I was ready to do a port to CUDA when the rumors were of double-precision capability on consumer cards, but I am not going to bother if it is limited to overpriced 'professional' cards that only open up what they are blocking on the affordable cards. I am sure they won't miss me, but I bet you can project the same onto more significant developers.

/rant

nBd




It all makes sense now..
By oopyseohs on 11/9/2007 6:58:55 PM , Rating: 2
So THAT's where the HD2900XTX's went. ;)




Quad Config
By Mitch101 on 11/10/2007 8:04:49 PM , Rating: 2
We know ATI has been working and must have quad nearly complete so this card in a quad config might just blow away NVIDIA's dual config. Looking forward to the benchies.




To Wolfgangster
By mediahype on 11/10/07, Rating: 0
Ne1 notice there no outputs on the card?
By Blood1 on 11/9/07, Rating: -1
By Homerboy on 11/9/2007 2:44:16 PM , Rating: 2
its not a video card
read the article


By Homerboy on 11/9/2007 2:46:03 PM , Rating: 3
I guess I also wanted to add people care do care about other technologies besides "fill rates" you realize right? This isn't a gaming website.

Cards and technology like this will help in the fight against diseases, build better strong lighter faster "stuff", help prevent deaths in earthquakes and natural disasters etc etc.


By FITCamaro on 11/9/2007 2:47:49 PM , Rating: 1
DYENHG it is to AEWL "anyone"?


"DailyTech is the best kept secret on the Internet." -- Larry Barber

DailyTech Poll
Do you use copy/paste on your smartphone? 




14 Comments









botimage
Copyright 2010 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki