AMD's new Firestream chip tops 1 teraflop The latest round of graphics card dueling between AMD and Nvidia isn't just over high-end gamers. The vendors will also exchange blows for the hearts and wallets of your friendly neighborhood medical imagers, seismic modelers, and computational fluid dynamicists. AMD is refreshing the FireStream processor line with a new …
"Seriously, the GPUs our there are beginning to look like super computer nodes."
Not surprising, as they have been for quite a long time.
To put this into perspective, the original super computers from Cray had a computational speed of 100 megaflops.
This "video card" is 10,000 times faster and it's sitting idle in the millions of computers around the world who have the hardware installed but aren't playing a 3d game at the moment.
As these chips become generally more integrated into the average desktop environment, how do you think the world will use the computational power they provide?
Sukoku?
For the FireStream 9170, it used IEEE 754 floats, but with some minor limitations on parts of the standard; i.e., the result of a division might not be the most accurate one possible.
I think the same will be true of the 9250. However, the 9170 was supposed to have shipped last month, and there seems nothing on the web about that having actually happened. The 9250 is an improvement; I do hope it will in fact be available for purchase in the third quarter of this year, as claimed.
GPU's use IEEE754 floats.
I wonder what the double precision performance is like, specially compared to the latest nVidia GPGPU.
NVIDIA don't say whether their performance figures are double or single precision and AMD specify they are single precision only.
For the vast majority of the HPC crowd I suspect double precision performance is more important.
It all depends on what you are doing - there are many applications where single precision is sufficient. Or with a bit of thought the computations can be re-arranged to avoid requiring extreme precision e.g. avoid taking small differences between two large numbers.
In the old days when FPUs were about 1/1000 (or less) of todays speed and C would only compile to double precision floating point ops I had to resort to writing a small set of C-callable assembly language single precision routines for vector arithmetic.
It was worth it, reduced run times from 4 hours to just over 1hr !
"As these chips become generally more integrated into the average desktop environment, how do you think the world will use the computational power they provide?"
Stanford University's Folding at Home project, using public distributed computing to unscramble the structure of bio-molecules, has been using networks of GPU cards and PS3s for quite a while. Not surprisingly, they are now getting far more teraflops from this small group of machines than from the much larger group of CPU donors.
Wouldn't it be grand if some über-hacker could great a POVRay fork that used nVidia GPUs? (Hint! Hint!)
Of course, that would require that nVidia does actually support double-precision FP. Years ago, when Intel developed the first SSE instruction set, the POVRay maintainers stated that experiments had shown that single-precision FP just couldn't produce renders worth the disk-space they took up.
Anyway, POVRay development seems a bit moribund these days. Come on guys! I demand that you spend all your free hours writing code, free of charge, so that freeloading dilettantes like myself can produce clumsy, amateurish pictures that they never actually finish!
Oh wait, maybe I understand their reluctance, now... :-(
"Of course, that would require that nVidia does actually support double-precision FP."
From the Tesla article...
"Key to the release today of the Tesla-10 Series processor is the presence of 64-bit, double-precision floating point support."
Although you may have been referring to normal graphics card?
Yer darn tootin' I was thinking of "normal" GPUs. I don't think my budget _quite_ stretches to a rack of high-end nVidia HPC kit (In fact, these days I need bank finance to buy a stick of chewing gum.)
Indeed, what's been getting me excited over the last few years is the thought that a consumer-grade piece of hardware can theoretically outperform a warehouse full of the old Crays. The mind boggles!
Whilst people like myself drool over such power, whither then the average desktop? Even fancier rendering for Windows Vista?
Lastly:- single-precision is useless for computational fluid dynamics; the truncation error must be small, since the overall error increases with every iteration.
The Tesla 10 numbers are for single precision. Only 1/8 of the 240 shaders are capable of double precision, so 30 of them, while the other 210 are single precision only. So since they went from 128 shaders in Tesla 8 to 240 in Tesla 10 at not that much higher clock rate and just about doubled performance, it has to only be for single precision. Double precision would be a lot lower, quite likely a bit lower than the ratio ATI/AMD is claiming on the 9250.
.... Rewind to the End BEGIN Again. Start ... Run Weighty Spin Waits.
"What if this universe in which we live is purely a simulation running on a couple of PS3s in some other-dimensional kid's basement?" .... By Anonymous Coward
Posted Wednesday 18th June 2008 20:05 GMT
School them well then would be Wise, AC, for that is an Awesome Power which Cannot Permit Abuse for that Destroys IT and the Abuser.
"What if this universe in which we live is purely a simulation running on a couple of PS3s in some other-dimensional kid's basement?" .... By Anonymous Coward
In which case I wonder if the universe would still exist if the PS3's were turned off? The only purpose in running the simulation would be for the other-dimensional kid to see the results.
The Mandelbrot Set exists anyway, the only purpose in running WinFract is so that we may see it.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
