We've seen a couple cautious
attempts at leveraging the raw floating-point capabilities of modern high-powered graphics cards, but NVIDIA is taking the gloves off with the launch of Tesla, its new general-purpose computing platform built on the 8-series graphics cards we all know and love. According to NVIDIA, the only way to skirt the inevitable collapse of Moore's Law is to join the GPU and CPU together, so two of the three Tesla configs are in the form of workstation upgrades -- a $1,499 single GPU PCI Express card and a $7,500 dual-GPU "deskside supercomputer" that plugs into a custom PCI controller. The truly crazy can pony up a full $12,000 for NVIDIA's first rack units, the four-GPU Tesla S870, which has a peak performance of 2 Teraflops. We're hearing the card and deskside unit will be available in August and that the servers will start shipping in November or December -- perfect for the
Engadget Folding@Home holiday rush.
posted Jun 20th 2007 6:10PM
I've gotta ask. Intel and AMD are both heading for hardware that really kicks butt on performance per watt. However in the GPU market ATI and Nvidia both are churning out hotter and hotter cards. When is the market going to start asking for better performance per watt GPU's? And when are the manufacturers going to respond?
hahah...
Do you have any idea how many flops per watt these cards get, and how amazingly efficent they are if that rating is taken into account?
Core 2 Quad has a max per cycle of 4 * (4 cores) FLOPS * speed (mhz, we will say 3.33Ghz) = 53.28 * 1,000,000,000 Hz, or 53,280,000,000 FLOPS. If we say the Core 2 quad wastes 135 watts, then we have a FLOPS/watt of 394,666,666.66... The 8800 ultra uses about 175 watts. Therefore, 2,000,000,000,000/175 = 11,428,571,428.57... Alot more efficient per watt, wouldnt you say?
Please take note that my math might be wrong, so i ask anyone to correct me. Realise that i gave the Core 2 Quad a greater chance to win by being lenient in its power usage and its FLOPS capability (im sure in the real world, its less than 16 per cycle) but this none the less illustrates the power that we all take for granted in these modern GPU's.
http://icl.cs.utk.edu/hpcc/hpcc_desc.cgi?field=Theoretical%20peak (I assumed 4 related per core, if its the entire CPU, then it does alot worse)
http://en.wikipedia.org/wiki/Intel_Core_2#Quad_Core_Kentsfield (the qx6800 is rated at 135w)
http://www.legitreviews.com/article/496/2/ (8800 ultra ratings, used as a blueprint. This might waste more, but who cares, it kicks the crap out of any CPU out today)
Benchmarks in Quake 4, Fear, etc?
What exactly is that and why doesn't it have outputs?
It's an off-video-card GPU.
All data transmitted through PCIE and/or (hopefully) secondary channels straight back to the video card?
Can it play Doom?
I wonder how long till we get that kind of processing power in a HANDHELD.
Can I use that with DAZ|Studio? :)
renderrenderrenderrenderrenderrender
wow u guys are special this isent a video card its a cpu/gpu. and for preformence per watt the fastes intel quad core runs at 130 watt and has about 30 gfps? this dose 2000 gfps with a max of 250 watts you do the math wait let me do it for u
the the nivada comes out with 8gfps per wat
the intell comes out with .230 gfps per wat out course the nivada cpu/gpu need specific programing it still has the potentionl to be 32x more powerfull
Wow, you all are hilarious!
The GPU will always have higher FLOP's power than a standard CPU, why you may ask? Because that is what it was designed to do. CPU's are not floating point processors by nature, they are general purpose processors that are capable of far more than the GPU alone.
This entire article is misleading, the GPU will _supplement_ the CPU only.
Comparing FLOPS of CPU vs GPU is just another worthless Apples to Oranges comparison.
Oh, theres no doubt about that, but the statements were meant to counter those that feel that GPU's are getting more and more energy hungry. Everyone knows that GPU's are parallel monsters (at least most know) and can never compare directly to a CPU, but it is evidently clear that given its power, it is energy efficient in comparison to a CPU.
While i agree that it wont happen anytime soon, GPU's definently have the capability of being CPU's (similar to the CELL, where the "texture units" can serve as integer and floating point units, of course completely remapped, but the concept of separate units holding true) but there is no need for that amount of processing power today or in the next 10 years, considering that everything else will bottleneck it.
@Ruben
You're comparison is fundamentally flawed. the FLOPS/watt argument is bunk in the first place.
These GPUs are very much similar to the Pentium 4, where by you could drastically increase performance by feeding it more energy rather than be higher transistor counts. many people consider the P4 a bust, but for what it was designed for (high GHz ranges), it was by no means.
simply put, the GPU will never match the efficiency of a CPU unless they are of the same process technology, which is currently not the case; and the fact the GPU requires as much power as it does to perform at the level it does says something about its core architecture as well.
"In terms of pure number crunching horsepower, Nvidia told us that one GeForce GPU can match the combined performance of 40 x86 processors. In addition to the raw performance, Tesla also makes a case for power efficiency: The C870 is rated at a maximum power consumption of 170 watts and the GPU server at 800 watts, which may sound a lot at first look. However, 40 low-power x86 processors would run at a typical 1600 watts. With a common power budget of about 25 kilowatts per rackserver, a Tesla GPU server rack has a theoretical maximum performance of more than 60 TFlops – which would put the floating point rating of such a device among the 15 fastest supercomputers currently ranked on the Top 500 Supercomputer list."
http://www.tgdaily.com/content/view/32557/118/
Oh, and its 518000000000/170 = 3,047,058,823.52...
Ruben, you are again trying to make an Apples to Oranges comparison to support your argument. the GPU is a specialized floating point processor, the CPU is not.
That last comment was a qoute. Apparently neither Nvidia nor AMD/ATI seem to have a problem making a comparison between the two.
I am well aware of the inner workings of a CPU and a GPU, and am well aware of their fundamental differences (which are vast) and do not need to be schooled on them, but still believe that with 170 watts, the GPU is quite the processor, and i still believe its quite efficient compared to a current CPU. It in no way compares to a P4, as the latest changes to the architecture have introduced parallelism that is unheard of.
Marrying the fundamental parts of a CPU with the "stream processors" to act as Integer and floating point units gives us possibilities that the Cell processor only hints at (the Fusion better be good, thats if AMD still exists in 2009/2010). As it is, the numbers are impressive, and i hope Nvidia can get current GPU's to be used as secondary processing units for physics and other processes.
Yeah, that's cool and all, but will it blend?
Im sure it will Jason, im sure it will...
But for now only time will tell.
If they keep putting those fans, of course they'll blend.
The only thing that is keeping up with Moore's law is nVIDIA's pricing. It doubles every 18 months.
Basically does this yield results like having SLi on steroids?