Well, we've gone a full month since the
last episode of NVIDIA's and Intel's
ongoing public feud, but it looks like Intel has now stoked the flames once again (albeit inadvertently) in a paper presented at the recent International Symposium on Computer Architecture. That attempted to debunk the "100X GPU vs. CPU myth," but it also contained the tidbit that GPUs are "only" up to 14 times faster than CPUs in running application kernels, which NVIDIA has more than a happily latched onto. In a blog post, NVIDIA's Andy Keane says that it's a "rare day" when a competitor states that their technology is
only 14x faster, and that he can't recall another time when he's "seen a company promote competitive benchmarks that are an order of magnitude slower." Of course, he then further goes on to note that Intel's tests were done with NVIDIA's previous generation GeForce GTX 280, and that the codes were simply run out-of-the-box without any optimization -- but, still, he seems more than happy to accept this bit of "recognition." In Intel's defense, however, the overall finding of the paper (linked below) is that the performance gap between a GTX 280 GPU and Core i7 960 processor is actually just 2.5X "on average," which NVIDIA hasn't highlighted for some reason.
posted June 24th 2010 1:52PM
The difference is that the GPUs take up more space, and use more power, if intel made a cpu with that much space (i.e. put lots of cores and ram into it) then they would kick nvidias butt
Keep mind--
GPGPUs still aren't as efficient as normal CPUs in actual benchmarks. If you follow the Top500 list, you'll notice #2 (a Chinese GPGPU-based system) is theoretically much faster than #1 (Jaguar, an AMD-based Cray at Oak Ridge National Lab). The Chinese system, however, only gets 42% peak, and Jaguar gets about 75%. (Full disclosure: I'm a tech-support guy for Jaguar, but check out top500.org if you don't believe me).
@Revolutionary
its like a difference between a intel pentium and celeron
i've seen 1.7Ghz pentium run great while the 2.4Ghz celeron keep maxing out cause it sucks!
uhg...Im getting nauseous from all this spin...
I say that everything will have to rely on the CPU first. It doesn't matter how fast your GPU is if you don't have a powerful enough CPU. GPU is great for many things. Hardware acceleration, Video, Graphics, etc. But the heart of your system relies on the CPU. I get sick of Nvidia always talking how much better their GPU is to a CPU. Get the GPU to run without the CPU then we'll talk. There are things that the GPU simply don't do as good as the CPU. I think Jen-Hsung needs to optimize Fermi to use a little less power and exhaust less heat. And come out witth a Quadro based on the Fermi design that will use up the whole SP and not just some of them.
@MManLA
What would be good is if an ATi Fusion version uses a full GPU, like the full Evergreen GPU with GDDR5 on-chip, and just uses a far larger footprint than the average CPU/APU. That would be sweet, as the bandwidth to push raw processing to the GPU would be huge and there would be no software layer slowing it down. Imagine that. AMD could deliver that, in the near future if they worked it right. Might not be power efficient, and might need the biggest CPU fan in known history, but with a smaller, energy-efficient CPU, that could be astounding!
@MManLA Look up FASTRA.
350 Intel cores exhaust far more heat and use far more energy than 8 nVidia GPUs.
In their current forms, GPUs are reliant on the rest of the system because they're primarily designed for games. Build a special-purpose system around a GPU, however, and it becomes entirely feasible to ditch the CPU altogether.
You wouldn't be able to run Windows on it, as it stands, and you'd have to rethink the entire OS/File system/Application hierarchy, but it's entirely possible, though you'd have to reinvent the PC from the ground-up, specialized for certain applications.
My respect for NVidia drops with each of these stupid cartoons they release.
I'll pass on the childish antics and stick with AMD/ATI - at least they have the smarts to act maturely and simply deliver solid products to their customers instead of picking fights with Intel.
Can't help but think Nvidia is run by a child.
I can't help but notice that nobody spoke about the power consumption.
Intel CPU's use much less power, when performing the same task, as a NVidia GPU.
Not to mention, CPU's are optimized for integer calculations while GPU's are optimized for floating-point calculations.
Each piece of hardware has it's place in your PC.
@MattWeiler Power consumption has been mentioned. It has been also mentioned that measuring the power usage of a single CPU and comparing it to a graphics card (which also has memory, buses, fans, you name it) is not a fair comparison, and that when you get down to it, the power consumption is never really an issue when you take into consideration the 10-fold (or more!) performance increase.
@MattWeiler Also, when you look at that FASTRA link on the first page of comments, you'll notice that the engineers there got amazing performance results:
350 standard PC CPU cores replaced with about €4k worth of off-the-shelf components and 8 nVidia GPUs.
In the words of Dr. Lazarus: By Grabthar's hammer... what a savings!
Wow. I was completely unaware of the 100x GPU vs CPU myth. That was interesting...
No one talking about buses? GPUs are faster instruction-wise in some cases, and the same can be said about the CPU. Ok...
So a GeForce GTX 480 has about 177GB/sec of memory bandwidth. A similarly high-end Intel proc has about 25GB/sec of bandwidth through QPI. So, you can take your "one number is 7x bigger" stat and stop thinking now if you like. Really though, whenever the GPU has to move data to any other component in the system, it's limited to a paltry 8GB/s because it does so over a PCI Express 2.0 16-lane bus. Even the PCI Express 3.0 spec, which won't see the light of day until next year, only doubles that to 16GB/s. In today's desktop computer, that's not an issue because the GPU isn't required to run the whole system, but it would be unreasonable to limit the whole system to an 8GB/s bus and expect earth-shattering performace.
The reason these video cards have such insane internal bandwidth is because they're specifically adapted to move large textures around in VRAM. Getting the texture on and off of the card is by far the most time-expensive operation, and in the fantasy world where your video card runs your system, every single piece of data that isn't generated as the result of an GPU operation comes over the motherboard bus. 8GB/s might be more than enough for a video card talking to the CPU during a game, but it will become quickly saturated when every single I/O device on your system is using it to communicate.