Intel has conclusively debunked the myth that GPUs are 100x faster than CPUs running application kernels. In the recent International Symposium on Computer Architecture (isca2010.inria.fr) that was held in Saint-Malo, France, Intel published a white paper titled “Debunking the 100X GPU vs. CPU Myth: An Evaluation of Throughput Computing on CPU and GPU” (PDF) and in it Intel concludes:
In this paper, we analyzed the performance of an important set of throughput computing kernels on Intel Core i7-960 and Nvidia GTX280. We show that CPUs and GPUs are much closer in performance (2.5X) than the previously reported orders of magnitude difference.
Two and a half times faster than a CPU. That’s a win for NVIDIA in my book. And this came from Intel no less. It gets worse. Intel’s Core i7-960 is one of the company’s latest and fastest desktop CPUs featuring four cores, a clock speed of 3.2GHz boostable to 3.46GHz, 8MB of cache with a sizable price of $562.00 per one thousand (source: Intel). On the other hand, NVIDIA’s GTX 280 is somewhat old and last-generation.
Not only are there faster 200 series GPUs (GTX 285 and GTX 295) there is a more recent 400 series (GTX 465, GTX 470, GTX 480) based on a new Fermi platform that came out earlier in the year in March. The GT 200 series had 1.4 billion transistors and the new Fermi-based 400 series feature 3.0 billion. A full GTX 480-based graphics card can be had for much less than a single Core i7-960 CPU. I am fairly certain if the much more powerful GTX 480 GPU was compared the conclusion would have been that GPUs are merely 5x faster than CPUs running application kernels. More computing power for less cost: thanks to Intel, I think that’s a major win for NVIDIA.