We prepared a comparison sheet of theoretical performance (number of transistors, peak performance, memory bandwidth) of CPUs and GPUs over the last 6 years.
Since it was quite an effort to collect all the data, we would like to share our results with you.
Maybe you have some comments to extend or correct it …
These kinds of lists are always useful! I always keep trying to compare different cards, often to see if some cheapo old card can still perform fast enough for an algorithm or if I need to tell a client to update their GPU.
There’s two more detailed lists here for NVIDIA and ATI cards.
attached are a table of HW from both AMD/ATI, Intel, and Nvidia, some numbers are still missing. There is also a graph that looks at the peak GFLOPs developments since 2000.
THe GPU is also surprisingly powerful for a several year old design. I believe it was loosely based on a suped-up version of the NVidia GTX7800. Some talk of the GTX380 coming in at around 3TFLOPS.
I don’t believe the 1,8 TFLOPS of the GPU. It was clocked at 550 MHz, had 300 million transistors, 24 pixel shader units and 8 vertex shader units. Its closest relative on the PC, the 7900 GTX, had the same number of shaders, roughly as many transistors, was clocked higher (at 650MHz) and achieved 280 GFLOPS. This would be a much closer estimate of its theoretical peak. It didn’t have unified shaders yet BTW, so programmability was questionable.
It’s CPU isn’t that strong either. Each SPU gives a theoretical peak of 25,6 GFLOPS, yielding about 150 GFLOPS (6 active SPUs in PS3) in single precision. The SPUs also rate at about 10,8 GFLOPS in double precision in total (6 * 1,8 GFLOPS). Additionally, the PPU gives 25,6 GFLOPS in single and 6,4 in double precision.
Double precision performance has been increased considerably (to 50% of SP) in a newer revisions but I’m not sure if Playstations received those.
You’re referring to the PowerXCell 8i chip, which is only used in IBM blade servers and PCI-Express coprocessor boards (which are several thousand dollars each). Those chips do not go into PS3s.
What you say does seem to make some sense and I was kind of hoping for an answer like this as the PS3’s performance simply doesn’t make sense. If an Nvidia-based GPU was capable of 1.8TFLOPS in 2006, then how come the latest GTX295 struggles to reach that in 2009 and at a cost that’s more than an entire PS3? Doesn’t seem to make sense.
And looking at these, I’m struggling to see where the extra performance comes from. But how can they get away with a claim that’s braisenly false all the same?