http://fah-web.stanford.edu/cgi-bin/main.py?qtype=osstats

OS Type	Current TFLOPS*	Active CPUs	Total CPUs
Windows	200	210498	2074291
Mac OS X/PowerPC	7	8419	116953
Mac OS X/Intel	27	8632	52888
Linux	74	43460	314731
GPU	895	8139	16013
PLAYSTATION®3	1335	47352	549907
Total	2538	326500	3124783

As you can see - GPUs produce 4 times the achieveable Tflops per CPU that the PS3 does and thats without the latest Teraflop cards from AMD/Nvidia fully online yet. *4850/70 etc*

Within a year I would expect the numbers for the GPU client to be at least 4 times higher than the PS3 client.

It looks like PCs will reign supreme in the floating point areas of computing if IBM/Sony/Toshiba don't do something about it soon.

It is only logical that GPUs would run better. There are a TON of thread processing abilities with each GPU. Plus I dont think that the PS3 version has been really optimized like it should be.

That a pretty big disproportion right there. 5 times as many ps3 active cpu's produce only 1.3-1.4 as many teraflops than those gpu's. Basically the cell can pack up and go home (pun intended) if the Gpu's so much as double their amount.

And last year the Cell seemed invicible..Ouch.

Yeah, the stream processors on ATi cards eat GFLOPS for breakfast.

Is that GPU number only for Radeons X1900 still? I remember it was the case back then... and even then, the number favored the GPUs by... like a lot.

Still different tasks: The GPU only supports a small set of work-unit types. The CELL is more flexible and normal CPUs can do all work tasks (but slower).

you forgot this

Active CPUS are defined as those which have returned WUs within 50 days. Active GPUs are defined as those which have returned WUs within 10 days (due to the shorter deadlines on GPU WUs). Active PS3's are defined as those which have returned WUs within 15 days.

 

"As you can see - GPUs produce 4 times the achieveable Tflops per CPU that the PS3 does and thats without the latest Teraflop cards from AMD/Nvidia fully online yet. *4850/70 etc*"

The active cpu/gpu is just a count of how many systems have folded within the last 10 or 15 days.

The current tflops is the measure of processing of an unknown amount of cpu/gpu of each platform at the time the chart was updated.

A direct comparison like yours is flawed.

 

I know that a high end gpu is better at it but your comparison is faulty.

We're comparing cards like the 3870 with the Cell, when the latest cards come online I would expect the performance difference to shift even further.

These are current Flops counted so the GPU is fast catching up. With dirext x 11 and much more programable shaders then we will see the difference really open up with more work units being done on the GPUs.

Cell = jack of all trades master of none here.

Kyros said: Still different tasks: The GPU only supports a small set of work-unit types. The CELL is more flexible and normal CPUs can do all work tasks (but slower).

^ Listen to him, he's right. Those TFLOPS numbers are apples to oranges, in FAH's case they favor the more specialized hardware (GPUs over Cell's SPEs over PC CPUs).

 

NJ5 said: Kyros said: Still different tasks: The GPU only supports a small set of work-unit types. The CELL is more flexible and normal CPUs can do all work tasks (but slower). ^ Listen to him, he's right. Those TFLOPS numbers are apples to oranges, in FAH's case they favor the more specialized hardware (GPUs over Cell's SPEs over PC CPUs).

Whilst its true, we have to wait until at least direct x 11 before we see any real gains here. The current card - RV770 produces 100gflops of the scientifically useful double precision variety. Which is about 5 times what the cell can do.

Theres going to be billions of dollars poured into this race in a 4 way shootout for control of this market. Amd vs Intel vs Nvidia vs IBM. AMD and Nvidia already have a multi-billion dollar GPU market for their goods and the Cell uptake hasn't been that great so far. So I would pick one of those two rather than Intel or IBM at present.

Coming soon - 4+32 cell vs the next generation Direct x 11/12 GPU hardware vs Larrabe - Many core X86.

 

This statement needs clarification. The CELL is a single-precision processor. Its double precision ALUs are small. This is done because many usecases only need single-precision computations (games, video conversion, many scientific algorithms). Calling only double precision scientifically useful is dangerous. This depends on the usecase.

There is or will be a double precision CELL btw. So if you need double-precision: Buy an IBM Cell blade. GPUs are great for many usecases but they are simply limited.

It boils down to:

FLOPS: GPU > CELL > x86
Flexibility: x86 > CELL > GPU.