IBM's benchmark for Cell = ONLY 42GFLOPS??

[Hardknock]

http://www.power.org/news/events/barcelona/11_chow.pdf

Could someone please review this, and explain it?? Ken Kutaragi was claiming 218 GFLOPS @ 3.2 Ghz at E3! What gives???

 

[gosh]

Which page are u referring to?

 

[jvd]

This is showing 1 spe

The cell has 7

In one test 1 spe gets 19 gflops . So 7 of them would get 133gflops . This is not counting the ppu .

The second test shows the whole cell chip 1x8 at 3.2 ghz getting 41.9 gflops on the test .


I don't see what the problem is . I believe we've talked about this actually before a month or so ago .

There will be code that the cell is piss poor at . But thee will be code that when the cell is properly programed for it will get stealler performance

Look at the first test . Going from library code to specilized code you gain more than double the performance out of the spe

 

[Bobbler]

It is true, in a sense.

One test program they used scored 42gflops and they compared it to brand X (which is probably a P4) which only got .4gflops. There are other "benchmarks" where each SPE got 19gflops.

All this shows is that your mileage varies with the code you use. This is the way it works.

 

[gosh]

how many Gflops is the PPU

 

[Shifty Geezer]

The 200+ GFlop is theoretical peak. No processor achieves its theoretical peak. The IBM benchmark shows real-world performance. It's very good!!

 

[PC-Engine]

These are basically benchmarks for specific types of math problems. Just like Linpack your GFLOPS throughput will vary according to the scientific problems being solved. It's more a measure of efficiency for Fast Fourier Transforms.

 

[gosh]

These are basically benchmarks for specific types of math problems. Just like Linpack your GFLOPS throughput will vary according to the scientific problems being solved. It's more a measure of efficiency for Fast Fourier Transforms.

Seems to then have around 55% efficiency.

 

[one]

old
http://www.beyond3d.com/forum/viewtopic.php?t=23845

 

[Sonic]

Very well, bring the old thread back form the dead.

 

[jvd]

See I knew we talked about it before