Panajev2001a (g_marocchi@yahoo.com) on 8/3/05 wrote:
---------------------------
>David Wang (dwang@RWTexpanded.com) on 8/2/05 wrote:
>---------------------------
>>fastpathguru (fastpathguru@yahoo.com) on 8/2/05 wrote:
>>---------------------------
>>
>>>Re: The Mystery of the Missing FLOPS, is it possible that it's solved by accounting
>>>for the extra meat in the PPE? I didn't see that mentioned as a possibility.
>>
>>You need 68 flops per cycle to get 218 GFlops @ 3.2 GHz.
>>
>>It's easy to get 56 of those flops from 7 SPE's, using FMADD, but then you'd need
>>12 flops per cycle out of the PPE to get to 68.
>>
>>Again, assume that you can crank 8 Flops per cycle out with FMADD, you still need
>>4 flops per cycle. The PPE is multi-threaded, but the impression that I got from
>>the IBM folks is that the scalar and vector FP units are actually shared so you
>>really can't get more flops out of the PPE either. Regardless, even if the FP units
>>aren't shared, you still can't get 4 flops per cycle out of the scalar side of thing.
>>
>>So no, the beefier PPE can't get to 12 flops per cycle even if it had independent scalar/vector FP units.
>
>Why not ? See Gekko ;).

I've forgotten about Gekko entirely.

http://www.hotchips.org/archives/hc13/2_Mon/06ibm-gekko.pdf

So the 12 Flops/cycle can be attained by the PPE if the PPE's scalar FPU can do a 2 wide SP FMADD in addition to a separate and independent Altivec unit. That does make things quite curious. I'm not sure how many programmers will be able to make use of this capability, but I think you're right, the Gekko ISA extension to the DD2 PPE explanation sounds to be the most reasonable explanation that I have come across about how Sony was able to claim the 218 GFlops rating with only 7 SPE's @ 3.2 GHz.

Perhaps you can write a rebuttal article and we can link it to the current article. I think the readership in general will benefit from this discussion.