kars on 11 March 2007
Kwaad said:
The C2D changed things for the better. But still once agian, it is not the way the cell works.
The cell has 64mb memory on each SPE. (dedicated) meaning that one SPE can access that memory at 100% speed. or all 8 SPEs can access their memory at 100%.
And I still rais the issue: check your numbers. The amount of local memory that you grant a single SPE is more than the amount of memory available on the whole cell. A single SPE has 512 KB of memory and its basic limitation lies in the interaction of its units and in task switches, and if i am not totally mistaken this will become a major hazard in the cell of the PS-3. Either you don't really use them efficiently or they are in the process of flooding the internal bus and their common memory interface. The cell has certain advantages and certain disadvantages.
Another thing I find funny is how much some people value special benchmarks. in reality ih the profesional world benchmarks are more or less seen as smoke screens. You really has to think about what you are measuuring and what kind of hazards you miss. The only realiable benchmarks are benchmarks with your real applications and real data.
