Has the Cell peaked? If no, would it work for next gen? - More questions inside

[zarx]

darkknightkryta said:

1.  It was the general flow of the discussion, not directed at you just the misinformation.

2.  Cell was designed for multi core programming.  A lot of the decisions made for Cell was directly for that (Lack of cache for the SPUs, in-order exectution).  Programming specicially for it or not does't change multi-core programming paradigm it encourages it.

3.  Cell is more powerful than most, if not all, PC processors, so yes it's still pretty powerful.  As powerful as more modern server processors?  No, not unless they do another update ontop of the PowerXCell 8i revision.  But no one is willing to invest.

4.  I was agreeing with the GPU point.

5.  I disagree, it will always be performance vs flexibility.  If flexibility was the way to go everyone would have jumped ship to Larrabee.  There will be the convergance, so you're right with that, but they're going GPU to CPU (Current video cards) vs CPU to GPU (Cell, Larrabee).

Cell is only better at certain things tho, sure it runs circles arround a traditional CPU in terms of raw flops but that doesn't really reflect all workloads. No one has really prioratised SIMD performance on x86 in a long time unless you count larrabee/Xenon Phi which were never meant to be used as a CPU successor but rather GPU/coprocessor to be used in conjunction with a traditional CPU.

The CELL's asymetric CPU design seems to be dead, tho I suppose including a GPU on the same die as the CPU could be seen as a successor to that design.  I don't see anyone using 2 different types of CPU cores in the same die tho, that design did seem to be a dead end or at least one that IBM didn't think was worth pursuing.

So far in terms of heterogeneous compute GPU bassed design does seem to be winning the war but they are still a very far away from the point where you could do away with traditional CPU cores in a console let alone a PC. CPUs aren't replacing dedicated GPU any time soon ether, Intel and AMD have instead gone for the aproach of just including a GPU on the same die (with the option of an addon card) rather than try to do graphics on CPU cores. Maybe in 10 years we will see a new architecture that really does offer the ultimate balance of flexibility and performance that lets it supersede the current CPU+GPU paradime for games, but not this generation.