What's the point of cell?

[Groucho]

Man. This topic (not this particular thread) should be a sticky one.

Let me try to paint a simple picture of the Cell, relative to its competition, and with regards to its duty as a part of the PS3:

Typical console game, single frame:

(A) CPU runs game logic
(B) CPU runs physics
(C) CPU does animation
(D) CPU sends completed animation data to GPU
(E) GPU does skinning of animated character skeletons
(F,) GPU processes vertexes into renderable triangles
(G) GPU renders textured triangles to screen
(repeat)

Typically, performance on these sections goes like:
(A) game logic (PS3 okay, X360 pretty good, PC awesome)
(B) physics (PS3 okay to awesome, depending on programming investment, X360 pretty good, PC pretty good)
(C) animation (PS3 awesome, X360 pretty good, PC pretty good)
(D) not really a big performance hit or difference, I just threw this step in there for clarity
(E) skinning (PS3 awesome with programming investment, X360 pretty good, PC awesome with modern GPU)
(F) vertex processing (PS3 pretty good, X360 pretty good, PC awesome with modern GPU)
(G) texel/pixel processing (PS3 okay to pretty good, X360 pretty good, PC awesome with a modern GPU).


The issue that makes PCs faster than the PS3 or X360 generally boils down to the fact that games are unilaterally toned down to run decently on the CPU for the lowest common denominator, and then are limited on the back end -- often mostly by step (G). Fill rate is one of the few things that the Cell cannot assist with -- its pretty much entirely GPU dependant. Since many, many games tend to be "fill bound" (that "G" step is the bottleneck), generally the fastest filling GPU "wins"... and that boils down to the PC in every case.

When it comes to raw number crunching, the Cell beats the socks off any PC CPU you can compare it against. The trouble is that number crunching really only comes into play during steps (B) and (C) above, and many games aren't even bottlenecked there to begin with (most games don't throw 100 bad guys at you at once, so you just don't need to animate that many). With enough programming effort, the Cell's SPUs can help the RSX (the PS3's GPU) out enough to best the performance of... certain competing GPUs, by simplifying the GPU processing pipeline at steps (E) and (F), but if the app is bottlenecked at (G), that still may not matter much. The performance of the GPU "hides" the CPU performance, in a sense, because usually the bottleneck at (G) is pretty serious.

The real benefit of the Cell, now that the yield is improved from the first half year or so, is that it has this extra power (admittedly minor in this first incarnation) without generating crazy amounts of heat, and in time, it should be slightly cheaper to make (it may be already) than its competition, because it uses less transistors than a typical design of the same performance would.

The next versions of the Cell family will have the same "tougher to write code for" issues that the current Cell does (maybe slightly less, as tools evolve), but, on a per-cost basis, they will be far more powerful than their peers.  That still won't solve the bottlenecks at step (G) -- the Cell could (and future versions very well might) kick every other gaming platform on the planet's behind at steps (A)-(F), and if there's still a bottleneck at (G), the end result will be the same.  You always have to have a good GPU to go with the CPU.   Because PCs are modular, and consoles are not, the PC will always "win" in this department... always.  Remember, a high-end PC GPU costs more than a PS3 does -- JUST the GPU.

The PS4 is set up to succeed, both financially and from a performance perspective, already. Assuming they stick with the Cell (and there's no reason for them not to, really), Sony fundamentally designed the future PS4 when they designed the PS3, and reduced their future HW costs to boot.  If Sony doesn't skimp on the GPU, the PS4 will be a great machine.  The Cell was a great choice, and it has very little to do with the PS3's issues in the marketplace at this time.