The PS3 Licked the Many Cookie

7 days ago

Copy Link

The PS3 failed developers due to its excessively heterogeneous architecture, which resisted composability.
Many-Core architecture trades off explicit programmer control for wide parallelism.
PS3's Cell processor had SPEs with 25 GFlops each, but even combined, they couldn't match the underpowered PS3 GPU (192 GFlops).
Xbox360's GPU had shared computational resources (240 GFlops), allowing better load balancing than PS3's separate vertex and pixel shading units.
Developers used SPEs for tasks like skinning and shader patching due to GPU limitations, requiring complex synchronization.
PS3's memory architecture was inflexible, with 256MB dedicated to graphics and limited CPU access.
SPEs had only 256KB local memory (effectively 128KB), making them restrictive compared to modern GPU compute.
Generic C++ code couldn't run on SPEs without significant modification, forcing handcrafted SPE-friendly code.
Naughty Dog showcased PS3's potential by tailoring their engine to its architecture.
Original PS3 design (4 Cell processors) could have been a more homogeneous Many-Core beast (~1 TFlop).
PS3's half-baked Many-Core implementation may have permanently damaged the concept's viability.

Hasty Briefsbeta