Next week in the IBM booth at Siggraph 2007 we will be showing a head to head comparison of our QS20 Cell blades and a Linux Playstation3 running the iRT ray-tracer. On head one a Playstation3 will be using its six available SPEs, the RSX will be sleeping comfortably, to ray-trace a 450K triangle scene directly into its 1080p resolution virtual framebuffer. On head two the IBM blade center will deploy 14 Cell processors or 112 SPEs to ray-trace, compress, and deliver frames to a second Playstation3 for display. I won't keep you in suspense, the blade center wins by a large margin. The point of this exercise is not to show that a blade center can out perform a game console but instead how a computational intensive algorithm can scale linearly in performance across many Cell processors and hundereds of SPEs. When the problem is broken down to exploit both data parallelism and Cell's on chip local store bandwith, performance scale very well across many blades. This exercise also shows the versatility of code bases created with the Cell SDK, the same executable targeted for the QS20 blade can be executed on the PS3 without modification.
The iRT now uses automatic code overlay support to switch shaders and read/write software caches to spill, sort, bundle, and evaluate secondary rays. These two coding features allowed the programming team to ignore the SPE's 256KB local store size for both code and data fetch and eliminated the need for programmer directed DMAs. Stop by the booth next week to see this technology demonstration and talk about Cell programming, ray-tracing, or whatever's on your mind.
Here is a Quicktime Movie (16MB) with voice over, showing some of our overlayable shaders and sample frames from this technology demonstration.
