For videos that have been viewed more than a month ago, the three PS3 are connected through routing, and can now implement real-time (or interactive) Ray tracing. Although I have known about cell before, it is shocking to have the results so quickly.
I took a look at paper. There are two common practices: the SPE is used as each stage of ray tracing and has different kernel. The same ray tracing kernel is run on the SPE to process different pixels on the screen. The first problem is that it is difficult to balance the load balance in each stage. To avoid stall, the design will be extremely complicated; the headache of the second method is that every SPE needs to access the geometric data of the entire scenario, which is obviously not kb ls, therefore, we need to design a software cache method to cover up annoying DMA operations. Fortunately, coherent for scenario data is relatively large between close pixels, so the software cache hit rate is also high, in addition, the software hyperthreading practice -- transfers the SPE after the cache miss to the pixels for processing data in other existing off-the-shelf scenarios, and processes the data after asynchronous DMA is reached.
For multiple cellbe instances, You can synchronize the scene data to the master memory of multiple cellbe instances through the network, and then use the Protocol to split the screen pixels for processing. To be honest, this kind of work is not small ......
From this point of view, with the huge success of Folding @ Home on the PS3, we have reason to believe that CPU similar to the cellbe architecture will become one of the mainstream high-performance computing in the future-but we still lack time and tools.
Digress: it seems that there are still few hands-on in China, and we hope to have more experienced friends to communicate with each other.