From the 60s to the 90s IBM and Cray battled each other in trying to build the fastest computers. Cray won pretty much every time, he raised the performance bar to the point that the only machines which eventually beat Cray’s designs were newer Cray designs.
IBM made flexible business machines, Cray went for less flexible and less feature rich designs in the quest for ultimate speed. If you look at what is planned for future GPUs [DirectX Next] it is very evident they are going for a flexible-features approach – exactly as you’d expect from a system designed by a software company. They are going to be using virtual memory on the GPU and already use a cache for the most commonly used data, in fact GPUs look like they are rapidly becoming like general purpose CPUs.
The Cell approach is the same as the Cray’s. Virtual memory takes up space and delays the access to data. Virtual memory is present in the Cell architecture but not at runtime, the OS keeps addresses virtual until a software Cell is executed at which point the real addresses are used for getting to and from memory. Cell also has memory protection but in a limited and simple fashion, a small on-chip memory holds a table indicating which APU can access which memory block, it’s small and never flushed, this means it’s also very fast.[/quote:971fdcccdb]
Viewing 0 reply threads
The forum ‘General Discussion’ is closed to new topics and replies.