Intel has revealed a lot more about Larrabee since then. If you are interested in GPGPU, you might want to check out the paper Intel presented at SIGGRAPH : "Larrabee: A Many-Core x86 Architecture for Visual Computing"
As a person currently working with both CUDA and the AMD Stream SDK, I am pretty excited about Larrabee. If nothing else, I hope Intel will simply provide better documentation and perhaps better software tools :|
I remember reading through some Nehalem articles mentioning that Intel will actually beat AMD to the CPU/GPU on the same package with AMD delaying Fusion.
Of course, if it ends up being a G45 class chip, it's not going to be much to write home about (performance wise). Hopefully it ends up being a Larrabee variant, but that might be hoping for too much.
I'm interested to see what AMD and Intel end up doing with an on core GPU. It could be really interesting from a GPGPU perspective as you could eliminate the current need to ship things from main memory to GPU memory.
Of course, if it ends up being a G45 class chip, it's not going to be much to write home about (performance wise).
This is the dirty secret of Fusion. Moving a dog-slow GPU from the chipset to the processor still leaves you with a slow GPU.
I'm interested to see what AMD and Intel end up doing with an on core GPU. It could be really interesting from a GPGPU perspective as you could eliminate the current need to ship things from main memory to GPU memory.
Unfortunately, a GPU tends to have ~150 GB/s of memory bandwidth while a processor socket has ~30 GB/s. Is it worth a 5x drop in bandwidth to avoid the PCI Express bottleneck? Maybe...
http://www.siggraph.org/s2008/attendees/program/item/?type=p...
As a person currently working with both CUDA and the AMD Stream SDK, I am pretty excited about Larrabee. If nothing else, I hope Intel will simply provide better documentation and perhaps better software tools :|