4 ms·
AMD needs to come out with a few AVX-1024 instructions for vector ops. Essentially make one core into a GPU that doesn't suck at branching.
by crb002 9y ago
AMD needs to come out with a few AVX-1024 instructions for vector ops. Essentially make one core into a GPU that doesn't suck at branching.
- sqeaky 9y agoI think there are some really hard limits on this now. I think the better at branching a chip is the worse it will be at SIMD and vice versa. Barring some crazy breakthrough I don't see what you want coming soon. Which is sad, because I want it too.
- vvanders 9y agoThere's a reason that GPUs suck at branching. Hint: It's because they're good at the vector stuff.
- opencl 9y agoA "GPU that doesn't suck at branching" is basically what the Xeon Phi is intended to be, with 72 AVX512-enabled Atom cores per chip. However it costs over $6000.
- pandaman 9y agoThe reason you think GPUs suck at branching is that they cannot branch on different lanes of an SIMD data (the 'SI' part guaranteed that). GPUs branch just fine between different instructions. There is nothing inherent to GPU making SIMD instructions to behave this way so SIMD instructions in any other PU will do the same.