3 ms·
HBM and HMC are already in the process of bringing most of those advantages to GPUs and probably some CPUs. Last year's high-end GPUs introduced HBM with a 4kb
by wtallis 11y ago
HBM and HMC are already in the process of bringing most of those advantages to GPUs and probably some CPUs. Last year's high-end GPUs introduced HBM with a 4kb (0.5kB) wide memory bus with 0.5TB/s bandwidth. The only reason they didn't have more memory bandwidth was because they didn't need it and the die area was better put to use for compute resources. FPGAs may be the only general-purpose chips so far to actually ship that kind of memory bus, but they aren't the only ones that can.
- melted 11y agoGPUs aren't custom logic. They come with significant and fundamental programming model limitations. So I don't see much of a functional overlap FPGAs and GPUs. In particular, GPUs suck mightily at anything that involves branching. They're also not ideal for things like SDN, crypto, search retrieval/ranking, realtime DSP, and so on.
- wtallis 11y agoMy point was that memory bandwidth won't sell anybody on FPGAs if they didn't already have strong reasons to be considering FPGAs.