4 ms·
A Breakdown of AI Chip Companies
- PragmaticPulp 5y agoThis overview seemed excessively cynical until I got to the end and realized that he’s actually trying to pitch his new AI Chip startup idea: Cherry Computers. Does this mean he’s done with the Comma self-driving product?
- Issaclabs 5y agoNo, I believe this will be like how Tesla made a chip of their own.
- syntaxing 5y agoIf his cherry line is suppose to work so good, why hasn’t anyone else done something similar? Is it the new expansion of RISC-V? Would love to hear the thoughts of others that specializes in this!
- baybal2 5y agoIt's simple, he misidentifies the problem as such. The problem all these AI companies have vs. Nvidia is not that nobody buys their chips because they are bad, but because nobody needs them. GPUs on other hand sell well because, well, videogames
- pclmulqdq 5y agoHPC also sells a lot of GPUs. A lot of ML folks don't seem to appreciate that neural net training (outside of Google) is actually a fairly small market for the GPU companies compared to games and simulators.
- pclmulqdq 5y agoI actually was a hardware developer for a while, and I'm not sure this will pan out like the author thinks. A 1 GHz CPU on an FPGA won't happen, even a 7 nm FPGA. Also, host memory latency and bandwidth will likely kill any hopes of putting the special- purpose CPU far from its memory. The best outcome is a new RISC-V matrix extension, though, which sounds pretty cool and useful given the Apple M1. Good luck!
- Jasper_ 5y ago1 instruction per cycle at 1GHz? With no custom memory architecture?? On an FPGA???? Keep dreaming, kid.