A Breakdown of AI Chip Companies
geohot.github.io
A Breakdown of AI Chip Companies
1–8 of 8 posts
Re: A Breakdown of AI Chip Companies
#2This 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?
Re: A Breakdown of AI Chip Companies
#3This 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?
No, I believe this will be like how Tesla made a chip of their own.
Re: A Breakdown of AI Chip Companies
#4If 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!
Re: A Breakdown of AI Chip Companies
#5If 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!
It'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
Re: A Breakdown of AI Chip Companies
#6I 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!
Re: A Breakdown of AI Chip Companies
#7If 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!
It'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
HPC 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.
Re: A Breakdown of AI Chip Companies
#81 instruction per cycle at 1GHz? With no custom memory architecture?? On an FPGA???? Keep dreaming, kid.