geohot is shilling his product, which aims to capitalize on making accelerator manufacturers compete with each other on FLOP / tensor core throughput.
the OP outlines what could be an entirely different compute paradigm for LLMs, hence the FPGA study. they just happen to also get impressive performance on GPUs making the most of the available interface.
Comments
Reminds me of ghotz's interview: https://youtu.be/wE1ZoMGIZHM
geohot is shilling his product, which aims to capitalize on making accelerator manufacturers compete with each other on FLOP / tensor core throughput.
the OP outlines what could be an entirely different compute paradigm for LLMs, hence the FPGA study. they just happen to also get impressive performance on GPUs making the most of the available interface.