Bitgrid - A new computing architecture, which seems absurd at first, but might be the one that democratizes access to Petaflops/second. Take an FPGA, rip out a

Bitgrid - A new computing architecture, which seems absurd at first, but might be the one that democratizes access to Petaflops/second. Take an FPGA, rip out all of the routing hardware, and basically have a sea of LUTs in a grid, clock them in two phases like the color on a chess board. It's a huge waste of transistors, but every single bit of compute is pipelined, so you could run an LLM with one output per cycle, for example. I'd estimate a 8192 * 8192 array could fit in a chip the same size as my i7 processor. (Assume an FP16 multiply fits in 1024 cells, there could be 65,536 of them, each giving an answer each clock cycle, at 2Ghz --> 131 TFlop FP16) If you manage to optimize the multiply to fit in a smaller area, you can immediately get more TFlops out, something that an FPGA can't do. I'll be glad to help anyone who wants to make this thing go.