Inference

Compute for fun projects, on a small budget

Cohese runs on very little money. Almost everything here was built on the Apple M-series Macs we already have; Shoal was written and benchmarked on an M5 laptop.

Some ideas need more than a laptop: training imagery models like DeepSky's, or drawing neutral maps for every state at once. Those want NVIDIA or AMD GPUs we don't own, so we go looking for compute wherever we can find it: grants, cloud credits, spare hours on someone else's machines.

If you have any of those to share, we'd love to hear from you.

A rendered green FPGA board seen from above: a lidless FPGA at the centre whose exposed die shows the real placement of one Shoal carving engine, two memory chips with length-matched traces, a power section, USB-C, headers, buttons and status LEDs. — A rendered board. The exposed die on the FPGA is real: one Shoal engine as nextpnr placed and routed it on a Lattice ECP5-85F (full placement).