Some proofs of work are built to resist ASICs. They lean on large memory, or change their work often enough that fixed logic falls behind a programmable processor. Every proof of work is still compute, and compute can be accelerated.
For these, Chipchain evaluates and designs accelerators that trade fixed logic against programmability, and on-die memory against external memory, and weighs their energy per hash, performance and return on investment against CPUs, GPUs and FPGAs before a mask set.
Between a fixed ASIC and a GPU lies a range of designs. An ASIP is a processor whose instruction set is cut to the algorithm. A CGRA is an array of coarse compute tiles that can be rewired between jobs. An eFPGA puts a block of programmable fabric on the chip, next to fixed logic. A structured ASIC fixes most layers and customises a few. Domain-specific accelerators and many-core arrays sit closer to the fixed end.
Memory is the other axis. Memory on the die is fast and costly in area; memory beside the die is cheap per bit but costs energy on every access. The right split depends on how the proof of work reads its data.
The balance moves with the market. In 2018 the company analysed Equihash and found the memory an Equihash chip needed cost more per unit of mining return than the register-dominated logic of a SHA-256 chip, so the team put its effort into SHA-256 then.