MemryX

A separate chip that does the AI, so the computer you already have does not have to. The MX3 is an at-memory dataflow accelerator — weights sit on-chip (10.5 MB per chip) next to the compute, which is what lets it run one image at a time at consistent latency instead of needing batches to be efficient. It ships as a chip for chip-down designs and as modules you plug in: the Cascade 100M (M.2), 100P (PCIe card), 100R (Raspberry Pi 5 HAT+) and 100U (USB). Their two selling points are "no retraining" — you compile the model you already have — and one software flow across every form factor. MX3+ and MX4 are on the site as the next generations.