[cec]

The paper says "we started a high-priority project to quickly produce a custom ASIC for inference (and bought off-the-shelf GPUs for training)."