Megakernels on Mac: What Fusion Saves and What It Costs
Persistent workers, memory reuse, and the dependencies that survive fusion. Why an improved megakernel still lost to optimized separate kernels.
Read the essay← All writing
Engineering field notes · 3 parts
Fast inference on Mac.
The goal was 200 tokens per second from a 27B model on one Mac. These are the mechanisms, the approaches that lost, and the measurements behind Pulsar.
Drafters and Speculative Decoding: Useful Tokens, Correct State
DFlash, DFlash 2 and DSpark: block proposals, target verification, retained state and the cost of a complete speculative cycle.
Read the essayInside Pulsar: Faster Mac Inference, Measured End to End
Making a 27B model fast on one Mac: the speed gains, the moving baseline, and the kernel and drafter experiments that failed.
Read the essay