§ Writing
Notes & derivations
Long-form technical posts on inference, hardware–software codesign, compilers and machine learning systems. Every number derived, most charts interactive.
Fig. 1 — Specimen acceleratorJun 2026
Codesigning an Accelerator from Agent Traces
Profile a tool-using Claude session, fit a hardware model, then design and optimize a custom accelerator — scale-up and scale-out — with every number proved.
Fig. 2 — Throughput frontierApr 2026
Understanding the SemiAnalysis Inference Curves
Rooflines, the KV-cache tax, TTFT and inter-token latency: a model for reading inference throughput frontiers, built around DeepSeek R1.