Skip to content
Work Writing About Contact

02 Writing

Notes

What the numbers did not cover, and what it cost to find out. Mostly measurement, mostly written after being wrong about something.

3 pieces

RSS feed
01

Edge LLMs that never block the detection path

Generation takes about eight and a half seconds. Dispatch takes 16.6 microseconds. Which of those two numbers you report decides whether the design is honest. LLMEdgeSystems 02

When frameworks lose on tiny models

TensorRT FP16 was slower than eager PyTorch. ONNX Runtime on the CPU beat several GPU paths. The model was small enough that none of the usual assumptions applied. SystemsTensorRTPyTorch 03

On measuring CUDA kernels without lying to yourself

A kernel I had explained away as floating-point noise was a data race. The explanation was wrong in a way I could have checked in one command. CUDAMeasurementResearch