home About Projects Culti7 Blog Courses Papers

Projects & Experiments

Things I've built and broken. Each one started with a question I couldn't answer without building something first.

field notes from the bench · every write-up ships with its code on GitHub

ESP32-S3 · Xtensa LX7

Complete

Matmul Deepdive on ESP32-S3

Does the $3 AI chip actually do AI?

A ground-up benchmarking study of matrix multiplication on the ESP32-S3. Four implementations built from scratch, from the naive three-loop version to cache blocking and PIE SIMD, measured with cycle-accurate timing. The short answer: yes, but only up to a point, and the reason why is more interesting than the headline number.

CSIMDCache blockingBenchmarking
Read the write-up
ESP32-C3 · RISC-V

Complete

How Numbers Behave on a Chip With No FPU

Checking the claims against real silicon.

An investigation that turns each claim about machine arithmetic into a tiny on-board probe and reads the verdict off the serial port. Floats are torn apart bit by bit, pushed off both edges into infinity and the subnormals, and set against integers and fixed-point, on a RISC-V chip with no floating-point unit, where every quirk shows plainly. A companion to Unit 1 of The Mathematics ML Runs On.

CIEEE-754Fixed-pointUndefined behaviour
Read the write-up

In progress

More coming

Something is being built. Check back soon.