Edge AI Coding — Part 3: Axelera Wingman
Axelera’s Wingman found the bug in my Metis benchmark code, taking ResNet-50 from 695 to 1601 FPS on four cores.
Posts
Edge AI Coding — Part 2: Creating Custom Skills
How I turned Part 1’s ‘skills from struggle’ insight into a working library of Edge AI vendor skills — one per accelerator.
Edge AI Coding — Part 1: Increasing Productivity with Claude Code
How I turned a hard-won OpenCV frame-navigation recipe — first written by hand, then re-discovered by AI coding the painful way — into a persistent skill that future AI sessions can reuse.
Edge AI Yocto Integration — Part 1: AMD PetaLinux 2025.2
A compile-survey of vendor Yocto layers (meta-hailo, meta-axelera, meta-deepx-m1, memx-yocto) against the scarthgap LTS release, on AMD Zynq-UltraScale+ platforms.
Edge AI Power Benchmarking — Part 6: Measuring the Power Efficiency of MemryX MX3
Reproducing MemryX’s published resnet50 benchmark on the MX3, with independent power measurement.
Edge AI Power Benchmarking — Part 5: Measuring the Power Efficiency of DeepX M1
Reproducing DeepX’s published resnet50 benchmark on the M1, with independent power measurement.
Bird by Bird — with Jeff Bier
A short story about my intersection with Jeff Bier, and how his advice still drives me after more than a decade.
Edge AI Power Benchmarking — Part 4: Measuring the Power Efficiency of Axelera Metis
Reproducing — and surpassing — Axelera’s published resnet50 benchmark on the Metis, with independent power measurement.
Edge AI Power Benchmarking — Part 3: Measuring Edge AI Power with INA228
Reproducing the Hailo-8 reference methodology with an Adafruit INA228 I2C power measurement board.
Edge AI Power Benchmarking — Part 2: Power Insertion with ElmorLabs
Reproducing the Hailo-8 reference methodology with an ElmorLabs PMD2 with PCIe to M.2 adapter with power insertion.