Latest · July 2026
Making GPUs go brrrr
Unfortunately this won't be a free article. GPU kernel engineer are expensive.
I'm a software engineer working in Physical AI. I use writing as a tool for my own recursive self-improvement. If any of it's useful to you too, even better.
Making CPUs go brrrr
Robots have extreme real-time and latency budgets — here's how I optimized an on-premise x86 system, a Siemens BX-39A IPC, to meet them.
Fast isn't the same as good when picking a QP solver
Worst-case latency and low-variance solve time matter more than average solve speed for real-time IK control.
Know your sources of jitter: how to instrument for performance and determinism
I'll guide you through instrumenting a real-time 125Hz control loop for tracking accuracy, jitter, and external client interactions.
So you want to build a QP-based inverse kinematics controller for end-effector tracking
Using Stephane Caron's PINK library to formulate a QP problem: input is a task space end-effector pose target, output is joint velocities.