Writing
Write-ups on HPC, GPU computing, infrastructure, and the projects behind them.
- Nov 16, 2026What NVML can tell you about a single kernelMy Kokkos energy tooling stopped at the scale of the whole GPU. On my own RTX 3080 Ti I measured how often NVML’s power and energy readings change, how exact its energy counter is, and whether CUPTI kernel timestamps can split board power between kernels. They can, as long as the kernels do not share the GPU.
- Nov 16, 2026Where the energy of a one-step image goessd-turbo makes a 512x512 image in a single denoising step. Measured stage by stage on my RTX 3080 Ti with the NVML energy counter, the step itself is not where most of the energy goes: at one step, the VAE decoder that turns the result into pixels takes 62% of the energy above idle and more than half of the time.
- Oct 15, 2026Reading GPU energy instead of guessing it, on an AMD MI210NVML only reports a power value every 100 ms, so my Kokkos energy connector could not see short kernels. AMD GPUs expose a cumulative energy counter instead. On an MI210 it moves about every millisecond, but one read costs a third of a millisecond, which decides how a profiler has to use it. Measured with it, ArborX fdbscan-densebox takes 38% less time and 42% less energy than fdbscan.
- Jun 18, 2026Tearing the GPU node down from Proxmox to bare DebianThe GPU node ran Proxmox. I wiped it for bare Debian 13 so the GPU would sit directly under the host kernel, then spent the evening in the NVIDIA driver gauntlet Trixie hands you. The part that caught me was Secure Boot. (It has since gone back to Proxmox.)
- Jun 12, 2026Charging GPU energy to the code that spent itA profiler tells you where GPU code spends time. I wanted to know where it spends joules, so I built a Kokkos Tools connector that samples power on a side thread and integrates it over each profiled region. On ArborX DBSCAN, the faster implementation saves more energy than time: 19% less time, 25% less energy.
- Jun 9, 2026A galaxy in a browser tab, computed on your GPU with WebGPUA real-time gravitational N-body galaxy, a brute-force O(N²) sum recomputed every frame in a WebGPU compute shader. 65k bodies, no server, no precompute.
Questions, corrections, or work: ethan.puyaubreau@gmail.com