● Field notes — 13 posts on CUDA, C++, benchmarking & AI tooling03 years shipping · CMAP · ONERA · CEA
20 cyclesper misprediction01MicrobenchmarksFinding the predictor in my own CPU−9 %fewer mispredictions02Branch predictionAccurate, but late: branch predictors at equal budget+18%bubble operand waits03BranchesBranches, measured: operand waits, prediction and the return stack12×stage threads vs ISS04SystemCA CPU in SystemC, without RTLAgent-firstlanguage vs Dijkstra05AgentsZeroLang Field Notes: An Agent-First Language Meets DijkstraTiny msgswhat drives overhead06MPITiny MPI Messages: The Cost of Convenienceup to 7×faster integer pow07C++ PerfBeating std::pow: Up to 7x Faster Exponentiation in C++Rerununtil stable08BenchmarksYour Benchmarks Are WrongLLM vsweeks of tuning09LLM × C++Can an LLM Optimize My Code Better Than I Can?Branchlessbeat my AVX210SIMDThe Fastest std::find: How Doing More Work Makes You Faster3 layoutsAOS · SOA · AoSoA11Memory layoutAOS vs SOA vs AoSoA: The Answer Is Not What You ThinkEvolvedcircle packing12EvolutionEvolving Algorithms With My Claude Code Plan: ShinkaEvolve + Meridian80×faster Python N-body13N-bodyDoes Shinka Rediscover My SoA Tricks? A Python N-body Warmup