Field
notes
Notes, tutorials, and deep dives into high-performance computing, GPU programming, and systems engineering.
Finding the predictor in my own CPU
The branch machinery of a laptop CPU, measured from user space on its two core designs: misprediction cost, pattern memory, branch target buffer and return stack.
−9 %fewer mispredictions02 · Branch prediction · Oct 13, 2026 · 12 min readAccurate, but late: branch predictors at equal budget
Six direction predictors at equal budget: the most accurate one loses cycles once its answer takes time.
+18%bubble operand waits03 · Branches · Oct 13, 2026 · 17 min readBranches, measured: operand waits, prediction and the return stack
Three branch rules added to a SystemC RISC-V pipeline: operands read in ID, a 2-bit predictor with a branch target buffer, and a return stack, each checked by hand and by an independent predictor.
12×stage threads vs ISS04 · SystemC · Oct 13, 2026 · 16 min readA CPU in SystemC, without RTL
A five-stage RISC-V core in SystemC: 42 of 42 riscv-tests pass, and two timing models that agree on every cycle count.
Agent-firstlanguage vs Dijkstra05 · Agents · May 20, 2026 · 16 min readZeroLang Field Notes: An Agent-First Language Meets Dijkstra
Testing an agent-first language by implementing Dijkstra and probing its limits.
Tiny msgswhat drives overhead06 · MPI · May 20, 2026 · 12 min readTiny MPI Messages: The Cost of Convenience
What really drives overhead when sending tiny messages with MPI.
up to 7×faster integer pow07 · C++ Perf · Apr 08, 2026 · 15 min readBeating std::pow: Up to 7x Faster Exponentiation in C++
Making integer exponentiation 5.6–7.6× faster in a C++ solver.
Rerununtil stable08 · Benchmarks · Apr 08, 2026 · 18 min readYour Benchmarks Are Wrong
Adaptive reruns that turn noisy measurements into reliable benchmarks.
LLM vsweeks of tuning09 · LLM × C++ · Apr 08, 2026 · 14 min readCan an LLM Optimize My Code Better Than I Can?
Comparing LLM-generated kernel optimizations with weeks of manual tuning.
Branchlessbeat my AVX210 · SIMD · Apr 08, 2026 · 12 min readThe Fastest std::find: How Doing More Work Makes You Faster
Why branchless C++ beat my hand-written AVX2 search.
3 layoutsAOS · SOA · AoSoA11 · Memory layout · Apr 09, 2026 · 42 min readAOS vs SOA vs AoSoA: The Answer Is Not What You Think
Benchmarking memory layouts—and why the winner changes in a real simulation.
Evolvedcircle packing12 · Evolution · Apr 10, 2026 · 11 min readEvolving Algorithms With My Claude Code Plan: ShinkaEvolve + Meridian
Connecting ShinkaEvolve to Claude Code for a circle-packing experiment.
80×faster Python N-body13 · N-body · Apr 10, 2026 · 13 min readDoes Shinka Rediscover My SoA Tricks? A Python N-body Warmup
An evolutionary optimizer rediscovers SoA and Numba, speeding up Python 80×.