About this collection
Systems claims become useful only when the level of explanation is clear. A throughput result may come from execution width, shared core resources, memory behavior, an operating-system cache or simply a benchmark that exercised a different path. This hub gathers investigations that keep those boundaries visible across CPU microarchitecture, SIMD, concurrency, storage I/O and container runtimes.
The curated path starts with modern core sharing and vector execution, then moves outward toward Linux I/O and transactional memory. The chronological list includes embedded processors, Ruby process models, text compression and zero-allocation servers because software structure often reveals the hardware contract more clearly than a specification table. These notes are intended for comparison and diagnosis: clock speed alone is rarely enough, a completed benchmark is not automatically representative, and a result from one kernel, compiler or SoC should not be promoted into a universal rule without new evidence.
Start here
- Inside Apple M4’s SME engine A technical review of Apple M4’s SME implementation, its shared cluster-level throughput, ZA pipeline behavior, precision limits, and software implications.
- SMT, Spatial Multithreading, and the battle for a CPU core How classic SMT, NVIDIA Spatial Multithreading, many-way server SMT, and non-SMT CPUs trade throughput against latency, isolation, and power.
- SVE, AVX-512, and vector-length agnostic code A Ruby-readable explanation of how Arm SVE's predicate-driven model differs from fixed-width SIMD such as AVX-512, with practical cautions about Apple M1.
- RWF_UNCACHED and buffered I/O without page-cache pollution A short explanation of Jens Axboe's RWF_UNCACHED patch set: what it tries to solve, how it differs from direct I/O, and why it matters for fast storage.
- Transactional memory after the hype Notes on hardware transactional memory, Intel TSX, software transactional memory, Clojure STM, and why transactional memory survived in narrower forms.
All investigations
- Inside Apple M4’s SME engine
- SMT, Spatial Multithreading, and the battle for a CPU core
- From A14 to M5: Apple’s wide cores and the x86 single-thread race
- How a game crash left 8 GB wired in the Apple GPU stack
- Using caffeinate on an unplugged Mac
- Making macOS Tahoe faster, without snake oil
- Days Gone hordes, machine guns, and polite cheating
- Transactional memory after the hype
- Neural text compression: prediction, range coding, and deterministic decoding
- Apple Silicon GPU memory limits and wired RAM
- Zero-allocation HTTP servers in Ruby terms
- ESP32, Xtensa, ARM, and the Pentium MMX comparison
- SVE, AVX-512, and vector-length agnostic code
- Docker on WSL2, NixOS, and runtime boundaries
- Deploying multiple Rails apps on one VM with Capistrano
- RWF_UNCACHED and buffered I/O without page-cache pollution
- SSE, proxy buffering, and long-polling fallbacks
- SIMD string processing in a Ruby C extension
- Ruby Fibers and cooperative concurrency
- Kafka, PostgreSQL, and consistency boundaries
- dRuby, Unix pipes, and forked worker pools
Related main-site work
- Apple Silicon GPU Wired-Memory Limit Benchmark A current systems-measurement project with explicit restoration and reporting boundaries.
- Rails Dependency Pruner A runtime and static-analysis experiment that measures memory against latency.
- Ruby Concurrency Lab Runnable examples for threads, Fibers, Ractors and process forks, linked from the main-site project archive.