peak resident memory per workload and api-driver session-create deltas
Official release binaries — minor releases, plus patch releases from v2026.13 on — re-measured on the current perf runner. Click a workload name for its full sub-counter breakdown.
Session floor — peak resident memory of doing nothing but starting the compiler (createGlobalSession + core module), paid by every compile. There are two, because a process peak is only comparable against a baseline from the same executable: slangc's floor is compiling an empty shader (the minimal workload), and the api-driver's is creating and destroying a session and nothing else (api_session_create). Workload panels — each tracked workload's OWN memory: its process peak minus the same point's floor for the binary it ran, i.e. what the workload's work added beyond compiler startup. Raw peak RSS is recorded for every workload in results.json; only the curated set (the realistic rt/mdl workloads and the microbenchmarks with meaningful own memory) is charted and alerted on. createGlobalSession RSS delta — measured inside the api driver: process memory immediately before vs after that one call. Panels are line charts, not stacked areas, because these are separate memory peaks — maxima do not add up to a total the way the time pages' phase buckets sum to compileInner.
Methodology: release points are the OFFICIAL prebuilt binaries, re-measured on the perf runner; daily tip-of-tree points are built on the runner with release-matched flags but a different MSVC toolset. Each page's points share one build provenance; absolute times are NOT comparable between the release and ToT pages (a uniform few-% offset, and more on individual hot loops).