RIGEL

Characterizing and optimizing the Metal 4.1 tensor compute path on Apple Silicon GPUs — open the black box Apple ships, then beat its defaults with free, hand-tuned kernels on hardware people already own.

Codename RIGEL: the bright blue supergiant within the constellation Orion. Lineage callback — the prior paper was Orion (the ANE); this is a star inside it (the GPU tensor path). Methodology lineage: Orion (arXiv:2603.06728).

Current Phase

G0 skeleton complete; G1 (Phase 0 capability gate) is the next work — confirm what actually compiles and runs on this M4 Max before designing around it. No experiments have run.

Research Thesis

• Act I — Characterize. Metal 4.1 documents the interface of its tensor/MPP path but deliberately hides the hardware behavior: the cooperative_tensor layout is "opaque" and "device specific" (§2.22.3), the SIMD-group-matrix element-to-thread mapping is "unspecified" (§6.8), and acceleration is punted to the external Feature Set Tables. RIGEL recovers what the PDF refuses to state on Apple M4 Max — dispatch target, accumulator widths, the opaque cooperative-tensor fragment layout, the acceleration-vs-emulation dtype truth-table, and the microarchitectural cliffs.
• Act II — Beat it. Use those discoveries to ship open-source hand-tuned Metal kernels that outperform the stock MPP path on M4 Max in the regimes characterization exposes.
• The discovery bar (§0.1): if a finding could have been written by reading the PDF, it is not a result; if a kernel isn't faster than the baseline it tuned against, it isn't a contribution.
• Null hypothesis the project must defeat: the MPP tensor path / cooperative_tensor layout / 4.1 low-precision rows / hand-tuned kernels outperform the cheap baselines — scalar-ALU roofline, naive threadgroup-tiled GEMM, simdgroup_matrix, and the stock fp16 path — on metric M (median GFLOP/s, J/FLOP). If the baselines match within ε, the performance claim is falsified (still publishable as characterization; no kernel ships against it).

Deliverables

• Paper — Act I characterization + Act II optimization (MLSys-class / arXiv-first, Orion cadence).
• rigel-bench — MIT-licensed open microbenchmark suite; emits the CSVs and regenerates every figure.
• rigel-kernels — MIT-licensed hand-tuned Metal kernel library (the headline contribution): packed low-precision GEMM, layout-aware tiling, cliff-avoiding schedules, fused epilogues, ANE/GPU hybrid router.
• Constraint catalog — Orion-format, numbered; headline metric = count of HIDDEN+CONTRADICTED entries.
• Decision guide — which Metal matmul path for which dtype/shape/fusion/energy budget, built on measured facts.

Repo Guide

Path	Purpose
AGENTS.md / CLAUDE.md	Operating contract (byte-identical)
CODEX.md	Codex pointer to AGENTS.md
STATUS.md	Current state
CHECKPOINT.md	Exact resume point
tasks/atomic-task-list.md	Canonical atomic task list
spec/RIGEL-spec.md	The science (design doc)
docs/	Decision log, experiment log, reproducibility, risks, runbook, gate reviews
notes/	Append-only audit trail
configs/	Locked, versioned sweep configs (per phase)
runs/	Run artifacts with full provenance (bodies gitignored)
harness/	Swift/Obj-C++ host orchestration (rigel-bench host; language TBD — spec open-Q #1)
kernels/	MSL kernels: simdgroup_matrix, MPP host-bound, MPP cooperative_tensor, rigel-kernels
analysis/	Python: YAML config → CSV → stats / roofline / plots (the testable layer)

Working Defaults

• Device under test (FIXED): Apple M4 Max, 64 GB unified memory, Mac Studio. Pre-neural- accelerator generation — no GPU "Neural Accelerators" (those arrive on A19/M5). Every finding is reported "on M4 Max"; cross-generation claims are out of scope.
• Toolchain floor: macOS 26.x (latest installed, 26.4+ preferred), Xcode 26.1+ (26.0 had the MPP header bug), Metal Toolchain component installed, Metal 4 project template.
• Data posture: repo is text-free; the archived Metal v4.1 spec PDF, raw measurement bodies, and checkpoints live on the external drive, gitignored. Per-cell CSVs + run manifests are committed (they are the citable results).
• Source of truth: the shipped MPP headers, not the PDF, where they disagree — and every disagreement is logged as a constraint-catalog entry.

License

RIGEL is released under the MIT License — see LICENSE. This covers the rigel-bench microbenchmark harness, the Metal kernels, the analysis layer, and the committed per-cell CSVs; the forthcoming rigel-kernels library ships under the same license.