Skip to content

Competitive HPC landscape (Li)

One-page map of who we track and why. Source of truth: benchmarks/competitive/registry.toml.

Li’s pillar order applies: provability before speed. Competitive intel informs gaps and bench honesty; it does not justify unproved parallel for or unsafe shortcuts.

Bench columns (harness today)

Ecosystem Track CSV lang What we compare Honesty
C++ / Clang OpenMP bench_tier2 cpp MD, matmul, SIMD micro, PDE Native reference
Rust bench_tier2 rust Same kernels via shared C core shared_c_kernel
Julia bench_tier2 julia Tier-2 physics + micro shared_c_kernel
Li bench_tier2 li Same fixtures mixed (pure_li vs shared)
OpenMP / races bench_tier0 — Tier-0 correctness, race rejects CI strict

Watch list (manual review)

Ecosystem Why track Li implication
Chapel PGAS HPC language Locale / distribution model → future G-*
Kokkos Performance portability Team policies, backends → LLVM/OpenMP alignment
SYCL / oneAPI Heterogeneous offload Watch; no Li GPU story in v1
Zig Low-level SIMD/comptime Patterns for codegen, not a bench column yet
NumPy / BLAS ML tier baselines Tier-3 only; label BLAS explicitly

Process

  1. Registry — add/change ecosystems in registry.toml.
  2. Validate — ./scripts/check-hpc-competitive.sh.
  3. Bench — bench.py --tier 12; read benchmarks/results/latest.csv.
  4. Review — quarterly manual release notes for watch rows; bump last_reviewed.
  5. Gaps — file master-plan G-* items when adoption needs a proof path.

Agents: use skill hpc-competitive-review. Rule: li-hpc-competitive.mdc.