HPC Acceleration Abstraction Contract

This contract preserves one runtime shape for every accelerator path and prevents incompatible hardware-specific behavior from fragmenting the public API.

Scope

The contract applies to:

• Apple integrated GPU (Metal on macOS)
• Discrete CUDA/ROCm-class GPU paths
• TPU via compiler-backed execution
• ASIC/custom-circuit feasibility

Standard interface

All accelerator tracks must preserve this contract:

• CPU is the reference implementation
• Output envelopes and diagnostics must be bit-for-bit equivalent (within documented tolerance) to the CPU reference
• Public APIs remain unchanged by backend selection
• Benchmark packets must include runtime metadata, workload fingerprint, and comparison fields
• On failure or unsupported paths, fallback to CPU must be automatic and explicit

Standardized acceleration candidates

1. Keep NumPy as the required portable baseline in-core implementation
2. Keep Apple Metal as the integrated-GPU first-class pilot path
3. Evaluate a standard compiler path for dense-kernel workloads before broad discrete GPU or TPU implementation

The immediate single-library candidates: JAX/XLA for dense tensor program lowering, PyTorch MPS/CUDA APIs for pragmatic incremental execution, and XLA-style IR boundaries where supported.

Transition criteria

Progression from feasibility to implementation requires:

• A preserved contract comparison packet (CPU + target device)
• Reproducible benchmark evidence with warm-up and repeatable timings
• Clear owner and next-step fields for unresolved tasks
• A public decision record in Conductor