Skip to content

Backend Support

Milestone 4 adds the initial hardware/backend layer used by the benchmark engine.

Built-in Backends

  • CPU backend for CPU-targeted execution.
  • GPU backend for GPU/CUDA-targeted execution.
  • Reference backend retained for deterministic validation tests.

Hardware Inspection

SystemHardwareInspector collects a portable host snapshot using standard library probes and optional GPU hints when a compatible runtime is present.

Backend Selection

BackendRegistryImpl resolves backends by name when available and otherwise selects a backend that supports the requested device. The benchmark service can use an injected registry and hardware inspector when a direct backend instance is not supplied.

Validation

Backends validate device compatibility and execution precision before benchmark work begins. GPU validation also requires an inspected hardware profile with an available accelerator. Structured errors include the cause, a suggested fix, and a documentation link.

The built-in CPU and GPU backends currently advertise fp32, fp16, and int8. Requests for unsupported precision modes are rejected during backend selection or preparation, before warmup and measured iterations can run.

Hardware Data

The hardware profile attached to each benchmark session records host, platform, CPU, memory, GPU, accelerator, driver, thermal, and power-policy fields. The standard inspector uses portable operating-system probes and optional runtime checks when packages such as PyTorch are installed.

Public API

from aihw_bench.infrastructure.backends import default_backend_registry
from aihw_bench.infrastructure.hardware import SystemHardwareInspector

registry = default_backend_registry()
hardware = SystemHardwareInspector().inspect()
backend = registry.select(configuration, hardware)