mlperc

mlperc is an open-source software project that provides benchmarking and profiling tools for machine learning workloads. It helps researchers and practitioners evaluate performance across models, datasets, hardware, and software stacks in a reproducible and transparent manner. Central to mlperc is collecting and analyzing latency, throughput, memory usage, and energy consumption, with percentile-based metrics (for example p50, p90, p95, and p99) to summarize tail behavior.

Key features include a benchmark harness to define workloads, support for common ML frameworks, configurable targets (CPUs, GPUs, accelerators), and instrumentation for timing, memory, and power measurements. Results are emitted in machine-readable formats such as JSON or CSV and can be rendered into reports or dashboards. The project emphasizes reproducibility through environment capture, deterministic seeding, and versioned workloads, and supports extensibility via plugins or modules for new workloads or backends.

Architecture is typically modular, with a workload driver, a measurement layer, and a results aggregator. It supports both synthetic microbenchmarks and real-model benchmarks and can orchestrate tests across multiple devices or nodes.

Status and governance are community-driven; licensing and release cadence vary by distribution. See also MLPerf, ML benchmarking.

Notes: publicly available information on mlperc may be limited, and the term may refer to different tools in different organizations.