clpeak

clpeak (compute latency peak) measures the peak compute throughput, memory
bandwidth and latency that a machine's CPUs, GPUs and NPUs actually reach.
It uses small, tight kernels alongside each vendor's own tuned libraries,
and runs the same tests through every API that exposes a device, so the
results line up side by side.

What it measures:

- Compute: fp32, fp16, fp64, mixed-precision, bf16, integer and
dot-product rates
- Matrix engines: cooperative matrices, tensor cores and CPU matrix units
- Bandwidth: device, local and image memory, host transfers, and CPU cache
and DRAM bandwidth per level
- Latency: kernel launch, memory latency per cache level, atomics and
branch mispredicts
- AI workloads: matmul, convolution and a transformer block per data
type, each with its numeric error against fp32

This snap has the Vulkan, OpenCL and native CPU backends, and the ONNX
Runtime and LiteRT backends, which use a runtime installed on the host
(--onnx-lib, --litert-lib). Builds with the CUDA, ROCm and oneAPI backends
are on the GitHub releases page.

It has two front-ends over the same engine: the clpeak command (clpeak
--help) and the desktop app, clpeak.clpeak-gui, which is also in the app
launcher on x86-64. Runs can be saved as JSON (-o run.clpeak.json) and
compared with a later run (--compare).

Classic confinement lets clpeak reach the GPU drivers and device nodes it
benchmarks.

Source, issues and reference results: https://github.com/krrishnarraj/clpeak