clpeak Live

clpeak (compute latency peak) measures the peak compute throughput, memory bandwidth and latency that your device's CPU, GPU and NPU actually reach. It runs small, tight kernels and real AI workloads through each API Android exposes, so the same test can be compared across them on one device. What it measures • Compute: fp32, fp16, mixed-precision, integer and int8 dot-product rates, on the GPU and across the CPU cores • AI workloads: matmul, convolution and a transformer decoder block (prefill, decode and per-token latency) in weight formats from fp32 down to int4, each with its numeric error against an fp32 reference • Bandwidth: GPU memory, local memory, image and host transfers; CPU cache and DRAM bandwidth per level • Latency: kernel launch, memory latency per cache level, atomics, and dispatch cost through the AI runtimes • CPU extras: AES, SHA and CRC32-C rates, string scanning and UTF-8 validation How it reaches the hardware • OpenCL and Vulkan on the GPU • Native CPU kernels, chosen for your processor's instruction set • LiteRT on the NPU (where the chip vendor's runtime supports it: Qualcomm, MediaTek, Google Tensor, Samsung), the GPU and the CPU • ONNX Runtime Using the app • Run everything with one tap, or choose devices, test categories and time budgets • Results stream in as they are measured, grouped by device and category, with a short explanation of every test • Every run is saved to History; rename runs and export them as JSON • The screen stays awake during a run, so later tests are not measured on a sleeping display Found a wrong number or a crash? Turn on Settings → Diagnostics → Verbose diagnostics, run again, export the JSON and attach it to an issue on GitHub.