StreamBench

Free plan LinuxMacWindows

What StreamBench is

StreamBench is a free, open-source benchmark for CPU and GPU memory bandwidth plus AI inference on CPU, GPU, and NPU hardware. Memory tests run the Copy, Scale, Add, and Triad STREAM kernels. AI tests report model load time, first-response time, and warm-run tokens per second through Microsoft Foundry Local, LM Studio, or Ollama. Foundry Local supports CPU, GPU, and NPU benchmarking; LM Studio and Ollama do not support NPU tests. The command-line interface displays results and exports JSON or CSV files. Range testing sweeps array sizes and saves consolidated CSV output, while custom profiles and automation are supported. StreamBench runs on Windows, macOS, and Linux, with Windows and macOS binaries in x64 and ARM64 variants. Linux users can use documented launch scripts or build from source. GPU testing uses dynamically loaded OpenCL without a GPU SDK dependency. Saved results may include hardware details and the hostname. The project is limited to single-node benchmarking.

Who it suits

StreamBench suits users comparing modern CPU, GPU, and AI hardware who want visual results, prebuilt binaries, custom profiles, or command-line automation.

What it does well

Free and open source
Measures CPU, GPU, and NPU workloads
JSON and CSV export
Custom profiles and range testing
Windows, macOS, and Linux support

What to know first

Focused on single-node benchmarking
NPU tests require Foundry Local
macOS Foundry Local is GPU-only
Saved files may include hostname

Verdict

StreamBench combines STREAM memory kernels with AI inference timing across major desktop platforms. Backend choice affects available devices, and exported files should be reviewed before sharing.

Facts from StreamBench's own pages

AI backends: The AI inference benchmark supports Microsoft Foundry Local, LM Studio, and Ollama. · github.com, 11 Oct 2026
AI integrations: It supports Microsoft Foundry Local, LM Studio, and Ollama as AI backends. · github.com, 11 Oct 2026
AI measurements: The AI benchmark reports model load time, first-run response time, and warm-run tokens per second. · github.com, 11 Oct 2026
Array size: The README recommends using arrays at least four times larger than the largest cache and notes that small sizes can be skewed by cache effects. · github.com, 11 Oct 2026
Automation: Range testing can sweep array sizes and save consolidated CSV output. · github.com, 11 Oct 2026
CLI or automation: Yes · github.com, 22 Sept 2026
Custom test profiles: Yes · github.com, 22 Sept 2026
Device coverage: Foundry Local supports CPU, GPU, and NPU benchmarking; LM Studio and Ollama do not support NPU benchmarking. · github.com, 11 Oct 2026
Downloads: The project offers prebuilt binaries for Windows and macOS, with x64 and ARM64 variants. · github.com, 11 Oct 2026
GPU testing: The GPU memory bandwidth backend uses OpenCL, which the project says is loaded dynamically without a GPU SDK dependency. · github.com, 11 Oct 2026
Intended users: The README describes it as aimed at users seeking a quick visual benchmark with prebuilt binaries, GPU coverage, and AI inference comparisons on modern hardware. · github.com, 11 Oct 2026
Limits: The project is focused on single-node benchmarking and does not include the MPI support listed for upstream STREAM. · github.com, 11 Oct 2026
Linux support: The README documents macOS and Linux launch scripts and says the project can be built from source on Linux. · github.com, 11 Oct 2026
macOS limitation: Foundry Local on macOS currently supports GPU mode only, without CPU or NPU device targeting. · github.com, 11 Oct 2026
Measured AI metrics: It reports model load time, first-response time, and warm-run tokens per second. · github.com, 11 Oct 2026
Memory tests: Its CPU and GPU benchmarks run the Copy, Scale, Add, and Triad STREAM kernels. · github.com, 11 Oct 2026
NPU limitation: NPU benchmarking requires Foundry Local; the README says LM Studio and Ollama do not support NPU benchmarking. · github.com, 11 Oct 2026
Operating systems: Windows; macOS; Linux · github.com, 22 Sept 2026
Platforms: Prebuilt downloads are offered for Windows and macOS, while the project also documents Linux use and source builds. · github.com, 11 Oct 2026
Privacy: Saved results and CLI summaries may include hardware details and the hostname, and the project advises reviewing files before sharing them publicly. · github.com, 11 Oct 2026
Purpose: StreamBench benchmarks memory bandwidth and AI inference across CPU, GPU, and NPU hardware. · github.com, 11 Oct 2026
Requirements: AI benchmarking requires at least one supported AI backend, which the setup script can install. · github.com, 11 Oct 2026
Result export formats: JSON; CSV · github.com, 22 Sept 2026
Results: The CLI displays benchmark results and can save them in JSON and CSV formats. · github.com, 11 Oct 2026

Where StreamBench stands on Freedom251

PC Benchmark Software#56 of 73
PC benchmark tools#58 of 70