About HWBench: who it is for
HWBench helps FAEs, integrators and engineers pick edge-AI hardware (GPU, NPU, FPGA, MPSoC, Versal) for vision and custom models first and local LLMs second, using cited, dated public benchmarks. Independent testing is coming.
Who it is for
- Field application engineers who need a cited number to answer “which board?” on a call.
- Integrators sizing a vision system against a power and cost budget.
- Engineers choosing between GPU, NPU, FPGA, MPSoC and Versal parts for a model they already have.
What it covers
- Vision and custom models first: detection, segmentation, pose, classification and related tasks.
- Local LLMs second: tokens per second where a source publishes them.
- 312 platforms and 7,347 cited results as of 2026-10-16.
How the numbers work
- Every figure is copied from a public source, linked and dated, and graded A–E for evidence quality (methodology).
- Comparisons are like-for-like by model-size class and batch, or flagged as mixed.
- Unknown power or price is shown as unknown, never estimated.
- Independent testing by HWBench is planned; until then nothing here was measured by us.
What it is not
Not a vendor site, not sponsored, and not procurement advice. See the disclaimer.