What exists—and what does not.
What is HASLAB?
HASLAB is an open-source project to build the computing engine for vision AI in devices such as cameras and robots. It is developing a checked path from a trained model, through software simulation, toward an FPGA prototype and eventually a chip. The software runs the pinned accelerator and declared host-tail path on tested inputs today. Read the plain-language overview.
Can I run HASLAB today?
Yes. Start with the numerical reference, simulator, conformance, and compiler/runtime tests. With the pinned model and calibration artifacts, you can reproduce nodes 0–217 and the declared Python host tail. Follow the getting-started guide.
Does HASLAB run YOLOv8n?
FLOAT and calibrated INT8 proxy accuracy are measured over COCO val2017. Accelerator nodes 0–217 execute through 412,092 HASLAB commands in the simulator. The synthetic diagnostic run matches 6,931,200 stored values; six further inputs match 37,900,800 stored values. COCO-wide command-path accuracy and hardware execution remain ahead. See the scope of each result.
Is an FPGA target available?
No. Board selection follows workload closure, compiler/runtime work, verified RTL, and measured resource probes on plausible targets.
Can I fabricate a chip?
No release is tapeout-ready. Process, SRAM, IO, power, clocking, DFT, physical verification, packaging, fabrication collateral, and bring-up evidence remain future work.
Is commercial use allowed?
The hardware and software licenses allow commercial use under their respective reciprocal terms. Third-party models, datasets, PDKs, IP, and tools retain separate licenses.
What should I contribute now?
Independent contract and conformance review, the remaining neck lowering, package/runtime validation, adversarial vectors, documentation, and reproducible verification are especially useful. Find current work.