Open-source AI hardware research

From AI models to hardware you can understand.

HASLAB is building the computing engine that will let cameras and robots run trained vision models near their sensors. Today, its software executes the pinned detector’s accelerator partition and declared host tail on tested inputs. COCO-wide accuracy through that path, an FPGA prototype, and silicon remain future work.

Current progress

YOLOv8n accelerator nodes 0–217 run through the compiler and simulator. Six diagnostic inputs match 37,900,800 stored integer values exactly; COCO-wide command-path accuracy and hardware are still open.

Check the evidence →
01 / What it does

Picture a camera spotting a defect.

A trained AI model can recognize the defect, but a device still needs a way to run that model. HASLAB is building that missing path: translate the model into instructions, check the answers in software, and eventually run those instructions on hardware near the camera.

01 / THE INPUT

A trained vision model

Start with a model that has learned to identify objects in images. HASLAB's first example is a small YOLO object detector.

02 / THE BRIDGE

Software that translates and checks it

The compiler turns part of that model into machine instructions. A simulator checks the results against an independent reference.

03 / THE GOAL

Hardware near the sensor

Once the complete path is proven, the design can move to an FPGA prototype and later to an open chip.

Why build it in public? Model conversion, data movement, and correctness can each make or break a real device. HASLAB makes those steps visible so others can inspect, test, and improve them. Read the project principles →

02 / System path

From model to machine.

The diagram shows the intended route. Today, pinned accelerator nodes 0–217 execute through the compiler, runtime, and simulator; the declared host tail runs in Python. COCO-wide command-path accuracy and hardware stages remain ahead.

InputFramework model
PinnedONNX export
Nodes 0–217Compiler + package
Nodes 0–217Runtime + simulator
PlannedFPGA target
Long-termOpen ASIC
03 / Workload evidence

Smaller integers. Measured accuracy.

Quantization represents the model with smaller integers. The calibrated INT8 software proxy stays within the project’s one-mAP-point accuracy-loss budget.

FLOAT baseline

28.50 mAP

YOLOv8n · COCO val2017 · 320 × 320

Exact report value: 0.284969
Calibrated INT8 proxy

27.61 mAP

Signed-symmetric INT8 proxy · 320 × 320

Exact report value: 0.276098
Within the 1.0-point budget

−0.89 mAP points

3.11% relative loss · 96.9% retained

Exact loss: 0.887 points

Scores use the familiar 0–100 mAP scale; exact 0–1 report values remain visible for traceability. YOLOv8n · 320 × 320 · 5,000 COCO val2017 images. Both software evaluations were repeated with identical predictions. The INT8 result uses an ONNX Runtime QOperator proxy; COCO-wide HASLAB command-path accuracy and hardware measurements remain future work.

04 / What exists

What you can use today.

The source includes a tested numerical model, a command simulator, 46 independent test cases, and a compiler/runtime path through the learned detection head of pinned YOLOv8n. Read the exact comparison and track the remaining work.

Tested

Numerical golden model

FP8/INT8 conversion, accumulation, tensor operations, layouts, rounding, saturation, and edge cases.

Read the reference model ↗
Tested

Functional command simulator

A byte-level simulator for the proposed 128-byte v0 command ABI, including memory spaces, completion, reset, and errors.

Inspect the simulator ↗
In progress

Contract + conformance

46 binary fixtures and strict validation exist; independent review and the stable ABI freeze remain open.

Review the corpus ↗
In progress

Compiler + runtime

Nodes 0–217 compile to diagnostic and release packages. Six inputs pass exact stored-value comparisons and the declared Python host tail; COCO-wide accuracy is next.

Follow the software path →
Six diagnostic inputs matched 37,900,800 stored accelerator values exactly; 1,058,400 decoded host-tail values passed tolerance. COCO-wide command-path accuracy remains open.
Exact integer agreement on tested inputs is software correctness evidence. Decoded host-tail values use atol=1e-4 and rtol=1e-5; neither result is a COCO mAP or hardware-speed measurement. Source report ↗
05 / Choose a path

Start at your layer.

You do not need to understand the entire stack before contributing. Pick the layer closest to your interests and follow a short reading path.

A / LEARN

Study the arithmetic

Explore quantization, layouts, activations, and exact numerical semantics.

Start here →
B / REVIEW

Audit the contract

Review the v0 ABI, error behavior, and independent conformance fixtures.

Read the contract →
C / BUILD

Choose a build path

Compare reference, FPGA, and future silicon profiles without overstating readiness.

Compare profiles →
D / HELP

Contribute evidence

Add review, test vectors, workload analysis, documentation, or tool support.

Find useful work →
!

Open work, built in public.

HASLAB is an active work in progress. The reference model, simulator, and conformance foundation are ready to study today, while every future hardware capability will arrive with reproducible evidence.