Live product

FPGA Studio v0.0.28

Run your filter on a real AMD Alveo U55C card, or check your own Verilog with Lattice and Gowin synthesis and place-and-route. Use it from Claude or ChatGPT.

EARLIER RESEARCH PREVIEW (BELOW) · DEMOS AND BUILD REQUESTS
Agent-native FPGA development

Your algorithm.
On real silicon.

Bring a reference algorithm. Build toward an FPGA implementation you can inspect, test, and iterate on.

Your code. Your acceptance criteria. An inspectable result.
PRODUCTION SURFACE / DATA EXPLORERCAPTURED DATA
FRAME 20056KolektorSDD2
230 × 648 px
ANNOTATION ON
Captured industrial surface, dataset sample 20056
DATASET LABEL

A defect worth
catching.

Locate the annotated region in an actual production image.

Human-provided annotation

This overlay is ground truth,
not a model prediction.

01 / 06
The engineering job

Turn camera frames into a tested inspection pipeline. Define defect recall, line rate, and latency before building.

Current: source-data review · full detector and board validation pending

Designed for teams moving algorithms
from software to dedicated hardware.

Neural signalsMachine visionSignal processing
Explore the work

Start with the job to be done.

Inspection, imaging, and scientific instruments need more than generated code. They need a pipeline that meets a measurable acceptance test.

Vision · Drag the divider above to compare the original photograph and the rectified output. The current preview is the CPU reference; vision hardware validation is incomplete.
The workflow

Keep the algorithm.
Close the hardware loop.

The product goal: less setup and rework between a working reference and an accepted FPGA implementation.

01 / DEFINE

Bring your reference.

Start with your code, representative data, target board, and the constraints that matter.

02 / ITERATE

Build against acceptance.

Compare implementations against a frozen reference, then check the actual board output.

03 / INSPECT

Own the result.

Review editable code, build artifacts, measurements, and the records needed to reproduce the outcome.

Agent first. Human friendly.

Part of your workflow.
Not another chat window.

A run-shaped API for your engineering agent. A clear workspace for your team. Submit inputs, follow events, and inspect the result.

VERSIONED INPUTSRUN EVENTSEDITABLE OUTPUTS
API / RUN CONTRACTDESIGN PREVIEW
POST /v1/runs

{
  "project_id": "your-project",
  "bundle_id": "frozen-input-revision",
  "case_id": "vision-warp"
}

GET /v1/runs/{id}/events

The public build console on this page queues real on-card runs today.
The full run API below is a design preview.

The result should be inspectable.

Research results have different evidence levels. Each demo keeps its input source, verification scope, and limitations attached.

Back to the visual demo ↗
What am I looking at in the visual demo?

A real captured OpenCV sample photo, transformed using manually marked planar corners. The 12 CPU reference frames match the C++ simulation. This is not physical lens calibration or an FPGA readback result. The vision hardware build was stopped before completion; no vision board result is available.

Industrial images come from KolektorSDD2 production captures. The threshold component misses many annotated defects; it is not a validated inspection system.

What is verified on an actual FPGA?

Recorded-audio FIR16: three clips, 20 launches each, 200,600 positive output samples checked against the reference on an AMD U55C. Input mutation and skipped-launch controls were checked. This is a filtering component test, with no audio-quality or CPU speedup claim.

Where does the Neuropixels case stand?

The existing MATLAB/C++/HLS consistency evidence uses synthetic data. The DANDI 000034 excerpt is MEArec simulation. This website does not add a real-acquisition on-card result for the separate Neuropixels lane.

Which real industry workflows inform the product?

R.A.M. industrial inspection and Intuitive Surgical image-warping prototyping are documented vendor user cases. They inform the workload choice; these organizations are not represented as our customers or partners.

Open the current visual receipt
CPU reference preview. Hardware receipt not available.
Build console

Verify your signal on real hardware.

Choose a seed for a new input signal and verify it on an Alveo U55C using our fixed FIR-16 bitstream. Every run checks the full integer reference and negative controls. Custom HDL and new bitstream synthesis are not available in this preview.

Each request has a tracked result. Hardware availability affects queue time. Timing covers kernel dispatch and completion; PCIe transfers and readback are excluded.

Build queue

LOADING

fetching live status…

Recorded reference runs