Start with certainty, not a sales pitch.
Our evaluation process gives you projected performance data before any commitment. Four stages, approximately 6–8 weeks, NDA-first.
The evaluation process.
From initial model submission to pilot decision in six to eight weeks.
Share your frozen model graph in ONNX or TorchScript format. We execute an NDA before receiving any model data. Your workload structure is treated as confidential engineering information.
Our team produces a workload analysis report: op type distribution, tensor dimensions, memory working set estimate, data dependency graph. You receive the full report regardless of whether you proceed.
We run your model profile through Procunit's silicon characterization model and produce a performance projection: throughput-per-watt estimate, latency floor, batch efficiency curves. Presented with methodology and confidence ranges.
Based on the projection, we jointly decide: access our pre-release evaluation board (shared hardware, your model loaded), or proceed to a low-volume custom pilot run with a timeline scoped to your workload.
Tell us about your workload.
Fill out the form and our engineering team will respond within two business days to schedule an initial consultation.
Common questions.
Procunit's compiler currently accepts ONNX (opset 17 and below) and TorchScript frozen models. PyTorch export via torch.onnx.export is the most common path. If your model uses custom ops or a different serialization format, contact our engineering team to discuss compatibility.
The full evaluation — from model submission to pilot decision — takes approximately 6–8 weeks. Graph analysis is typically complete within 2 weeks of receiving your model. Performance projection follows in 2–3 weeks. The final discussion meeting is scheduled at your convenience.
Yes. We execute a mutual NDA before any model data is transmitted. Your model architecture, operator graph, and weight tensors constitute proprietary engineering information. Procunit's evaluation process is confidential by default — we do not reference customer workload details in any public communications.
A pilot run is a low-volume initial silicon tape-out optimized for your specific model graph. It produces a small quantity of functional parts for performance validation and integration testing. Timelines and scope are determined jointly based on the evaluation results. An alternative is our pre-release evaluation board — shared hardware where your model is loaded and run against the projection numbers.
Procunit is in the evaluation-silicon phase. We are not announcing production availability dates. Production silicon timelines for ASIC development are model-specific — they depend on the complexity of the compute graph, the process node, and the foundry schedule. What we can commit to: teams that complete the evaluation process are first in the pilot allocation queue, and characterization data gathered during evaluation directly informs the RTL synthesis target for production silicon. The case for starting an evaluation now is that six months of workload stability data is worth more than six months of waiting.