sequenced.ai
Articles/Models & infrastructure/Blueprint//7 min read

Quadric makes the programmable part of an AI chip a design choice

Quadric licenses an inference processor and the tools to program it. An edge-chip team can test an entire vision pipeline before committing to silicon.

By Sequenced deskAI-assisted, source-led · how we work
Visit Quadric website ↗
ChimeraProcessor IPA programmable core for integration into a chip.
ONNXModel inputThe graph compiler accepts exported networks.
C++Custom operatorsDevelopers can implement additional kernels.
SimulatorEvaluation toolProfile a proposed configuration before tape-out.
Quadric mark
Quadricquadric.ai · independent research

Represent this company? Verify your work email to access its workspace, or send the desk a factual correction.

The hard part of choosing an AI accelerator for a camera can be the few operations on either side of the network. A sensor delivers an image in one layout, a model expects another, and its output needs turning into an event the camera can use. Quadric licenses the design of a processor intended to run that wider workload. Its Chimera proposition asks a chip designer to evaluate a program from input to output before deciding what to put in silicon.

In brief
  1. 01The product Processor intellectual property, verification material and a software development kit for embedded inference.
  2. 02The tradeoff A programmable core can absorb new work in software, provided that the deployed chip still has enough memory, bandwidth and time.
  3. 03The evaluation Compile an actual model and its surrounding code. Compare output quality and execution estimates with the device’s requirements.

01 / ProductWhat a licensee receives and what still has to be built

Quadric’s licensing FAQ describes a package of register-transfer level design material, verification IP and the Chimera SDK. A semiconductor team uses these deliverables when integrating the processor into a system on chip. A camera maker that buys finished modules is making a different purchase. It needs to identify a shipping chip or board, its device interfaces and the software its supplier will support. The presence of an IP license offer alone cannot answer those questions.

The architecture description puts scalar, vector and matrix operations on a programmable core. Matrix work is central to much neural inference. The remaining operations matter when a pipeline includes filtering, indexing, control decisions or a custom calculation. Bringing more of that work onto one processor can avoid a handoff to a different engine. How much this helps depends on the amount of data crossing each boundary, the selected core configuration and the way the model is compiled.

The accompanying SDK exposes a route from an ONNX model to a compiled program and a simulated execution. It also provides a way to write custom operators in C++. This matters over the life of a device. A firmware team may be able to implement a new postprocessing step after the chip design is fixed. A software update still has to fit the resources that were physically built. The right design question is how much room the product roadmap needs.

02 / AudienceA choice for teams that can influence the silicon

Consider a manufacturer designing a vision processor for successive generations of inspection cameras. It anticipates a second inspection model before a new chip would normally ship. A programmable inference core gives the manufacturer something concrete to evaluate. It can compile both the current model and a candidate successor, then test how much of the imaging pipeline can be shared. If the next product changes camera resolution as well as its model, include both changes in the experiment.

A team delivering a few dozen cameras against a short deadline faces different economics. It may get further by testing a purchasable module, where board bring-up, power delivery and drivers are already part of the product. Hailo’s approach to edge inference offers that adjacent decision. Verify the model against the module just as carefully. The work being bought includes hardware the team can deploy, while the Quadric route here begins upstream with the chip design.

03 / WorkflowFollow a single inspection image all the way to the reject signal

For a proposed internal evaluation, take a representative frame from a line inspecting printed packaging. The edge device must locate a label, decide whether its print is acceptable and send a decision before the package leaves the inspection point. Write down that deadline in elapsed time from camera capture. Include frames with glare, rotated labels and nearly acceptable print. The number of images a classifier can process in isolation is a poor proxy for whether the actual line rejects the right packages in time.

Save the floating-point model’s outputs as a reference. The SDK describes importing ONNX, quantizing with representative calibration data, compiling and running the result through its instruction-set simulator. Compare decisions as well as numerical outputs. A small change close to a reject threshold may alter the production decision. Use separate images for final comparison so the same few calibration examples are not also deciding whether the accelerator is good enough.

Next include the conversion from the sensor’s pixel layout and the logic that finds the relevant label region. Check whether each operation is captured in the compiled graph, implemented as a custom kernel or performed elsewhere on the proposed chip. For each handoff, account for the data that must be written and read again. The team can then explain which part of an end-to-end deadline is spent computing, moving an image or waiting for another component.

The multicore documentation distinguishes distributing independent work from dividing work inside a model. Two simultaneous camera streams might use separate cores. A single larger frame may demand cooperation between cores. These are different design choices even if both consume the same number of cores. Profile the stream that produces the latest acceptable decision and the combination that creates the most contention. Label simulator results as estimates until a representative system confirms the timing.

04 / PricingWhat the evaluation license permits before chip integration

StageDocumented routeWhat to settle
Internal evaluationSDK terms allow noncommercial internal evaluation of a possible licenseAuthorized evaluators and handling of results
Chip integrationIP license covers RTL, verification IP and SDKConfiguration, engineering support and commercial terms
External results or product shipmentThe standard evaluation rights are limitedObtain the necessary rights before distribution

Commercial scope reviewed 19 November 2026. SDK evaluation agreement and IP licensing FAQ.

The SDK agreement limits the downloadable material to internal evaluation of a potential commercial license. It also restricts generating benchmark results for publication or transmission outside the organization. Plan the proposed packaging study as an internal engineering exercise. A consultancy delivering results to an external client should clarify permission first. The public agreement does not itself grant the rights needed to ship the design in a sellable product.

Quadric describes support through tape-out in its IP FAQ. The cited pages do not give a universal dollar price or royalty schedule. Obtain a proposal that names the implementation, permitted uses, support coverage and rights for later product variants. A license for a first design can be commercially attractive yet a poor match for a buyer planning several chips if follow-on rights require another negotiation. Treat those terms as inputs to the cost comparison.

05 / DistinctionsA programmable inference core gives the compiler a larger role

Arm’s processor-IP model is a useful reference for understanding why a buyer evaluates software alongside the hardware design. The interface between code and silicon will outlive the initial demonstration. In Chimera’s case, one revealing test is to add an operator the original network did not use. Record how a developer implements it, whether the compiler integrates it, what evidence explains a slowdown and who maintains that implementation after delivery.

A named example establishes the level of engagement that public evidence supports. Quadric announced that TIER IV licensed the SDK to evaluate and optimize future Autoware work. That is meaningful evidence of an engineering evaluation. It leaves open which hardware a future vehicle will contain and when it might ship. Readers comparing design wins should distinguish a software evaluation agreement from a production chip program with committed volume.

06 / QuestionsWhere does an estimate stop being an answer?

Estimates from a core simulator are useful during architecture selection. A finished camera also has shared memory, interfaces, a thermal envelope and competing work. Ask which of these are represented in a particular simulation and which belong to the integration team’s system model. A favorable estimate on an otherwise idle core may leave the most important uncertainty untouched if image capture and inference compete for the same memory bus.

Ask the person responsible for future firmware to reproduce one change without the original evaluation engineer. Give them the approved tools, the reference output and the performance budget, then have them explain a regression. This measures a practical element of programmability that a specification cannot settle. The manufacturer needs people who can use the additional flexibility over several product cycles, especially after the team that chose the IP has moved on.

07 / DecisionChoose the next commitment the evidence warrants

01

Designing a new system on chip

Evaluate one full pipeline and one plausible successor. Keep correctness and timing results together so a favorable cycle estimate cannot conceal a changed classification.

Request an IP proposal for the configuration you actually evaluated.
02

Choosing a board for an imminent deployment

Identify an available hardware product, its supported runtime and its interfaces. An IP specification leaves purchasing and integration questions open.

Start with a shipping platform that can meet the project schedule.
03

Publishing or sharing evaluation results

Check the SDK agreement and any separately negotiated permissions before distributing materials or measurements to customers, partners or conference audiences.

Secure rights for the intended audience before the study begins.
What should we explore next?

A business worth understanding.

Suggest your business or one you find interesting. Tell us what you want to understand about its product, positioning, design or workflows.

Suggestions are free. Selection and publication stay with the desk.

Sources

Continue reading

All in this category