Expedera and Cadence Pair an NPU With a DSP for Edge AI
Expedera and Cadence have demonstrated an AI subsystem that pairs Expedera’s Origin Evolution neural processing unit with the Cadence Tensilica NeuroEdge 130 AI Co-Processor, running in pre-silicon simulation. The simulation shows AI workloads distributed and executed across both IP blocks, with the aim of improving performance and power efficiency while reducing latency, cost and complexity in edge AI inference. The pairing is the interesting part, and the reason for it appears in a single clause of the Cadence quotation: the co-processor handles pre-processing, post-processing and operator fallback. That one phrase explains why a specialised accelerator benefits from a general processor beside it.
Why an NPU needs a co-processor
A neural processing unit is efficient because it does one thing. It executes the operations its designers built it for, matrix multiplication and convolution and the standard activations, at high throughput and low power, and that focus is exactly why it is fast.
The same specialisation is a limitation. A model contains operations outside that set. Some are mundane: resizing an image before inference, normalising pixel values, decoding a video frame. Others change as research moves, such as a new activation function, an attention variant or an operator that a model published last month introduced. An NPU designed two years ago does not know them.
Operator fallback is the mechanism that handles this. When the NPU meets an operation it cannot execute, the work passes to the co-processor, which is a DSP and therefore general enough to run it. The model still runs end to end and only the unsupported fraction runs slower. That is why the pairing matters more than either block alone, and it is the difference between an accelerator that runs the models its designers targeted and a subsystem that runs the models a customer actually has.
The packet-based architecture
Expedera’s architectural claim is that its NPU uses packet-based scheduling rather than the layer-based approach its competitors take, and the release states the benefit in one sentence: the patented architecture eliminates the inefficiencies of conventional layer-based designs, delivering superior utilization without sacrificing model accuracy.
The distinction is worth unpacking because it is the company’s core differentiator. A layer-based NPU processes a neural network one layer at a time, so the processing elements wait at each boundary for the previous layer to finish. Where a layer is small the array sits partly idle, and where a layer’s shape does not match the array dimensions part of the array goes unused.
A packet-based design breaks the work into packets that a scheduler can distribute across the array independently of layer boundaries, so a small layer can sit alongside work from another part of the network. That is what the word utilization refers to, and it explains why the claim concerns efficiency rather than peak throughput. Peak numbers assume every array element is busy; utilization describes how often that is true.
Why the edge is where this matters
The shift the release describes is real and is the reason the collaboration exists. AI inference is moving from cloud to edge because latency, bandwidth and privacy all argue for processing near the data rather than in a distant data centre.
That move changes what an accelerator has to do. A cloud inference chip optimises for throughput per dollar in a rack with abundant power and cooling. An edge device optimises for efficiency per watt inside a phone, a car or a camera, where the thermal budget amounts to single watts and the product must run for years on a battery. Specialisation pays off far more at the edge, which is why an NPU architecture tuned for utilization has a natural market there.
What a reference design is worth
A reference design is the currency of the IP business, and this release reads as one. A company selling an NPU core does not ship a product; it ships a block that a customer integrates into their own system on chip. The customer’s risk is that the block works in isolation and not alongside the rest of their design.
A validated pairing removes part of that risk. If Expedera and Cadence can show the NPU and the co-processor distributing work between them as intended, an SoC team starts from a known-good configuration rather than from two datasheets and an assumption. For an edge AI programme that is worth more than a better specification, because it shortens the schedule, and the schedule is where chip projects fail.
The Cadence side of the arrangement is worth stating too, because it is not one-directional. Tensilica DSPs are widely licensed, and a DSP on its own is a general processor that needs a compelling reason to sit in an AI design. Pairing with an NPU supplies that reason: the DSP becomes the component that handles what the accelerator cannot, which makes it easier to justify in a bill of materials and harder to design out.
What pre-silicon means here
The qualifier in the release deserves attention. This is a pre-silicon simulation, which places the work at a specific point in the development cycle.
Pre-silicon means the design exists as a model rather than as manufactured silicon, and the simulation demonstrates that the two IP blocks can distribute and execute workloads together as intended. That is a genuine and necessary milestone: it is the point at which a customer can see the pairing works before either party commits to a chip.
It is also the beginning of a programme rather than the end of one. No silicon has taped out, nobody has measured performance on hardware, and the release gives no figures from the simulation itself. What it establishes is feasibility and a reference design, which is exactly what an IP customer needs in order to begin their own integration with confidence.
Ten million devices and an award
The most concrete evidence in the release is a deployment figure: Expedera’s Origin Evolution NPU solutions ship in more than 10 million consumer devices. For a semiconductor IP company that number is the real proof of concept, because IP ships inside someone else’s product and adoption is the only measure that matters.
The company also holds a third-party validation worth noting. The Origin Evolution NPU IP took the Best Edge AI Processor IP award in the 2026 Edge AI and Vision Product of the Year Awards, which is an assessment made outside the company rather than a self-description.
Together those two facts do more for the case than any architectural claim, because they describe adoption and external judgement rather than intent.
The claim about the vendor’s own engineering
One phrase in the Expedera quotation deserves a separate note. Sharad Chole says that with models running on the combined subsystem and agentic AI workflows helping the company deliver customized IP more quickly, customers have a faster and lower-risk path to production.
That describes the vendor using AI in its own engineering rather than in the product it sells, and it points at a specific operational problem. Delivering customizable IP quickly means each customer wants a variant tuned to their model mix and power envelope, and the configuration work is where schedules slip. If agentic workflows genuinely shorten that cycle, the benefit reaches the customer as a shorter integration. That is a more interesting claim than it first appears, and it is the one I would most like to see developed: a figure for how much shorter the integration cycle becomes would be worth more to a customer than any specification.
The figures that would complete it
The one thing a reader will want next is performance data. The release says the combined subsystem improves performance and power efficiency and reduces latency, and gives no figure for any of the three: no throughput, no efficiency in operations per watt, no latency in milliseconds and no utilization percentage for either architecture.
Those numbers would let an engineer evaluate the design rather than understand it. They would also make the packet-based claim checkable, because utilization is precisely the metric where a packet-based scheduler should show an advantage over a layer-based one.
The natural next publication is a benchmark from the simulation, and the release’s own logic points at it. If packet-based scheduling delivers better utilization, the utilization figure is the number that proves it, and publishing it would convert a well-argued architecture into a demonstrated one.
Why the timing matters
ICCAD-Expo runs in Beijing on 19 and 20 November and Expedera will be at booth F091, which means the collaboration is going to the design community directly rather than being announced into a vacuum.
For a company selling semiconductor IP that is the right venue. The people who choose an NPU are the architects and integration engineers who attend these events, and a pre-silicon demonstration with a named partner is exactly the artefact that starts a technical conversation with them.

What to watch
Whether Expedera and Cadence publish simulation results, because that is what turns a validated architecture into an evaluated one. Whether the pairing extends to the other Tensilica DSPs the release says the approach can reach, since a family of validated subsystems is more useful to a customer than a single reference. And, whether silicon follows, because a pre-silicon demonstration opens a programme rather than closing it. And whether the packet-based approach shows its advantage at the model sizes edge devices actually run, which is the question that decides how much the architecture matters in practice.

