Products
AI accelerators shaped around your workload.
A general-purpose chip spends power on work your model never asks for. A custom accelerator does the same job within a smaller budget, and we design them to order.
- You bring
- Your models, target throughput, power budget and cost
- You get
- A chip specified, designed and taped out for your product
- Process nodes
- TSMC direct account; FinFET design at 16 nm, 12 nm, 7 nm and 6 nm
What goes into the chip
Built from the workload up.
Workload analysis
Profiling your models to find the operators, data types and memory traffic that matter.
Architecture
Compute arrays, on-chip memory and interfaces sized to the workload, not to a benchmark.
Numerics
Integer and reduced-precision formats chosen against your accuracy targets.
Memory system
SRAM banks and external memory planned around bandwidth, not capacity alone.
Interfaces
Host and sensor interfaces matched to the system the chip goes into.
Software path
A compiler and runtime plan agreed before RTL starts, so the chip is usable on day one.
Questions about custom AI silicon
When does a custom accelerator pay off?
When volumes are high enough, or power and cost limits tight enough, that a general-purpose chip leaves too much on the table. We work through that trade-off with you before any design starts.
How do you supply AI chips?
By project: every chip is designed for one customer's product.
Who owns the design?
Ownership of the design and its data is set in the contract for each project.
Have a model and a power budget?
Send them with your volumes and schedule, and we will reply.