Paths to a chip
From GPUs to your own ASIC.
A GPU runs anything, and you pay for that in price and power; when one settled workload runs at scale, a chip built for it does the work for less. We assess the case and design the accelerator, for AI inference and for any other settled computation.
- For
- Teams whose GPU spend or power budget is set by one workload: AI inference, or another computation repeated at scale
- You provide
- The workload as it runs on GPUs today: its throughput, the hardware it uses and what it costs
- You receive
- A feasibility report first, then a chip specified, designed and taped out for your product
- Our record
- Compute-accelerator ASICs designed for customers on 28 nm, 16 nm, 8 nm and 7 nm, delivered as GDS and built by the customers; 6 nm AI full-custom chip design
Three tests decide it
Whether a workload deserves its own chip.
Will the workload stay?
A chip fixes what it computes. The models, operators or kernels have to be stable enough to outlast the design time, or the chip has to keep the right parts programmable.
How far ahead of a GPU can a chip get?
The gain comes from dropping what the workload never uses: narrower arithmetic, memory sized to the data, and a datapath shaped to the computation. The less of the GPU you use, the larger the gain.
Does the scale carry the one-off cost?
The saving per unit of work, multiplied by the work you will run over the chip's life, has to exceed the engineering and the masks.
What the chip changes
Built around one workload.
Numerics
Number formats chosen against your accuracy targets, including reduced-precision formats such as INT8 and FP8.
Memory
On-chip SRAM and external memory planned around bandwidth, and not around capacity alone.
Compute arrays
MAC arrays, multipliers, adder trees and pipeline registers, which is where full-custom design does the most.
Software path
A compiler and runtime plan agreed before RTL starts, so the chip runs your workload from its first day.
How it runs
Five steps, the first one on paper.
Feasibility study
The three tests, on your numbers.
Architecture and performance model
A model of the chip, checked against your measured GPU results.
RTL and FPGA prototype
RTL written and verified, and proven on FPGA where the product calls for it.
Physical design
A standard-cell flow, with full-custom blocks where they pay.
Tape-out and production
Mask data, the hand-off to the foundry, then production and supply.
Questions about moving from GPUs to an ASIC
When does an ASIC beat a GPU?
When one stable workload runs at high volume. While the model or the algorithm is still changing, GPUs are the right tool; once it settles and the volume is there, a dedicated chip does the same work for less power and cost.
Is this only for AI?
No. The same reasoning holds for any computation that is fixed, heavy and repeated: inference, video and signal processing, search, simulation kernels. AI inference is the most common case today.
Do we have to give up GPUs?
No. Most teams keep GPUs for development and for workloads that still change, and move the settled, high-volume part to the chip.
Which processes do you design on?
We are a TSMC direct account, and our design work covers FinFET nodes at 16 nm, 12 nm, 7 nm and 6 nm.
Who owns the design?
Ownership of the design and its data is set in the contract for each project.
Related pages
A GPU bill set by one workload?
Tell us the workload, its volume and what it costs today.
- Fill in the project form
- We reply by email within a few working days
- Onboarding: NDA, design review and a staged plan
The form is the fastest way to reach us. A person reads every request.