Summary
Migrate the part of the system that the FPGA is bad at, and keep the part it is good at
An intelligent RF system, one that senses, decides and reconfigures itself in real time, almost always starts life on an FPGA. It should. The architecture is uncertain, the algorithms are changing and a tape-out is not a sensible way to find out whether an idea works.
The question is what happens when the FPGA prototype has proved the architecture. The usual framing is "FPGA or ASIC", and it is the wrong one. The right question is which functions move to silicon, in what order, and on what evidence. This article sets out how we answer it.
The gap
The FPGA penalty is real, measured, and unevenly distributed
Kuon and Rose measured the gap between FPGAs and standard-cell ASICs on the same process in 2007. For logic-only designs the FPGA needed about 35 times the silicon area, ran at a third to a quarter of the speed and consumed about 14 times the dynamic power. Hard blocks such as multipliers and memories narrowed the area gap substantially, to roughly 18 times when they were used. Process has moved on since, and Boutros and Betz's 2021 survey of FPGA architecture describes how much of the FPGA's progress has come from hardening more of the common functions, but the shape of the result has not changed: the penalty is largest for arbitrary logic and smallest for functions the FPGA vendor has already hardened.
Two further observations matter for RF systems. First, the FPGA cannot do analogue. Everything in the signal chain before the ADC and after the DAC is already off-chip, in discrete or custom parts, however good the FPGA. Second, where the vendor has integrated converters, as AMD did on the Zynq UltraScale+ RFSoC with 14-bit RF-ADCs to 5.9 GSPS and RF-DACs to 10 GSPS, the integration is excellent and fixed: you get the converters they chose, at the power they consume, in the package they ship.
For an RF product the penalty shows up as heat in a sealed enclosure and as a bill of materials that grows with every element. A radio unit on a mast or a surface on a wall has no fan and no spare board area. The same design that is a comfortable laboratory prototype on a development board becomes a thermal and mechanical problem at the point of deployment, and that is usually what starts the migration conversation.
What moves
Four candidates, in the order they usually earn their place
The analogue and mixed-signal chain
Bias generation, data conversion, gain control, frequency generation, calibration. These are already off the FPGA, usually as many discrete parts. For a system with hundreds of controlled elements, the discrete version does not scale in size, power or cost, and it is the first thing a custom die should absorb.
The fixed, high-rate datapath
Filtering, decimation, correlation, detection: the parts of the signal processing that have stopped changing and run at the sample rate. This is where the FPGA area and power penalty is paid every cycle, and where hardening buys most.
The secure control plane
Key storage, cryptographic engines, the root of trust and the compartment boundaries around configuration. These benefit from silicon because the security argument is stronger when the boundary is a property of the die rather than of a bitstream that can be reloaded.
The algorithms, last or never
Beam optimisation, inference, tracking, anything that is still being improved. These stay in fabric until they are stable, and for many systems they stay there for the life of the product, either on a companion FPGA or as embedded programmable logic on the ASIC itself.
What it costs
Prototype silicon is affordable. Production silicon is a business decision
The cost of a first mixed-signal test chip is lower than most teams assume. Europractice publishes its multi-project wafer prices. For 2026 it lists a 22 nm FD-SOI run at 16,200 to 17,820 euros per square millimetre with a four square millimetre minimum, a 28 nm run at 11,300 to 12,430 euros per square millimetre, and a 55 nm BCD process at 4,700 to 5,170 euros per square millimetre with a nine square millimetre minimum. A small test chip on a shared shuttle is a six-figure exercise in fabrication alone, before design, IP, packaging and test, and the design effort is the larger number.
Production is a different question. A dedicated mask set at an advanced node, qualification, and the volume commitment that makes unit cost attractive are decisions about the market, not the technology. Intermediate routes exist: structured ASICs such as Intel's eASIC family, which the company launched in 2020 claiming up to 50 percent lower core power than an FPGA with lower non-recurring engineering than a cell-based ASIC, and embedded FPGA blocks that keep a programmable region on an otherwise hardened die.
The honest framing for a board paper is this. Migration is justified when the FPGA system has proved the architecture, when the volume or the envelope makes the FPGA penalty a product problem rather than an engineering irritation, and when the functions being moved have stopped changing. If any of the three is missing, the answer is to stay on the FPGA and say why.
The evidence
Decide from measurements of the prototype, not estimates of the ASIC
Measured FPGA resource and power by functionNot the whole bitstream. Each candidate function on its own, so the saving from hardening it can be stated.
The analogue bill of materials and its board areaWhat the discrete mixed-signal chain actually costs in parts, power and square centimetres at the element count the product needs, extrapolated honestly.
Stability of each functionHow many times it changed in the last six months of prototype development. A function that is still moving is not ready for silicon.
A floorplan and an area estimate from the chosen processUsing the foundry's design kit, not a scaling rule from a paper.
A verification plan that reuses the FPGA test benchesThe prototype has already generated the stimulus and expected results. The same vectors should exercise the RTL destined for silicon, and the FPGA should remain the golden reference for the ASIC bring-up.
What this does not settle
A migration study is a costed route, not a guarantee
The study tells you which functions to move, what they are expected to cost and save, and what the risks are. It does not tell you the chip will work first time, and no study can promise that. Mixed-signal silicon in particular rewards a test-chip step between the FPGA and the product die, and we plan one in.
This is the route we plan for our own programmable RF work: FPGA prototype first, then a mixed-signal test chip designed to tape-out, then independent evaluation. The numbers in this article are from the cited literature and price lists.
Next step
Have an FPGA prototype that has proved its point?
Talk to an engineer
Start a conversation
Bring the resource reports, the power measurements and the bill of materials. We will tell you what should move, what should stay, and what it would cost to find out properly.
The capability
FPGA prototyping and pre-silicon validation
Prove the architecture before you fund the silicon.
