Skip to content

FPGA or ASIC for cryptographic acceleration

An FPGA is reconfigurable after manufacture; an ASIC is fixed. For cryptography the trade is agility against efficiency, and the deciding factor is volume against the probability that the algorithms will change. During a standards transition, that probability is not low.

The four axes

  • Non-recurring cost. An ASIC carries substantial design and mask cost before the first unit. An FPGA design carries engineering cost and no mask cost.
  • Unit cost. At volume the ASIC is far cheaper per unit. The crossover depends on the design, and below it the FPGA wins on total cost.
  • Efficiency. An ASIC is faster and lower power for the same function — reconfigurability is not free in area or in switching energy.
  • Agility. An FPGA design can change after deployment. An ASIC cannot, at any price short of a respin.

Time to market

Often the deciding factor and rarely the one on the comparison table. An FPGA implementation can be in a customer's hands while an ASIC is still in design, and during a standards transition the window for a design win is set by when equipment gets opened, not by when silicon is optimal.

What tips it for post-quantum work

The change probability. FIPS 203, 204 and 205 are final, and that is not the same as settled: parameter sets can be revised, national schemes diverge, hybrid constructions will change, and cryptanalysis continues against young assumptions.

An ASIC implementing a specific parameter set is a bet that nothing changes for the life of the device. For a ten-year appliance in a regulated estate that is a long bet — the full argument.

Where the ASIC still wins clearly

High volume, cost-dominated, power-constrained, and implementing something stable. AES is a good example: the algorithm is not changing, and a fixed implementation is the right answer.

The conclusion should follow your volume and your change probability rather than any vendor's roadmap. A design shipping tens of millions of units with a stable algorithm should not be talked into programmability, and an appliance estate in the middle of a migration should not be talked into fixed silicon.