PQC hardware acceleration, ML-KEM and ML-DSA in silicon
Post-quantum algorithms spend their time differently from classical ones. ML-KEM and ML-DSA are dominated by polynomial arithmetic, sampling and hashing rather than modular exponentiation, so hardware that accelerated RSA and elliptic curve accelerates the wrong things.
What is actually expensive
- Number-theoretic transforms. Polynomial multiplication in the ring is the core operation, done via NTT. Highly parallel, regular, and well suited to dedicated hardware.
- Sampling. Generating polynomials from distributions — uniform, and centred binomial for errors. Consumes randomness and involves rejection sampling with data-dependent behaviour.
- Hashing and extendable output. SHAKE and SHA-3 appear throughout, for expansion and derivation. A surprising fraction of total time is hashing.
- Large data movement. Keys, ciphertexts and signatures are large enough that memory bandwidth becomes a real constraint rather than an afterthought.
Why existing accelerators do not help
Hardware built for classical public-key cryptography accelerates big-integer modular arithmetic. Post-quantum lattice schemes barely use it.
So an appliance with a strong RSA and elliptic-curve accelerator can find that its acceleration contributes almost nothing to post-quantum operations, and the work falls back to the general-purpose CPU. This is a real and under-appreciated consequence of the migration, and it is the sort of thing an inventory should be catching.
What dedicated hardware can do
NTT is the clearest target: regular, parallel and repeated. Hashing benefits from a dedicated Keccak core, which is well-understood hardware. Sampling benefits from a fast entropy path feeding it directly rather than through the host.
That last one is the connection this pillar cares about. Sampling consumes randomness continuously, and an accelerator whose entropy arrives across a host boundary has a bottleneck and an exposure that one with a local source does not — which is the platform argument.
What a realistic speed-up looks like
Shaped by Amdahl rather than by the accelerator. If hashing is a third of the time and you accelerate only NTT, the ceiling is set by what you left behind. Useful designs accelerate the whole operation rather than its most interesting part.
Measure where your own time goes before choosing what to accelerate. The split differs by parameter set, and by whether the workload is dominated by key generation, encapsulation or signing.