The performance cost of PQC, bigger keys, longer handshakes, more entropy
Post-quantum algorithms cost bandwidth and memory more than CPU. Keys and signatures are substantially larger, handshakes carry more data, hybrid modes perform both exchanges, and entropy consumption rises per transaction. Measure your own workload rather than trusting a benchmark table.
This page names what to measure and the shape of each cost. It does not quote figures, because the numbers depend on parameter set, implementation, platform and workload, and a stale benchmark is worse than none.
Size, not speed
The reflexive assumption is that new cryptography is slower. For ML-KEM the computation is fast — frequently faster than the elliptic-curve operations it replaces. The cost is in how much data moves.
Public keys, ciphertexts and signatures are all substantially larger than their classical equivalents. That shows up as bandwidth, as memory in constrained devices, and as storage wherever certificates and signatures accumulate.
Handshake size and what it breaks
A hybrid TLS handshake carries both a classical and a post-quantum key exchange, which pushes the ClientHello well beyond what it used to be. Real interoperability failures have come from middleboxes and implementations that made assumptions about handshake size.
This is the cost most likely to surprise a deployment, and it is testable now on your own path rather than in theory.
Entropy per transaction
Two effects push it up. Hybrid means two key exchanges, each consuming randomness. ML-DSA in hedged mode consumes randomness per signature rather than per key.
For a high-volume signing service that changes entropy from a provisioning concern to a continuous one, and a reseed budget sized for the previous workload may not hold. Redo the arithmetic rather than assuming headroom.
What to measure on your own workload
- Handshake latency at your actual percentiles, not the mean, and across your real network paths.
- Bytes on the wire per connection, and whether that crosses any packet or buffer boundary that matters.
- Memory high-water mark on the most constrained device in scope.
- Signature verification cost on the weakest verifier, which is often forgotten in favour of signing cost.
- Entropy consumption per second at peak, including the reconnection burst.
The quiet conclusion
For most server workloads the cost is absorbable. For constrained devices and for high-volume signing it is not always, and that is where dedicated execution capacity stops being optional — the acceleration argument follows from measurement rather than from assertion.
Keep reading
- Crypto agility
What is crypto agility, and why it matters now
Crypto agility is the ability to change cryptographic algorithms without redesigning the system that uses them. It matters now because post-quantum standards are settling rather than settled: parameter sets are revised, national schemes diverge, and a scheme could be weakened by cryptanalysis.
- Crypto agility
Crypto agility in hardware, why programmable beats fixed silicon
If cryptographic algorithms are committed to fixed silicon, every standards revision becomes a board respin and a new qualification cycle. In programmable hardware the same revision is a firmware update to a card that is already deployed and already qualified.