Skip to content

Cryptographic acceleration: hardware for faster, more efficient security

Cryptographic acceleration moves cryptographic work off the application CPU onto hardware built for it. It has usually been sold on throughput, but the arguments that matter now are trust boundary, energy per operation, and the fact that post-quantum algorithms cost more per transaction than the ones they replace.

22 articles in 5 sections

Cryptography has always consumed compute. What changes the conversation is that the per-operation cost is going up for the first time in years, at the same moment that estates are being re-keyed and power budgets are fixed.

Three arguments, only one of which is speed

Contention. Cryptographic work on the application CPU competes with the application. Under load the security work is what gets squeezed, which is exactly backwards. The argument in full.

Boundary. Work done inside a hardware boundary is work whose key material never reaches host memory. That is a security property rather than a performance one, and it is the reason offload survives even when the CPU is fast enough.

Energy. Fixed power envelopes make performance per watt the binding constraint in data centres and at the edge, and post-quantum operations cost more per transaction than their predecessors. That trend only moves one way.

Where this pillar connects

Agility in hardware and acceleration turn out to be the same design decision seen from two directions, because both require cryptography to be executed by something that can be changed.

A cryptographic boundary that reports on itself is the same idea as an entropy source that reports on itself, applied to more of the system — which is where verification stops being about one component.

Where it leads

At the RPU: entropy generation, runtime verification and cryptographic execution inside one boundary, in one slot, in one power envelope.

Infrastructure and efficiency

Where cryptographic load accumulates across infrastructure, and why performance per watt is becoming the binding constraint.

Entropy and execution together

The Randomness Processing Unit, hardware that reports on itself, and what putting generation and execution in one boundary buys.