Benchmarked, Not Marketed
The numbers, as measured.
Every figure on this page comes from a dated benchmark run on hardware we can name. Nothing extrapolated, nothing aggregated across cores without saying so. When a new run supersedes an old number, the page changes — not before.
Measurement basis
Every cryptographic figure is the statistical mean of 100 timed samples per operation, compiled with full optimizations, on a single core of a commodity AMD Ryzen 9 9950X workstation. Signature, key-exchange, and key-hierarchy figures measured 2026-07-27; hashing and symmetric encryption 2026-05-07. Mesh and edge-node figures come from an instrumented live circuit — not a simulation — measured 2026-04-02.
Standing qualifier: every per-second figure is a single-core, per-operation cost — not system capacity. Multiply by cores for aggregate throughput, and say so when you do.
Digital Signatures · NIST FIPS 204
ML-DSA-65
The algorithm that signs every agent credential, tool invocation, and delegation approval in the Aethyr stack. Verification is the hot path — it runs on every credential check — and at 9,671 verifications per second per core, the crypto is never the bottleneck.
| Operation | Mean time | Throughput (per core) |
|---|---|---|
| Key generation | 164.60 µs | 6,075 /sec |
| Sign | 451.58 µs | 2,214 /sec |
| Verify | 103.40 µs | 9,671 /sec |
Measured 2026-07-27 · single core
Key sizes: 1,952 B public key · 3,309 B signature · 32 B seed.
Key Exchange · NIST FIPS 203
ML-KEM-768
Establishes encrypted sessions between nodes without trusting a third party. The full exchange — keygen, encapsulate, decapsulate — completes in 90.93 microseconds: the total cost of a quantum-resistant channel between two agents.
| Operation | Mean time | Throughput (per core) |
|---|---|---|
| Key generation | 23.66 µs | 42,261 /sec |
| Encapsulate | 22.66 µs | 44,139 /sec |
| Decapsulate | 44.61 µs | 22,418 /sec |
| Full key exchange | 90.93 µs | — |
Measured 2026-07-27 · single core
Hashing & Key Derivation
BLAKE3 and the key hierarchy
Every agent DID is BLAKE3 of the ML-DSA-65 verifying key, and every derived key in the hierarchy uses BLAKE3-KDF. At millions of derivations per second, the identity layer adds no perceptible overhead.
| Operation | Mean time | Throughput (per core) |
|---|---|---|
| BLAKE3 hash (1 KiB) | 0.98 µs | 1,016,356 /sec |
| BLAKE3 hash (4 KiB) | 1.08 µs | 923,127 /sec |
| Derive identity key | 0.16 µs | 6,257,128 /sec |
| Derive session key | 1.12 µs | 893,062 /sec |
Symmetric Encryption
XChaCha20-Poly1305
Payload encryption for the AethyrWire Protocol and credential backup. Block size determines throughput — small frames pay fixed per-operation overhead, bulk encryption runs at memory-bandwidth-class speed.
| Operation | Mean time | Throughput (per core) |
|---|---|---|
| Encrypt 1 KiB | 1.63 µs | 614,653 /sec |
| Encrypt 16 KiB | 10.57 µs | 94,578 /sec |
| Decrypt 16 KiB | 10.49 µs | 95,340 /sec |
| Encrypt 1 MiB | 638.17 µs | 1.64 GB/s |
Measured 2026-05-07 · single core
Live Mesh Circuit
Agent teleportation
One agent, moved around a full three-hop circuit across two CPU architectures, with every hop ML-DSA-65 signed and verified. Migration bundle is roughly 5–10 KB — the agent's memory, personality, and skills travel with it.
| Hop | Route | Time |
|---|---|---|
| 1 | ws → jetson | 140.8 ms |
| 2 | jetson → core | 221.1 ms |
| 3 | core → ws | 54.7 ms |
| Full circuit | 416.6 ms | |
| Average per hop | 138.9 ms | |
Instrumented live circuit · 2026-04-02 · agent registered and verified at every hop
Edge Node · ESP32-S3
Cold boot to post-quantum session: 2.1 seconds
A $5 microcontroller goes from power-on to an ML-KEM-768-secured session in 2.1 seconds, in an 820 KB firmware. This is the one benchmark you can reproduce without asking us anything — the firmware is open source.
| Milestone | Elapsed |
|---|---|
| CPU start | 273 ms |
| ML-KEM-768 keypair generated | 442 ms |
| WiFi connected (direct AP) | 774 ms |
| TCP connected to upstream node | 1,887 ms |
| Session key derived | 2,116 ms |
| First encrypted frame | 2,121 ms |
Instrumented boot · 2026-04-02 · direct AP topology, no router in the path
Flash it yourself — github.com/aethyrai/esp32-awp-edgePerspective
The "too slow" myth
Post-quantum crypto is three to four orders of magnitude faster than the LLM call it protects. The overhead argument was valid for early PQC implementations; it is not valid for optimized implementations on modern hardware.
| Operation | Time |
|---|---|
| LLM inference (4B model, simple query) | 500–2,000 ms |
| HTTP round-trip to external API | 50–500 ms |
| ML-DSA-65 verify (credential check) | 0.10 ms |
| ML-KEM-768 full key exchange | 0.09 ms |
Provenance
Methodology, standards, and what's open
The post-quantum primitives come from audited, optimized open-source implementations of the NIST standards — we didn't reinvent the math, and won't claim to have. What Aethyr built on top is the part that makes them usable for agents: hierarchical key derivation, session management, memory that zeroes itself when keys are dropped, and cross-implementation verification so every platform derives identical keys from identical seeds.
| Standard | Status |
|---|---|
| NIST FIPS 204 (ML-DSA-65) | Implemented |
| NIST FIPS 203 (ML-KEM-768) | Implemented |
| CNSA 2.0 | Aligned (ML-KEM / ML-DSA); CMVP validation not yet obtained |
| W3C DID v1.0 | Implemented |
| W3C VC 2.0 | Implemented |
Server crypto
100 timed samples per operation, fully optimized builds, single core. Run dates shown on every table.
Edge node
Open source firmware. Flash an ESP32-S3 and reproduce the boot timeline independently.
Mesh circuit
Instrumented live three-node circuit, every hop signed and verified. Not a simulation.
Prefer the narrative version? Read the benchmark write-up.
Numbers you can hold us to.
CNSA 2.0 sets a January 2027 procurement deadline for post-quantum cryptography. The measurements above are what running it today actually costs.