Backends at a glance

Backend	Implementation	Relative speed	When to use
vortex-pqc-native	C extension (compiled)	~10× faster	Production, benchmarks, services
vortex-pqc-pure-python	Python reference (schoolbook mul)	Baseline	CI without compiler, education, correctness checks

Running benchmarks

from vortex_pqc import benchmark_throughput, native_backend

print(f"Backend: {native_backend()}")
results = benchmark_throughput(operations=50)

for op, stats in results.items():
    ops = stats["mean_ops"]
    ci  = stats["confidence_interval"]
    print(f"{op:8s}  {ops:8,.0f} ops/s  (± {ci:,.0f})")

Example output (native, Apple Silicon — your numbers will vary):

Backend: vortex-pqc-native-aarch64
keygen      1,200 ops/s  (± 50)
encaps      1,100 ops/s  (± 45)
decaps      1,150 ops/s  (± 48)

Where time is spent

Operation	Dominant cost	Notes
Keygen	1× poly mul per rotation + XOF	1 XOF call (vs 4 in Kyber-512)
Encaps	K poly muls + hashing	Fresh randomness each call
Decaps	K poly muls + re-encryption	FO check adds one full encaps path

VORTEX vs Kyber-512 efficiency

Metric	Kyber-512	VORTEX-256	Advantage
Key expansion XOF calls	4	1	VORTEX
Frobenius rotations	N/A	K−1 permutations	VORTEX (free vs XOF)
Private key size	1 632 B	1 248 B	VORTEX
Poly mul count (encaps)	Similar	Similar	Tie
NTT optimisation	Available	Planned	Kyber (today)

Optimisation roadmap

Optimisation	Status	Expected gain
Schoolbook poly mul (C)	✅ Shipped	Baseline native
NTT-based poly mul	🔜 Planned	5–10× on poly operations
AVX2 / NEON SIMD	🔜 Planned	2–4× on NTT
ARM64 -mcpu=native	✅ Shipped	Platform-specific tuning

Tuning for your deployment

Maximise throughput (servers)

✓  Install with C compiler present (native backend)
✓  Use pip install without --no-binary
✓  Pre-generate key pairs (keygen is amortised)
✓  Batch encapsulations if your protocol allows

Minimise binary size (embedded)

✓  Link only libvortex_pqc.a (no Python)
✓  Strip symbols: cc ... -s
✓  Pre-provision keys offline (skip keygen on device)

CI without a compiler

✓  Pure Python backend works everywhere
✓  All 26 tests pass without _native.so
✓  Slower but cryptographically identical

C benchmark

cd c
make lib
# Optional: build and run bench test if available
make -C c test