FHE Performance Myths Debunked
Real production numbers versus outdated academic assumptions about homomorphic encryption speed
The conventional wisdom about FHE is that it is too slow for production, a research curiosity with no practical deployment. This wisdom is wrong by a factor of a million. The gap between what people believe about FHE performance and what it actually achieves is the largest perception-reality gap in modern cryptography.
H33 processes 2,293,766 authentications per second. Each completes in 38 microseconds. The pipeline includes BFV homomorphic encryption, STARK zero-knowledge proof verification, and three-family post-quantum signing. This is sustained throughput on AWS Graviton4, not a microbenchmark.
Myth 1: FHE Takes Seconds Per Operation
Earliest implementations (2011-2015) did take seconds due to bootstrapping. Modern implementations do not bootstrap for most practical computations. H33's BFV pipeline uses a single-level modulus providing sufficient noise budget without bootstrapping. The 32-user batch completes in 943 microseconds -- under one millisecond for 32 operations.
Most practical workloads are shallow circuits. Appropriate parameter selection eliminates bootstrapping entirely. The myth applies to deep computations, not the batched shallow computations dominating real use cases.
Myth 2: Research Only
This confuses FHE libraries with FHE systems. IBM HElib, Microsoft SEAL, OpenFHE are research libraries. H33 is a production system: FHE + ZKP + PQ signing + caching + monitoring in a deployable binary. Sustained 2,293,766 auth/sec for 30+ seconds with health checks, metrics, and graceful degradation.
Myth 3: Impractical
FHE overhead is 1,000x-1,000,000x relative to plaintext. But if plaintext takes 1 nanosecond and FHE takes 1 microsecond, overhead is 1,000x but absolute latency is 1 microsecond -- invisible to any application. H33's 38 microseconds is faster than a network round trip, a database query, or a video frame.
Myth 4: Specialized Hardware Required
H33's numbers are on commodity AWS Graviton4. No custom hardware, FPGAs, or ASICs. Performance comes from algorithmic optimization, Rust engineering, and pipeline design (4,096 SIMD slots, fused stages, zero serialization).
Myth 5: Cannot Handle Real-Time
38 microseconds per authentication satisfies sub-millisecond requirements. A 32-user batch completes in 1.4 milliseconds. For streaming at 10,000 auth/sec with 32-user batches, total latency is approximately 4.6 milliseconds. Real-time by any authentication standard.
Why These Numbers
Multiplicative factors: SIMD batching (4,096 slots), ARM-optimized math on 192 vCPUs, fused pipeline stages, CacheeEngine sub-microsecond ZKP verification, lock-free structures, Rust zero-cost abstractions. Each provides 2x-10x. Combined: million-ops-per-second.
The myths are legacy beliefs from a decade-old landscape. FHE runs in production, at scale, on commodity hardware, at sub-50-microsecond latency. Contact support@h33.ai to see benchmarks on your own data.