Baseline. OpenFHE EvalFBT, one core: 127 µs per lookup.
B0Exact math on encrypted data.
Rented GPUs. Your keys. Nothing in the clear.
TFHE-rs 1.8.2 · H100 · one exact 4-bit lookup
54.2µs
twinflower · same H100
0.072µs
~750×
Same card. Same session. Their benchmark.
0 wrong.
Every output decrypted and checked.
Hundreds of millions of them.
Three minutes.
Real card data. No record ever seen.
twinflower_demo.mp4
● enc
0:00
2:59
1×
cost per exact lookup vs OpenFHE, one CPU core · sep 22
Nothing. Five escape hatches, all closed.
killedFirst run on a rented GPU.
H1 · RTX 4090Composable, end to end.
RTX 4090Same code, cheaper card.
RTX 4090Moved to a 24 GB gaming card.
RTX 3090Exact end to end. Also 5× slower.
killedThe big jump.
RTX 3090Under 1%. Not worth the code.
killedFraud model, encrypted. AUC 0.9534, same as plaintext.
engine v0At 138.5-bit security. Fits a 24 GB gaming card.
138.5-bitHead to head, same H100. 0.072 µs per lookup.
H100 SXMU.S. provisional filed.
USPTOKeys stay with the client. Worker stays warm.
v1.1Outside review. Independent reproduction. First paid pilot.
The results.
What we sent to FHE.org. Every number, the conditions behind it, and what failed.
| against | card | them | us |
|---|---|---|---|
| TFHE-rs 1.8.2, 4-bit bootstrap | H100 | 54.2 µs | 0.072 µs |
| Cheddar bare bootstrap (no lookup) | A6000 | 2.25 µs | 0.160 µs |
| OpenFHE method, GPU port | RTX 3090 | 1.486 µs | 0.138 µs |
| OpenFHE, one CPU core | rental $ | 1× | ~496× cheaper |
| Encrypted fraud model, AUC | RTX 3090 | 0.9534 plain | 0.9534 |
Fine print: per-lookup time, amortized over 65,536 values · batch, not single-value latency · fixed models · one host per measurement · 138.5-bit estimated security · outside reproduction in progress.