Benchmarks
Latency by scheme
and backend.
Compare measured performance across schemes, operations, and hardware backends. Each run names the hardware it was measured on; inspect its parameters, precision, and execution settings.
CKKS bootstrapping · S2C-first
Latency by backend
Reproduce it ↗Lower is better ↓
Median bootstrap latency · AMD Ryzen 9 9950X and Apple M5 Pro · hardware shown per run
Preset n16_d35_k720_p19_s2cParameters
- Ring dimension N
- 65,536 (2¹⁶)
- Complex values per bootstrap
- 32,768 (2¹⁵)
- Secret Hamming weights
- Dense 1,024 · Sparse 32
- Max. ciphertext modulus
- 1,382 bits
- Max. modulus including keys
- Dense 1,680 · Sparse 120 bits
08.76 s17.5 s26.3 s35 s
Select a backend row to inspect its hardware, configuration and precision. Speedup is relative to the single-thread portable backend for the same pipeline on the same hardware.
BackendHardwareTransformMedianSpreadSpeedupVersion
Reading the results
Compare the configuration, too.
Compare the same pipeline and preset, and account for the hardware and thread count: the runs were not all measured on the same host, and each row names its own. Each configuration has 20 measured iterations after three warm-ups. Timings include bootstrapping and return to application scale 2³⁵; setup, key generation, encryption and decryption are excluded.
Latency is the median; spread is median absolute deviation divided by median. Speedup compares each configuration with the single-thread portable backend for the same pipeline on the same hardware; it is not given where that backend was not measured on that host. These results were measured with Poulpy v0.9.0.
The GPU backend is in development, and Binary FHE results are not yet available. Pending results are excluded from rankings and configuration counts.