bench.md (5120B)
1 Benchmarks 2 ========== 3 4 Operations per second, higher is better. Tree mode at path length 32, so one 5 signature covers 32 subtrees. 6 7 | host | cores | backend | 8 | ------------------------- | ----- | -------- | 9 | AMD Ryzen 7 9800X3D | 8/16 | avx512 | 10 | Intel Xeon D-1541 | 8/16 | avx2 | 11 | Cortex-A520 + A720 | 12 | arm-sha3 | 12 13 gcc 14.2.1, `-Wall -Wextra -O2`, best of 5 on the Ryzen, 3 on the Xeon, 2 on 14 the ARM. The Xeon shares its host with other services, so treat it as a floor. 15 16 Only keygen and signing use the worker pool. Verifying and single mode run on 17 one core either way, so they are listed once. 18 19 ## SHA3-256, the default 20 21 | host | keygen/s | sign/s | verify/s | 22 | ------------- | -------- | ------ | -------- | 23 | 9800X3D | 18 | 0.56 | 40 | 24 | 9800X3D x16 | 186 | 4.72 | 40 | 25 | D-1541 | 2.0 | 0.063 | 12 | 26 | D-1541 x16 | 15 | 0.49 | 12 | 27 | A520/A720 | 1.1 | 0.034 | 8.0 | 28 | A520/A720 x12 | 7.0 | 0.21 | 8.0 | 29 30 ## Every hash size 31 32 Single core, then the same host threaded. 33 34 | host | hash | keygen/s | sign/s | verify/s | 35 | ------------- | -------- | -------- | ------ | -------- | 36 | 9800X3D | SHA3-224 | 20 | 0.63 | 43 | 37 | 9800X3D | SHA3-256 | 18 | 0.56 | 40 | 38 | 9800X3D | SHA3-384 | 12 | 0.38 | 26 | 39 | 9800X3D | SHA3-512 | 9.4 | 0.29 | 20 | 40 | 9800X3D x16 | SHA3-224 | 208 | 5.38 | 43 | 41 | 9800X3D x16 | SHA3-256 | 186 | 4.72 | 40 | 42 | 9800X3D x16 | SHA3-384 | 125 | 3.14 | 25 | 43 | 9800X3D x16 | SHA3-512 | 96 | 2.30 | 20 | 44 | D-1541 | SHA3-224 | 2.3 | 0.071 | 13 | 45 | D-1541 | SHA3-256 | 2.0 | 0.063 | 12 | 46 | D-1541 | SHA3-384 | 1.4 | 0.043 | 7.8 | 47 | D-1541 | SHA3-512 | 1.0 | 0.032 | 6.2 | 48 | D-1541 x16 | SHA3-224 | 17 | 0.57 | 13 | 49 | D-1541 x16 | SHA3-256 | 15 | 0.49 | 12 | 50 | D-1541 x16 | SHA3-384 | 10 | 0.34 | 8.0 | 51 | D-1541 x16 | SHA3-512 | 7.2 | 0.25 | 6.2 | 52 | A520/A720 | SHA3-224 | 1.2 | 0.038 | 8.8 | 53 | A520/A720 | SHA3-256 | 1.1 | 0.034 | 8.0 | 54 | A520/A720 | SHA3-384 | 0.49 | 0.015 | 3.5 | 55 | A520/A720 | SHA3-512 | 0.55 | 0.017 | 4.1 | 56 | A520/A720 x12 | SHA3-224 | 7.7 | 0.23 | 8.8 | 57 | A520/A720 x12 | SHA3-256 | 7.0 | 0.21 | 8.0 | 58 | A520/A720 x12 | SHA3-384 | 4.9 | 0.14 | 5.2 | 59 | A520/A720 x12 | SHA3-512 | 3.8 | 0.11 | 4.1 | 60 61 ## Single mode 62 63 One WOTS key, no tree, no path. Signs and verifies on one core. 64 65 | host | SHA3-224 | SHA3-256 | SHA3-384 | SHA3-512 | 66 | --------- | -------- | -------- | -------- | -------- | 67 | 9800X3D | 1395 | 1303 | 1055 | 626 | 68 | D-1541 | 453 | 424 | 315 | 197 | 69 | A520/A720 | 283 | 262 | 142 | 126 | 70 71 ## Chains 72 73 `make --directory test bench-chain`, one subtree of chains advanced 256 steps, 74 in millions of hashes per second on the 9800X3D. Higher is better. 75 76 - `plain`: an untweaked chain over the batch API 77 - `glue`: the tweak over the batch API, rebuilding records each step 78 - `fused`: the chain kernel 79 80 | hash | backend | plain | glue | fused | 81 | -------- | ---------- | ----- | ----- | ----- | 82 | SHA3-224 | scalar | 5.14 | 4.98 | 5.11 | 83 | SHA3-224 | scalar+bmi | 5.92 | 5.70 | 5.84 | 84 | SHA3-224 | avx2 x4 | 11.81 | 10.77 | 13.43 | 85 | SHA3-224 | avx512 x8 | 26.13 | 22.13 | 39.83 | 86 | SHA3-256 | scalar | 5.14 | 4.97 | 5.10 | 87 | SHA3-256 | scalar+bmi | 5.91 | 5.66 | 5.86 | 88 | SHA3-256 | avx2 x4 | 11.74 | 10.64 | 13.50 | 89 | SHA3-256 | avx512 x8 | 25.68 | 21.42 | 40.35 | 90 | SHA3-384 | scalar | 5.11 | 4.83 | 5.13 | 91 | SHA3-384 | scalar+bmi | 5.83 | 5.45 | 5.86 | 92 | SHA3-384 | avx2 x4 | 11.20 | 9.56 | 13.53 | 93 | SHA3-384 | avx512 x8 | 23.99 | 17.35 | 40.21 | 94 | SHA3-512 | scalar | 5.02 | 4.94 | 5.14 | 95 | SHA3-512 | scalar+bmi | 5.70 | 5.60 | 5.85 | 96 | SHA3-512 | avx2 x4 | 10.92 | 10.50 | 13.59 | 97 | SHA3-512 | avx512 x8 | 22.33 | 20.79 | 40.90 | 98 99 A chain step is one permutation at every hash size, so the fused kernel holds 100 about 40 Mh/s on avx512 from SHA3-224 to SHA3-512, where an untweaked chain 101 falls from 26.1 to 22.3. Reaching the chain through the batch API instead costs 102 a transpose and a scalar absorb per step, which is the `glue` column. 103 104 The scalar rows have one lane, so the batch API has nothing to overlap and its 105 `plain` column lands on the same one-permutation loop as `fused`. `scalar+bmi` 106 is the fallback on x86 without AVX, about 15% ahead. 107 108 ## Notes 109 110 Signing scales with path length: at path 32 each signature builds 32 subtrees, 111 and a shorter path is proportionally faster. Verifying does not, since it only 112 walks the path it is given. 113 114 Zen 5 has a full-width AVX-512 datapath. Intel parts that double-pump 512-bit 115 operations will sit closer to the avx2 numbers.