Skip to content

BENCH3.md

Latest commit

 

History

History
193 lines (156 loc) · 15.2 KB

BENCH3.md

File metadata and controls

193 lines (156 loc) · 15.2 KB
Raw

ITB Triple Ouroboros Benchmark Results

Security notice. ITB is an experimental symmetric cipher construction without prior peer review, independent cryptanalysis, or formal certification. The construction's security properties have not been verified by independent cryptographers or mathematicians.

PRF-grade hash functions are required. No warranty is provided.

No bespoke cryptography. ITB introduces no cryptographic primitive of its own — no custom S-box, permutation, or round function. It is a construction over existing primitives, much as PGP composes standard ciphers rather than defining one. Such constructions are not the object of algorithm-level cryptographic certification: national regimes (NIST CAVP/FIPS in the US, GOST/FSB in Russia, OSCCA's SM-series in China, IC3S in India, SOG-IS/EUCC and national lists in the EU, ASD's ISM in Australia, CRYPTREC in Japan, KCMVP in South Korea) certify primitives and the modules built on them, not compositional schemes. Eligibility for regulated use is therefore inherited from the primitives ITB is configured with, not conferred by ITB itself.

Results below were collected at ITB_NONCE_BITS=512 with ITB_GOMEMLIMIT=512MiB + ITB_GOGC=20 capping the Go runtime heap. Every PRF-grade primitive in the shipped hash registry dispatches through hand-written AVX-512 / AVX2 chain-absorb ASM kernels (each primitive family at its natural active register width); AES-ITB-128 dispatches through AES-NI kernels instead: the fused cascade auto-selects the widest VAES tier the host offers (ZMM on VAES + AVX-512F, YMM on VAES + AVX2 without AVX-512F, VEX / legacy-SSE XMM on the remaining AES-NI hosts), and the per-round VAES tiers stay built and force-reachable via ITB_FORCE_HASH_TIER.

Reproduction:

ITB_NONCE_BITS=512 ITB_GOMEMLIMIT=512MiB ITB_GOGC=20 \
  go test -bench='BenchmarkExtTriple.*_(1MB|16MB|64MB)$' -run='^$' -benchtime=5s -count=1

Intel Core i7-11700K 8C/16HT

The AES-ITB-128 rows below are measured on this host only; the other host sections do not carry an AES-ITB-128 row until a full-fleet rebench.

ITB Triple 512-bit (security: P × 2^(3×512) = P × 2^1536)

Hash Width ITB Width Crypto Encrypt 1 MB Encrypt 16 MB Encrypt 64 MB Decrypt 1 MB Decrypt 16 MB Decrypt 64 MB
AES-ITB-128 128 512 NPRF 131 295 368 193 489 678
Areion-SoEM-256 256 512 PRF 130 224 279 292 329 400
Areion-SoEM-512 512 512 PRF 134 237 293 304 356 442
BLAKE2b-256 256 512 PRF 104 162 184 197 210 235
BLAKE2b-512 512 512 PRF 106 164 189 193 212 246
BLAKE2s 256 512 PRF 89 128 144 142 159 173
BLAKE3 256 512 PRF 100 151 166 175 172 211
AES-CMAC 128 512 PRF 132 217 265 280 305 380
SipHash-2-4 128 512 PRF 116 184 213 226 244 284
ChaCha20 256 512 PRF 115 171 198 208 223 254

ITB Triple 1024-bit (security: P × 2^(3×1024) = P × 2^3072)

Hash Width ITB Width Crypto Encrypt 1 MB Encrypt 16 MB Encrypt 64 MB Decrypt 1 MB Decrypt 16 MB Decrypt 64 MB
AES-ITB-128 128 1024 NPRF 139 279 346 238 445 597
Areion-SoEM-256 256 1024 PRF 114 180 207 214 235 269
Areion-SoEM-512 512 1024 PRF 119 199 235 243 272 319
BLAKE2b-256 256 1024 PRF 83 116 126 126 134 145
BLAKE2b-512 512 1024 PRF 84 118 130 131 142 150
BLAKE2s 256 1024 PRF 64 85 90 90 96 100
BLAKE3 256 1024 PRF 74 103 111 112 122 128
AES-CMAC 128 1024 PRF 108 171 192 203 224 252
SipHash-2-4 128 1024 PRF 90 132 149 144 161 176
ChaCha20 256 1024 PRF 91 125 134 133 144 158

ITB Triple 2048-bit (security: P × 2^(3×2048) = P × 2^6144)

Hash Width ITB Width Crypto Encrypt 1 MB Encrypt 16 MB Encrypt 64 MB Decrypt 1 MB Decrypt 16 MB Decrypt 64 MB
AES-ITB-128 128 2048 NPRF 144 254 308 262 393 497
Areion-SoEM-256 256 2048 PRF 92 126 138 145 153 164
Areion-SoEM-512 512 2048 PRF 104 149 167 179 184 204
BLAKE2b-256 256 2048 PRF 57 70 75 76 80 82
BLAKE2b-512 512 2048 PRF 58 75 78 79 84 85
BLAKE2s 256 2048 PRF 43 50 52 51 54 54
BLAKE3 256 2048 PRF 51 63 66 65 69 72
AES-CMAC 128 2048 PRF 83 119 131 130 140 152
SipHash-2-4 128 2048 PRF 67 85 91 91 96 101
ChaCha20 256 2048 PRF 58 81 81 85 90 92

AMD EPYC 9655P 96C/192HT

ITB Triple 512-bit (security: P × 2^(3×512) = P × 2^1536)

Hash Width ITB Width Crypto Encrypt 1 MB Encrypt 16 MB Encrypt 64 MB Decrypt 1 MB Decrypt 16 MB Decrypt 64 MB
Areion-SoEM-256 256 512 PRF 397 469 513 679 888 1078
Areion-SoEM-512 512 512 PRF 393 488 519 765 908 1095
BLAKE2b-256 256 512 PRF 349 402 475 532 770 878
BLAKE2b-512 512 512 PRF 370 412 472 566 800 919
BLAKE2s 256 512 PRF 338 373 442 476 654 787
BLAKE3 256 512 PRF 358 382 472 518 746 865
AES-CMAC 128 512 PRF 378 454 503 673 912 1031
SipHash-2-4 128 512 PRF 381 427 497 629 842 964
ChaCha20 256 512 PRF 367 405 491 567 807 892

ITB Triple 1024-bit (security: P × 2^(3×1024) = P × 2^3072)

Hash Width ITB Width Crypto Encrypt 1 MB Encrypt 16 MB Encrypt 64 MB Decrypt 1 MB Decrypt 16 MB Decrypt 64 MB
Areion-SoEM-256 256 1024 PRF 364 411 483 576 808 913
Areion-SoEM-512 512 1024 PRF 378 429 488 660 772 916
BLAKE2b-256 256 1024 PRF 312 333 414 429 573 723
BLAKE2b-512 512 1024 PRF 316 353 423 459 611 764
BLAKE2s 256 1024 PRF 273 300 355 363 468 593
BLAKE3 256 1024 PRF 303 328 399 414 555 662
AES-CMAC 128 1024 PRF 372 418 469 571 765 900
SipHash-2-4 128 1024 PRF 336 371 433 506 658 803
ChaCha20 256 1024 PRF 326 361 415 454 620 764

ITB Triple 2048-bit (security: P × 2^(3×2048) = P × 2^6144)

Hash Width ITB Width Crypto Encrypt 1 MB Encrypt 16 MB Encrypt 64 MB Decrypt 1 MB Decrypt 16 MB Decrypt 64 MB
Areion-SoEM-256 256 2048 PRF 320 358 425 467 619 727
Areion-SoEM-512 512 2048 PRF 347 351 397 525 695 822
BLAKE2b-256 256 2048 PRF 248 272 338 319 432 502
BLAKE2b-512 512 2048 PRF 250 277 334 338 445 525
BLAKE2s 256 2048 PRF 203 226 269 258 326 383
BLAKE3 256 2048 PRF 235 259 303 303 386 476
AES-CMAC 128 2048 PRF 299 331 400 450 563 678
SipHash-2-4 128 2048 PRF 275 298 365 388 480 584
ChaCha20 256 2048 PRF 260 293 347 344 449 559

ARM64 Graviton 4 16C/16HT

ITB Triple 512-bit (security: P × 2^(3×512) = P × 2^1536)

Hash Width ITB Width Crypto Encrypt 1 MB Encrypt 16 MB Encrypt 64 MB Decrypt 1 MB Decrypt 16 MB Decrypt 64 MB
Areion-SoEM-256 256 512 PRF 43 49 51 46 54 57
Areion-SoEM-512 512 512 PRF 50 57 59 56 63 66
BLAKE2b-256 256 512 PRF 41 44 46 42 49 50
BLAKE2b-512 512 512 PRF 42 47 48 44 52 53
BLAKE2s 256 512 PRF 33 36 37 34 39 40
BLAKE3 256 512 PRF 20 23 23 19 21 25
AES-CMAC 128 512 PRF 53 60 61 57 67 69
SipHash-2-4 128 512 PRF 58 66 68 76 75 78
ChaCha20 256 512 PRF 6 22 31 7 20 32

ITB Triple 1024-bit (security: P × 2^(3×1024) = P × 2^3072)

Hash Width ITB Width Crypto Encrypt 1 MB Encrypt 16 MB Encrypt 64 MB Decrypt 1 MB Decrypt 16 MB Decrypt 64 MB
Areion-SoEM-256 256 1024 PRF 32 36 36 34 38 39
Areion-SoEM-512 512 1024 PRF 39 44 45 45 48 49
BLAKE2b-256 256 1024 PRF 27 32 31 28 34 34
BLAKE2b-512 512 1024 PRF 29 33 33 31 36 36
BLAKE2s 256 1024 PRF 20 24 24 23 25 26
BLAKE3 256 1024 PRF 11 14 12 14 13 15
AES-CMAC 128 1024 PRF 41 45 46 43 50 51
SipHash-2-4 128 1024 PRF 48 53 54 54 59 61
ChaCha20 256 1024 PRF 4 13 20 4 11 20

ITB Triple 2048-bit (security: P × 2^(3×2048) = P × 2^6144)

Hash Width ITB Width Crypto Encrypt 1 MB Encrypt 16 MB Encrypt 64 MB Decrypt 1 MB Decrypt 16 MB Decrypt 64 MB
Areion-SoEM-256 256 2048 PRF 20 23 23 23 24 24
Areion-SoEM-512 512 2048 PRF 26 30 30 28 32 32
BLAKE2b-256 256 2048 PRF 18 20 20 19 21 21
BLAKE2b-512 512 2048 PRF 19 21 21 21 22 22
BLAKE2s 256 2048 PRF 13 14 14 14 15 15
BLAKE3 256 2048 PRF 7 8 8 8 8 8
AES-CMAC 128 2048 PRF 27 31 31 28 33 33
SipHash-2-4 128 2048 PRF 35 39 39 39 42 43
ChaCha20 256 2048 PRF 2 7 11 2 6 11

Intel Core i7-11700K 8C/16HT — New 48-bit Interlocked ITB vs Old ITB (Lock Soup + Lock Batch mode) — Delta

Interlocked ITB with the nonce widens from 128 bits to 512 bits (secure default), and the overlay moves from an opt-in 24-bit Lock Soup + Lock Batch mask (roughly 2^33 mask space per chunk group) to an always-on 48-bit Interlocked Barrier (~ 2^70 mask space per chunk group). Previous "Lock Soup + Lock Batch" mode on the same i7-11700K host is the fair comparison point — both sides carry an overlay derivation cost per chunk, and the ratio isolates the ~ 2^37 mask-space widening plus the 4× nonce widening.

Encrypt at 64 MB (MB/s per primitive per width; new / old ; ► marks ratios ≥ 100%):

Primitive 512-bit E 64 MB 1024-bit E 64 MB 2048-bit E 64 MB
Areion-SoEM-256 279 / 198 (141%) ► 207 / 179 (116%) ► 138 / 141 (98%)
Areion-SoEM-512 293 / 226 (130%) ► 235 / 198 (119%) ► 167 / 162 (103%) ►
BLAKE2b-256 184 / 118 (156%) ► 126 / 84 (150%) ► 75 / 54 (139%) ►
BLAKE2b-512 189 / 174 (109%) ► 130 / 134 (97%) 78 / 88 (89%)
BLAKE2s 144 / 130 (111%) ► 90 / 91 (99%) 52 / 59 (88%)
BLAKE3 166 / 130 (128%) ► 111 / 101 (110%) ► 66 / 69 (96%)
AES-CMAC 265 / 165 (161%) ► 192 / 153 (125%) ► 131 / 120 (109%) ►
SipHash-2-4 213 / 158 (135%) ► 149 / 126 (118%) ► 91 / 90 (101%) ►
ChaCha20 198 / 110 (180%) ► 134 / 86 (156%) ► 81 / 58 (140%) ►

Decrypt at 64 MB (MB/s per primitive per width; new / old ; ► marks ratios ≥ 100%):

Primitive 512-bit D 64 MB 1024-bit D 64 MB 2048-bit D 64 MB
Areion-SoEM-256 400 / 248 (161%) ► 269 / 215 (125%) ► 164 / 165 (99%)
Areion-SoEM-512 442 / 280 (158%) ► 319 / 240 (133%) ► 204 / 187 (109%) ►
BLAKE2b-256 235 / 135 (174%) ► 145 / 93 (156%) ► 82 / 56 (146%) ►
BLAKE2b-512 246 / 211 (117%) ► 150 / 152 (99%) 85 / 96 (89%)
BLAKE2s 173 / 144 (120%) ► 100 / 100 (100%) ► 54 / 62 (87%)
BLAKE3 211 / 148 (143%) ► 128 / 110 (116%) ► 72 / 74 (97%)
AES-CMAC 380 / 211 (180%) ► 252 / 177 (142%) ► 152 / 137 (111%) ►
SipHash-2-4 284 / 187 (152%) ► 176 / 143 (123%) ► 101 / 98 (103%) ►
ChaCha20 254 / 133 (191%) ► 158 / 98 (161%) ► 92 / 64 (144%) ►

Every shipped primitive now sits at or above the old ITB line at 512-bit width on both Encrypt and Decrypt, and every shipped primitive on the 1024-bit line except BLAKE2b-512 stays at or above baseline (BLAKE2b-512 lands at 97-99%, essentially at baseline; BLAKE2s 1024-bit E sits at 99% and D at exactly baseline). The full ASM optimisation cycle (pack56 batched inversion + chain-kernel narrowing per family + ChaCha20 fused68 + interlock fill batching per family + bridge-free ZMM interlock kernel rewrite) delivers a fleet-wide throughput uplift that fully amortises the widened nonce envelope + always-on 48-bit interlock mask.

ChaCha20 gains largest (+80% at 512-bit E / +91% at 512-bit D / +56% at 1024-bit E / +61% at 1024-bit D) — the fused68 dual-compression kernel plus interlock fill batching remove the two dominant bottlenecks the old ITB line paid on this primitive; the interlock kernel rewrite compounds the win on the decrypt lane. AES-CMAC climbs to +61% at 512-bit E / +80% at 512-bit D and +25% / +42% at 1024-bit E/D — the AVX-512 VAES chain kernels amortise cleanly across the widened overlay + nonce cost. BLAKE family climbs from residual (65-90% pre-refresh) to above baseline at all 512-bit widths on all four variants, with BLAKE2b-256 sitting at +39-74% across all three widths thanks to combined narrowing + fill batching. Areion-SoEM family collects +30-58% on 512-bit Decrypt from the interlock kernel rewrite compounding on top of the pack56 batched inversion and the VAES-YMM 4-lane ChainAbsorb landings.

2048-bit residual is now split: AES-CMAC, BLAKE2b-256, ChaCha20, and Areion-SoEM-512 all sit at or above baseline at 2048-bit width (+3-46% D); SipHash-2-4 matches baseline (101-103%); Areion-SoEM-256 and BLAKE3 land at 96-99% (essentially at baseline); BLAKE2b-512 and BLAKE2s are the sole cells 11-13% below baseline at that width. The 2048-bit line reflects the wider container's higher per-chunk overhead relative to the per-byte hash cost — the pattern is comfortably absorbed by the wider security envelope, and the mid-tier residual is confined to two primitives rather than fleet-wide.

Further rows for other µarchs are scheduled — this table is a first-pass baseline pending maintainer-assisted runs on additional hardware.