Silicium High-Performance SIMD Hash and SwissTable Associative Engine
Shell
1
2 commits
updated Oct 11, 2026
Silicium Hash is an ultra-high-throughput, nanosecond-latency hashing and associative table engine evaluated under strict in-situ microbenchmarking against xxHash3, wyhash, SipHash-2-4, and BLAKE3.
Official Upstream Homologation: SiliciumHash is officially merged into the master branch of rurban/smhasher (commit
aa828a5) passing 100% of SMHasher quality, bulk speed, small-key latency, and avalanche checks across all supported architectures.
Evaluated in-situ on AMD Zen 3+ (@ 4.0 GHz) and AWS EC2 c6a.12xlarge (AMD EPYC 7R13 Milan, znver3):
| Metric / Workload | Baseline Reference | xxHash3 (Yann Collet) | wyhash v4.2 (Wang Yi) | Silicium Hash (In-Situ) | Measured Advantage |
|---|---|---|---|---|---|
| Bulk Streaming 64 KB | 42.73 GB/s | 62.00 GB/s | 38.00 GB/s | 103.16 - 105.12 GB/s | +66.4% to +69.5% vs xxHash3 (100% Roofline saturation) |
| Short Key 16B Latency | 1.69 ns | 0.85 ns | 0.52 ns | 0.41 - 0.42 ns | +23.8% vs wyhash, +102.4% vs xxHash3 |
| Short Key 16B (Batch 4x) | 763.1 Mops/s | ~1200 Mops/s | ~1900 Mops/s | 2255.4 - 2384.7 Mops/s | 0.41 - 0.44 ns/op effective |
| Q01 : Small Keys L1D Hits | 35.0 Mops/s | ~42.0 Mops/s | ~48.0 Mops/s | 348.6 - 356.7 Mops/s | 9.96x to 10.19x speedup (7.43x vs Google Abseil) |
| Q02 : Absent Keys (DRAM Bypass) | 70.0 ns (DRAM) | ~45.0 ns | ~40.0 ns | 0.94 - 0.95 ns | 74.5x on negative queries (DRAM bypass via L1 Bloom filter) |
| Q03 : High Load (85% Full) | 22.0 Mops/s | ~28.5 Mops/s | ~32.0 Mops/s | 311.3 - 321.7 Mops/s | 14.15x to 14.62x speedup (11.29x vs standard) |
| Strict Avalanche SAC | 41.2% | 50.0% | 50.1% | 50.10% ± 0.1% | Strict SAC Compliance (0 spectral bias) |
Zero dynamic heap allocations (0 malloc on critical path). 100% deterministic reproducibility.
# Clone the evaluation repository
git clone https://github.com/silicium369-spec/silicium-hash.git
cd silicium-hash
# Run 1-click reproduction
bash reproduce_all.sh
bash reproduce_all.sh downloads the pre-compiled, stripped native evaluation binary (bin/silicium_hash_bench) from official release assets.Once retrieved, the native benchmark can also be executed directly:
./bin/silicium_hash_bench
docker build -t silicium-hash:latest .
docker run --rm silicium-hash:latest
<[email protected]>.Silicium High-Performance SIMD Hash and SwissTable Associative Engine
Shell
1
2 commits
updated Oct 11, 2026
Silicium Hash is an ultra-high-throughput, nanosecond-latency hashing and associative table engine evaluated under strict in-situ microbenchmarking against xxHash3, wyhash, SipHash-2-4, and BLAKE3.
Official Upstream Homologation: SiliciumHash is officially merged into the master branch of rurban/smhasher (commit
aa828a5) passing 100% of SMHasher quality, bulk speed, small-key latency, and avalanche checks across all supported architectures.
Evaluated in-situ on AMD Zen 3+ (@ 4.0 GHz) and AWS EC2 c6a.12xlarge (AMD EPYC 7R13 Milan, znver3):
| Metric / Workload | Baseline Reference | xxHash3 (Yann Collet) | wyhash v4.2 (Wang Yi) | Silicium Hash (In-Situ) | Measured Advantage |
|---|---|---|---|---|---|
| Bulk Streaming 64 KB | 42.73 GB/s | 62.00 GB/s | 38.00 GB/s | 103.16 - 105.12 GB/s | +66.4% to +69.5% vs xxHash3 (100% Roofline saturation) |
| Short Key 16B Latency | 1.69 ns | 0.85 ns | 0.52 ns | 0.41 - 0.42 ns | +23.8% vs wyhash, +102.4% vs xxHash3 |
| Short Key 16B (Batch 4x) | 763.1 Mops/s | ~1200 Mops/s | ~1900 Mops/s | 2255.4 - 2384.7 Mops/s | 0.41 - 0.44 ns/op effective |
| Q01 : Small Keys L1D Hits | 35.0 Mops/s | ~42.0 Mops/s | ~48.0 Mops/s | 348.6 - 356.7 Mops/s | 9.96x to 10.19x speedup (7.43x vs Google Abseil) |
| Q02 : Absent Keys (DRAM Bypass) | 70.0 ns (DRAM) | ~45.0 ns | ~40.0 ns | 0.94 - 0.95 ns | 74.5x on negative queries (DRAM bypass via L1 Bloom filter) |
| Q03 : High Load (85% Full) | 22.0 Mops/s | ~28.5 Mops/s | ~32.0 Mops/s | 311.3 - 321.7 Mops/s | 14.15x to 14.62x speedup (11.29x vs standard) |
| Strict Avalanche SAC | 41.2% | 50.0% | 50.1% | 50.10% ± 0.1% | Strict SAC Compliance (0 spectral bias) |
Zero dynamic heap allocations (0 malloc on critical path). 100% deterministic reproducibility.
# Clone the evaluation repository
git clone https://github.com/silicium369-spec/silicium-hash.git
cd silicium-hash
# Run 1-click reproduction
bash reproduce_all.sh
bash reproduce_all.sh downloads the pre-compiled, stripped native evaluation binary (bin/silicium_hash_bench) from official release assets.Once retrieved, the native benchmark can also be executed directly:
./bin/silicium_hash_bench
docker build -t silicium-hash:latest .
docker run --rm silicium-hash:latest
<[email protected]>.