TCB-Core v0.1: Public T81 Benchmark Report
This document is auto-generated by the benchmark_runner.
Last Updated: 2026-02-28 01:23:23 UTC Branch: main Commit: 23fda56
- SIMD capabilities (compile target): AVX2=1, SSE4.2=1, NEON=0
- Native benchmark mode: SIMD acceleration may fall back to scalar when unsupported by target/features.
- CanonFS in-memory binary read baseline is pinned to
baseline=map-read for cross-run comparability.
- CanonFS persistent rows are labeled
comparison=pipeline-advantage (systems-path comparison, not strict microkernel fairness).
Executive Summary
- Comparable benchmark rows (ratio computed): 12 / 87
- Verdict split (comparable rows): T81 wins 4, Binary wins 7, Comparable 1
- Data quality: missing baseline 51, semantic mismatch 4, defect rows 0
- Largest T81 advantage:
BM_Llama_Block (74.62x)
- Largest binary advantage:
BM_CanonFS_ReadThroughput_InMemory (2.38e-03x)
What To Use Today
| Category |
Comparable Rows |
T81 Wins |
Binary Wins |
Comparable |
Recommendation |
| CanonFS I/O |
4 |
0 |
3 |
1 |
Prefer binary for throughput-critical path |
| Compiler & Language |
2 |
0 |
2 |
0 |
Prefer binary for throughput-critical path |
| Inference Kernels |
5 |
4 |
1 |
0 |
Prefer T81 on this host |
| Tensor Math |
1 |
0 |
1 |
0 |
Mixed; benchmark workload path |
Top Signal Rows
| Benchmark |
Category |
T81 Iteration |
Binary Iteration |
Throughput Ratio |
Verdict |
Fairness |
Confidence |
| BM_CanonFS_ReadThroughput_InMemory |
CanonFS I/O |
12.01 ms |
28.52 µs |
2.38e-03x |
Binary wins |
baseline-pinned (map-read) |
high |
| BM_CanonFS_ReadThroughput_Persistent |
CanonFS I/O |
12.00 ms |
58.81 µs |
4.90e-03x |
Binary wins |
systems-path (pipeline-advantage) |
medium |
| BM_T81LangCompile |
Compiler & Language |
91.70 ms |
1.20 ms |
1.31e-02x |
Binary wins |
apples-to-apples |
high |
| BM_Llama_Block |
Inference Kernels |
86.48 µs |
6.45 ms |
74.62x |
T81 wins |
apples-to-apples |
high |
| BM_CanonFS_WriteThroughput_Persistent |
CanonFS I/O |
12.70 ms |
2.48 ms |
7.86e-02x |
Binary wins |
systems-path (pipeline-advantage) |
medium |
| BM_Llama_RMSNorm |
Inference Kernels |
1.24 µs |
4.67 µs |
3.77x |
T81 wins |
apples-to-apples |
high |
| BM_Lexer_AllTokens |
Compiler & Language |
646.01 µs |
182.22 µs |
2.82e-01x |
Binary wins |
apples-to-apples |
high |
| BM_Llama_RoPE |
Inference Kernels |
2.12 µs |
767.04 ns |
3.62e-01x |
Binary wins |
apples-to-apples |
high |
| BM_Llama_Softmax |
Inference Kernels |
38.70 µs |
72.41 µs |
1.87x |
T81 wins |
apples-to-apples |
high |
| BM_Llama_SiLU |
Inference Kernels |
37.45 µs |
61.67 µs |
1.65x |
T81 wins |
apples-to-apples |
high |
Category Detail
CanonFS I/O
| Benchmark |
T81 Iteration |
Binary Iteration |
Throughput Ratio |
Verdict |
Fairness |
| BM_CanonFS_ReadThroughput_InMemory |
12.01 ms |
28.52 µs |
2.38e-03x |
Binary wins |
baseline-pinned (map-read) |
| BM_CanonFS_ReadThroughput_Persistent |
12.00 ms |
58.81 µs |
4.90e-03x |
Binary wins |
systems-path (pipeline-advantage) |
| BM_CanonFS_WriteThroughput_Persistent |
12.70 ms |
2.48 ms |
7.86e-02x |
Binary wins |
systems-path (pipeline-advantage) |
| BM_CanonFS_WriteThroughput_InMemory |
214.60 ms |
206.65 ms |
9.63e-01x |
Comparable |
apples-to-apples |
Compiler & Language
| Benchmark |
T81 Iteration |
Binary Iteration |
Throughput Ratio |
Verdict |
Fairness |
| BM_T81LangCompile |
91.70 ms |
1.20 ms |
1.31e-02x |
Binary wins |
apples-to-apples |
| BM_Lexer_AllTokens |
646.01 µs |
182.22 µs |
2.82e-01x |
Binary wins |
apples-to-apples |
Inference Kernels
| Benchmark |
T81 Iteration |
Binary Iteration |
Throughput Ratio |
Verdict |
Fairness |
| BM_Llama_Block |
86.48 µs |
6.45 ms |
74.62x |
T81 wins |
apples-to-apples |
| BM_Llama_RMSNorm |
1.24 µs |
4.67 µs |
3.77x |
T81 wins |
apples-to-apples |
| BM_Llama_RoPE |
2.12 µs |
767.04 ns |
3.62e-01x |
Binary wins |
apples-to-apples |
| BM_Llama_Softmax |
38.70 µs |
72.41 µs |
1.87x |
T81 wins |
apples-to-apples |
| BM_Llama_SiLU |
37.45 µs |
61.67 µs |
1.65x |
T81 wins |
apples-to-apples |
Tensor Math
| Benchmark |
T81 Iteration |
Binary Iteration |
Throughput Ratio |
Verdict |
Fairness |
| BM_TensorMatMul_Naive |
10.53 ms |
9.76 ms |
9.26e-01x |
Binary wins |
apples-to-apples |
Known Gaps
- Rows hidden from this public view: 75
- Missing baseline rows: 51
- Semantic mismatch rows: 4
- Defect rows: 0