Performance
Determinant
MIND's int8 GEMM is byte-identical by construction, so identical inputs reproduce identical bits run to run, a property the vendor libraries structurally lack. It is also faster than the reference it is measured against.
2.0×
MIND int8 against single-core OpenBLAS f32 on commodity x86, 1024³, byte-exact. The shipped CPU path.
Source: bench/RESULTS-int8-2026-06-08.md
1.28×
GPU int8 against cuBLAS int8 on an Ampere-class GPU, byte-exact. This path runs in the commercial mind-runtime.
Source: docs/optimization-frontier.md
Exact-integer accumulation
Products sum in integers, so the result does not depend on the order the hardware adds them in.
Fixed tiling
The blocking is fixed at compile time, so every run walks the matrices the same way.
No atomics
No atomic adds race to a shared total, so there is no scheduling-dependent rounding.
Across the stack
Smaller benchmarks across the stack. Each card opens its full table and method.
Your GPU
A register-tiled WGSL GEMM shader and ONNX Runtime Web's WebGPU backend, same operation, same GPU.
1.8 to 15.5 µs
Parse, typecheck and IR lowering per program, Criterion.rs, 100 samples.
12.4×
Fewer bytes: the same reference IR is 90 B as mic@3 binary and 1,117 B as JSON.
4.3×
Fewer tokens than JSON-RPC for an agent session: 58 against 251.
Determinant, CPU. MIND int8 VNNI GEMM at 1024³ on one pinned physical core, median of 64 runs, against a single-core OpenBLAS f32 reference. The byte-identity gate passed on every configuration. The best configuration recorded 2.02×, rounded to 2.0× above.
Determinant, GPU. The int8 tensor-core kernel runs in the commercial mind-runtime. The 1.28× figure is cited in the public optimization notes; its raw record is not public, which is why it carries the Reported label. Cross-device byte identity across GPU substrates is not part of the public gate.
Compiler frontend. In-process Criterion.rs timings of parse, typecheck and IR lowering. They do not include binary emission. Earlier versions of this page quoted compile-time ratios against other frameworks; the benchmark record classifies those as cross-harness observations, not tier-matched speedups, so they are not repeated here.
MIC and MAP. Byte and token counts for one 6-node MLP IR reference and one agent session. Token counts use the cl100k_base tokenizer unless marked as the len/4 heuristic.
Compiler benchmarks reproduce from source. Results vary by hardware.
The reproduction guide lists the exact commands, inputs and environments behind each benchmark.
