Performance

Benchmarks

Verified performance data across the MIND stack — compiler, runtime, serialization, and protocol. WebGPU benchmarks run in your browser. Compiler benchmarks are reproducible from source.

Determinant — Deterministic BLAS

int8 · byte-exact

MIND's deterministic int8 GEMM — byte-identical by construction (exact-integer accumulation, fixed tiling, no atomics), so identical inputs reproduce identical bits run-to-run — a property the vendor libraries structurally lack. The CPU path is shipped and measured on commodity x86; the GPU path (Ampere-class) runs in the commercial mind-runtime. Cross-device byte-identity across GPU substrates is on the roadmap.

2.0× single-core OpenBLAS f32 (CPU)·1.28× cuBLAS int8 (GPU · commercial runtime)·byte-exact & deterministic (integer path)

→

GEMM — Matrix Multiplication

WebGPU

Tiled GEMM on WebGPU (1024–4096). Compares MindLang AOT-compiled WGSL shaders against ONNX Runtime Web's WebGPU backend performing the identical operation. Measures dispatch time, GFLOPS, and optionally includes compile overhead.

7–19x faster·~4.5 TFLOPS peak·8×4 register tiling·f32 precision

→

Compiler — MIND vs PyTorch 2.10

Criterion.rs

Frontend compilation (parse → typecheck → IR) vs PyTorch torch.compile full cold-start pipeline. Ubuntu 24.04, commodity x86 CPU, Ampere-class GPU, CUDA 12.8.

35,000–176,000× faster·176,000× conv2d·122,000× simple_mlp·1.8–15.5 µs frontend

→

Compiler — MIND vs Mojo 0.26.1

Criterion.rs

Frontend compilation vs Mojo mojo build full LLVM compilation to native binary. Same platform, February 2026 verified.

135,000–458,000× faster·458,000× scalar_math·280,000× matmul·135,000× mlp

→

Compiler — MIND vs JAX 0.9

Criterion.rs

Frontend compilation vs JAX jax.jit() cold-start XLA compilation (cache disabled). Ampere-class GPU, CUDA 12.8, JAX 0.9.0.1.

21,200–95,100× faster·95,100× large_matmul·58,600× simple_mlp·43,100× small_matmul

→

MIC Format — Token Efficiency

Serialization

MIC (MIND IR Compact) — the canonical IR wire format — benchmarked against JSON, TOML, and TOON for AI agent workflows. Latest formats: mic@2 (compact text) and mic@3 (binary IRModule). Measures byte size, token count, and parse speed against a 6-node MLP IR reference.

mic@2: 10.3× fewer tokens vs JSON·mic@3 binary: 12.4× smaller bytes·2.26 µs parse·90 B vs 1,117 (JSON)

→

MAP Protocol — Agent Communication

Protocol

MAP (MIND Agent Protocol) compared against JSON-RPC for AI agent session communication. Measures wire size, token count, and per-session overhead.

4.3× fewer tokens vs JSON-RPC·77% smaller wire size·234 bytes vs 1,004

→

Reproduction guide Performance overview

WebGPU benchmarks require Chrome 113+ or Edge 113+. Compiler benchmarks reproducible from source. Results vary by hardware.