nx_gpu_bench.nx

nx_gpu_bench.nx -- SOVEREIGN three-way GPU-compute benchmark harness: OURS vs CUDA vs Vulkan, the no-wave measured-exceed instrument (operator 2026-06-20: "vm test ours vs cuda vs vulkan ... test all three on this machine" + "how does unsloth or other tools do this"). METHODOLOGY (grounded in how Unsloth / llama-bench / the kernel-bench world actually measure): - WARMUP before timing; MANY reps, report the MEDIAN(p50)+P95 (nx_bench_stats); GPU-clock timing for GPU lanes / monotonic-ns here; CORRECTNESS IS A GATE (a faster-but-wrong lane is DISQUALIFIED); report ABSOLUTE numbers + full config. ROLE: this sovereign harness is the JUDGE/AGGREGATOR. It runs the sovereign lanes (a naive FP32 ref that self-validates the instrument + the REAL fast integer-SIMD GEMM = __i16x16_madd, DETERMINISTIC), INGESTS the incumbent yardstick results (gpu_bench_cuda.out = cuBLAS FP32+BF16), and writes the unified scoreboard + honest verdict. The integer lane proves the DETERMINISM EXCEED: bit-exact (scalar==SIMD), which cuBLAS float CANNOT be. NO fake greens. license_tier: ORIGINAL