Your float32 data gets compressed. SPCMP also gives it a canonical form, a reversible record, and a Merkle commitment.
python3 prove_divergence.py
Note: prove_divergence.py is a broad environment test. If your two machines return a hash I’ve already published, they behaved identically on this test, which does not prove identical hardware. If they return a unique hex value, that is new data. For path-dependent divergence on GPUs (FMA, fast-math, atomic scheduling), see the GPU harness.
Your float32 data compresses. SPCMP also gives you a Merkle root. Same size reduction, plus tamper-evidence.
SPCMP is a reversible float32 container format that enforces a deterministic total order.
IEEE 754 does not guarantee bit-identical results across platforms, and NVIDIA’s own CUDA documentation says so. The records below show what that looks like in practice.
One script, five environments, three answers. A GPU harness shows the same effect from compiler flags and thread scheduling.
IEEE 754 does not guarantee bit-identical results across platforms, and NVIDIA’s own CUDA documentation says so. The records below show what that looks like in practice.
Two binaries. Four commands between them. Every run checks in against the license server before it does anything.