NeuroBrix detects your hardware when you do not pass --hardware. What it can then run depends on the level of support below. When in doubt, run neurobrix doctor.
Install the PyTorch build that matches your card before neurobrix, inside your virtual environment.
| Card family | PyTorch wheel index |
|---|---|
| RTX 30/40, A100, H100 | cu124 |
| Older stable option | cu121 |
| V100 (Volta, sm_70) | cu126 |
pip install torch --index-url https://download.pytorch.org/whl/cu124
pip install neurobrix
Replace cu124 with your row. Expected result: neurobrix doctor lists your GPU, its compute capability and driver, and reports no blocking problem.
PyTorch wheels built for cu128, cu129, cu130 and cu132 do not carry sm_70. A default pip install can pull one of those and the first run then fails with an error that looks unrelated. On Volta use cu126:
pip install --force-reinstall torch --index-url https://download.pytorch.org/whl/cu126
neurobrix doctor compares each GPU's sm_XY with the architectures your PyTorch was built for, and says so when they do not match.
No minimum driver version and no minimum compute capability is stated yet. The engine ships configuration for Volta, Turing, Ampere, Ada and Hopper; Volta is the family validated on real hardware. neurobrix doctor reports what your machine has.
Supported through MPS, using the compiled execution path. Install in a virtual environment as usual:
pip install neurobrix
neurobrix doctor does not check MPS devices. Use neurobrix info --hardware to see what was detected.
A supported path. Hardware detection selects a CPU profile and small models run without configuration. To avoid downloading the CUDA build of PyTorch:
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install neurobrix
Expected result: neurobrix doctor reports that no GPU is present and that CPU is a supported path. Start with the smallest models in the catalogue.
Detecting a vendor is not the same as running on it. The honest levels:
| Level | Hardware |
|---|---|
| EXECUTES today | NVIDIA CUDA (with the Triton branch), Apple Silicon through MPS (compiled), CPU |
| WRITTEN, NOT LIT | AMD ROCm / CDNA; Triton on Apple GPUs (not on the main branch) |
| RECOGNISED only | Intel, Tenstorrent, Moore Threads, Biren, Iluvatar, Hygon DCU, Cambricon |
"Written, not lit" means the code exists but has not been run on the hardware. "Recognised only" means detection exists and nothing more. Do not plan a deployment on either. Details: Hardware support.
© 2025–2026 Hocine Benkelaya. NeuroBrix engine released under the Apache License 2.0.
Developed by Hocine Benkelaya · Supported and funded by Neural Network Holding Ltd · Commercial distribution by WizWorks OÜ