They are viable in different ways. AMD is a general-purpose alternative for standard frameworks. Tenstorrent is an open-source option with a smaller footprint. Groq's LPUs now show up inside NVIDIA's platform rather than against it.
AMD
Instinct GPUs run PyTorch and JAX through ROCm, and FP8 is available on recent Instinct parts. The cost is CUDA-only libraries, as covered in whether AMD is better than NVIDIA for AI.
Tenstorrent
The software stack is open source and compiles PyTorch through TT-Forge. Blackhole cards carry up to 32GB of GDDR6 each, per Tenstorrent.
Groq
Groq's own site now describes the NVIDIA Groq 3 LPX rack, which connects 256 LPU accelerators to NVIDIA's Vera Rubin platform. Our write-up is in the NVIDIA Groq 3 LPX guide.
How to decide
Validate on your own model before committing. The full landscape, including Google TPU, AWS Trainium and Intel Gaudi, is in NVIDIA's main competitors in AI. To run a workload on NVIDIA hardware as the baseline, see GPU pricing and availability.
