Not across the board. AMD Instinct GPUs are a real alternative for common PyTorch and JAX workloads, but "better" depends on which software and memory limits your model hits.
Where Instinct holds up
AMD documents framework support on Instinct: PyTorch through tested wheels and validated Docker images, and JAX on MI210 through MI355X. FP8 is available on the MI300X, MI325X and MI355X.
Where NVIDIA still leads
CUDA is, in NVIDIA's words, "the software layer that enables applications to harness the power of GPUs", with a CUDA-X library collection built on top. AMD's own HIP guide says it is "not intended to be a drop-in replacement for NVIDIA CUDA", so CUDA-only libraries and custom kernels are where a port costs time.
Memory is the other axis
NVIDIA lists 141GB of HBM3e at 4.8TB/s on the H200, against 80GB at 3.35TB/s on the H100 SXM. Compare the AMD figures on the MI325X guide with what your model needs.
The full comparison is in AMD vs NVIDIA GPUs for AI, and the competitor map is in NVIDIA's main competitors in AI. To test on NVIDIA hardware, see GPU pricing and availability.
