> Unfortunately, whether you use CUDA or not, NVIDIA hardware is still the only realistic performant option. > Maybe the move away from CUDA dependence for the

> Unfortunately, whether you use CUDA or not, NVIDIA hardware is still the only realistic performant option. > Maybe the move away from CUDA dependence for these frameworks will make other manufacturers step up competition for accelerating ML workloads. But they should have been doing that anyway... That is entirely down to SW, not the HW. For example consider these two 24GB cards: Nvidia 4090: 82.58 TFLOPS/$1,599 = 51 GFLOPS/$ AMD 7900 XTX: 61.42 TFLOPS/$999 = 61 GFLOPs/$