If you're into LLM's you need to consider using Apple Silicon so you can continue developing on your own machine with MLX. MLX isn't just a replacemnet for CUDA, no way - it's made for the Apple arch. The new M5 chip will likely have serious ANE and ability to use the unified memory. Then we're talking powerhouse. Current Mac Studios are already more cost effective than say RTX 6000.
That's just a unified memory architecture. Strix Halo has this, too. It's not only Apple. Memory bandwidth is lower than the ultra series but faster or on par for every other M series processor.
If you're into LLM's you need to consider using Apple Silicon so you can continue developing on your own machine with MLX. MLX isn't just a replacemnet for CUDA, no way - it's made for the Apple arch. The new M5 chip will likely have serious ANE and ability to use the unified memory. Then we're talking powerhouse. Current Mac Studios are already more cost effective than say RTX 6000.
That's just a unified memory architecture. Strix Halo has this, too. It's not only Apple. Memory bandwidth is lower than the ultra series but faster or on par for every other M series processor.