Buying a m3 max with 128gb of RAM while will underperform any consumer NVIDIA GPU it will be able to load larger models in practice but slowly. I think a way f

Buying a m3 max with 128gb of RAM while will underperform any consumer NVIDIA GPU it will be able to load larger models in practice but slowly. I think a way for the m series chips to aggressively target GPU inference or training would need a strategy that increases the speed of the RAM to start to match GDDR6 or HBM3 or use it directly.