For decode, memory bandwidth is the main bottleneck, so these machines will likely perform well even without a ton of GPU horsepower. Not as well as Blackwell,

For decode, memory bandwidth is the main bottleneck, so these machines will likely perform well even without a ton of GPU horsepower. Not as well as Blackwell, but I expect they will be a reasonable choice in terms of price/performance if you want to run large models with a lot of context. The main place they are a bit behind is in the number formats they support natively. Iirc M5 doesn't have native FP8 support, so you will take a speed penalty on quants where other architectures get better acceleration.