Hacker News new | ask | show | jobs
by ivankra 105 days ago
But memory bandwidth (bottleneck for LLM inference) is only marginally improved, 614 GB/s vs 546 GB/s for M4/M5 Max - where is this 4x improvement coming from?

I think I'll pass on upgrading.
2 comments
singhrac 105 days ago
It’s prompt processing so prefill - that’s compute bound not memory.
link
0x457 104 days ago
4x is on Time To First Token it's on the graph.
link