M4 Pro trails RTX 5090 bandwidth

- Benchmarks flag memory bandwidth as the dominant bottleneck for Q4 27B models, comparing Apple M4 Pro at 273 GB/s to an RTX 5090 at 1,790 GB/s. (x.com) - The note shows that despite similar compute peaks, the M4 Pro’s 273 GB/s sustained bandwidth forces more off‑chip stalls versus the RTX 5090’s 1,790 GB/s. (x.com) - For 27B inference and fine‑tuning, the results imply memory‑centric hardware choices will beat raw FLOP counts for throughput. (x.com)