[outcoldman]

Marketing:

M5 announcement [1] says 4x the peak GPU compute performance for AI compared to M4. I guess in the lab?

Both iPad and MBP M5 [2][3] say "delivering up to 3.5x the AI performance". But all the examples of AI (in [3]), they are 1.2-2.3X faster than M4. So where this 3.5X is coming from? What tests did Apple do to show that?

---

1. https://www.apple.com/newsroom/2025/10/apple-unleashes-m5-th...

2. https://www.apple.com/newsroom/2025/10/apple-unveils-new-14-...

3. https://www.apple.com/newsroom/2025/10/apple-introduces-the-...

[storus]

M5 is supposed to support FP4 natively which would explain the speed up on Q4 quantized models (down from BF16).

[relativeadv]

Its not uncommon for Apple and others to compare against two generations ago rather than the immediately preceding one

[outcoldman]

I referenced everything about comparing to M4. I left outside the comparison with M1.