I guess it comes from this:
Geekbench link:
https://browser.geekbench.com/v6/cpu/5827749
Note the entry is marked as 'Invalid'.
If it’s 2.1GHz at that score that’s very good. The question though is what does the peak frequency look like for the phone, and then what does the performance per watt curve look like (for the whole mobo/phone like with Andrei/Geekerwan testing).
E.g if they can only get it to 3.6GHz, then they’ll have about a 2742 GB6 ST — in practice lower than that by some margin depending on timing stuff, L2/3 and memory, but let’s say 2650 @ 3.6GHz.
Okay, cool. That’s great. But what does power look like at that point? N3E is an easy + 10% boost iso-power from N4P, but they’ll need some real architectural changes and probably more cache to keep data movement lower in order to improve power. Right now on SpecInt2017 workloads the X4 is basically matching the A14 albeit at ~ 20% higher power - at least at 3.3GHz or whatever (maybe they have pushed their cores more than Apple by design).
Apple for all the talk about wide and slow has higher clocks than Android phones and is now pulling 4GHz in Macs at still low power, and I think nodes only explain so much. The A16 was doing 3.4-5GHz on plain N4, but the ST GB5 Geekerwan power was like 4.5W for a 20% perf uplift over A14 on plain N5, which is sure mainly frequency driven but with minimal node change that’s exactly why it’s impressive.
They did change to LPDDR5, and the SLC is 50% bigger than A14, and L2 is 100% bigger. Probably this is exactly how they lowered power in addition to the phydes changes from A15 to A16.
But that also goes to show how good the A14 really was.
Plain N5, 8MB of L2, 16MB SLC, LPDDR4x 4266, and yet the N4P 8 Gen 3 with plenty of SLC & L3 is just about matching it in Spec with 20% more power
I think Apple’s big L1 and humongous shared L2 + maybe some other phydes tricks are really killer as architectural decisions to reduce power. Even if Arm catch up on Perf/GHz for instance, and they might, it remains to be if they can achieve similar perf/W curves with similar process constraints.
Better than AMD/Intel? I think likely, even with similar phydes choices, yeah. But I am kinda concerned they won’t narrow this gap with Apple, and probably it will require some cache spending.