4 ms·

Yea, M1's win is more in locking up TSMC's 5nm supply than its genius design. AMD has been struggling to get their slice of TSMC 7nm capacity.

hajile 5y ago
AMDs chip peaks at 4.4GHz vs 3.2GHz. Together they imply 30-40% better IPC. 5nm only offers 15% better performance OR 30% lower power consumption vs 7nm. In truth though, the A14 design isn’t too much faster than A13 or even A12 despite them being 7nm. Most of its advancements were from clock ramps. I wonder if the hardware layer for x86 kept a lot of resources tied up on that instead of advancing the architecture performance more.
- enragedcacti 5y ago
  Someone correct me if I'm wrong, but comparing clock speeds cross-manufacturer and especially cross-architecture doesn't really make much sense, right? As far as I understand it, just because AMD and Intel chips are in the 5GHz range doesn't mean it is realistic to assume that Apple could also scale M1 up to that frequency so looking at just frequency or just IPC mean next to nothing. We pretty much have to resort to frequency x IPC i.e. benchmarks to gauge the relative difference. You're right that M1/A14 perf jump is relatively small and seems to effectively be a port of A13 to 5nm in terms of core design, so if we wanted to guess at AMD's perf at 5nm we could try to use that as a yard stick. I remember seeing someone do the benchmarks on that but I can't seem to find it now. He lowered ryzen 5000s wattage to match M1 with an offset to account for 7nm vs. 5nm and found that AMD was still behind in perf/w but by something like 20-30%. I'll keep looking to see if I can find it. Edit: Found it: https://old.reddit.com/r/Amd/comments/k3nbn3/psa_performance_doesnt_scale_linearly_with/ https://old.reddit.com/r/Amd/comments/k3nbn3/psa_performance... He arrives at the M1 having a 1.23x perf/w advantage in Cinebench when accounting for 7nm vs. 5nm. Without that it has a 1.34x advantage. Note that this is measuring the core directly. Ryzen's IO die is a huge power sucker so this really shouldn't be used to decide what laptop to buy or whatever, it's just a peek into how close their core architectures are to each other.

hajile 5y ago

AMDs chip peaks at 4.4GHz vs 3.2GHz. Together they imply 30-40% better IPC. 5nm only offers 15% better performance OR 30% lower power consumption vs 7nm. In truth though, the A14 design isn’t too much faster than A13 or even A12 despite them being 7nm. Most of its advancements were from clock ramps. I wonder if the hardware layer for x86 kept a lot of resources tied up on that instead of advancing the architecture performance more.

enragedcacti 5y ago
Someone correct me if I'm wrong, but comparing clock speeds cross-manufacturer and especially cross-architecture doesn't really make much sense, right? As far as I understand it, just because AMD and Intel chips are in the 5GHz range doesn't mean it is realistic to assume that Apple could also scale M1 up to that frequency so looking at just frequency or just IPC mean next to nothing. We pretty much have to resort to frequency x IPC i.e. benchmarks to gauge the relative difference. You're right that M1/A14 perf jump is relatively small and seems to effectively be a port of A13 to 5nm in terms of core design, so if we wanted to guess at AMD's perf at 5nm we could try to use that as a yard stick. I remember seeing someone do the benchmarks on that but I can't seem to find it now. He lowered ryzen 5000s wattage to match M1 with an offset to account for 7nm vs. 5nm and found that AMD was still behind in perf/w but by something like 20-30%. I'll keep looking to see if I can find it. Edit: Found it: https://old.reddit.com/r/Amd/comments/k3nbn3/psa_performance_doesnt_scale_linearly_with/ https://old.reddit.com/r/Amd/comments/k3nbn3/psa_performance... He arrives at the M1 having a 1.23x perf/w advantage in Cinebench when accounting for 7nm vs. 5nm. Without that it has a 1.34x advantage. Note that this is measuring the core directly. Ryzen's IO die is a huge power sucker so this really shouldn't be used to decide what laptop to buy or whatever, it's just a peek into how close their core architectures are to each other.