Also, here's more evidence that Apple's advantage isn't simply a matter of process node. The Snapdragon 865 and Apple's A13 were both made on TSMC's N7P process:
At the bottom of that page are the aggregates, where we see the A13 does burn more power, but also delivers a lot more performance. In SPECint 2006, it delivers 29% better efficiency (4.74 vs. 3.67 SPEC marks per kJ). In SPECfp 2006, it delivers 11% better efficiency (11.18 vs 10.09 SPEC marks per kJ). However, do note the disclaimer about their power figures, at the top of the page, which calls into question the accuracy of Snapdragon's energy usage metrics. There's no denying the A13's performance advantage (58.5% and 36.8% for int & fp, respectively), but its efficiency could actually be even better (or worse) than what these figures indicate.
Of course, the cores in the M1-series chips are A14-derived. The only reason to look at the above benchmarks was to investigate the claim that Apple's advantage is simply due to process node.
Of course, there are other factors in Apple's favor, which have been mentioned in preceding pages. They can afford to make larger chips with bigger caches, for instance. Perhaps we could independently investigate that question, if we knew how well individual SPEC tests responded to cache size and could look only at those which didn't. However, I think it's generally well-established that SPEC 2006 suite, overall, benefits rather little from increased cache sizes.
At the bottom of that page are the aggregates, where we see the A13 does burn more power, but also delivers a lot more performance. In SPECint 2006, it delivers 29% better efficiency (4.74 vs. 3.67 SPEC marks per kJ). In SPECfp 2006, it delivers 11% better efficiency (11.18 vs 10.09 SPEC marks per kJ). However, do note the disclaimer about their power figures, at the top of the page, which calls into question the accuracy of Snapdragon's energy usage metrics. There's no denying the A13's performance advantage (58.5% and 36.8% for int & fp, respectively), but its efficiency could actually be even better (or worse) than what these figures indicate.
Of course, the cores in the M1-series chips are A14-derived. The only reason to look at the above benchmarks was to investigate the claim that Apple's advantage is simply due to process node.
Of course, there are other factors in Apple's favor, which have been mentioned in preceding pages. They can afford to make larger chips with bigger caches, for instance. Perhaps we could independently investigate that question, if we knew how well individual SPEC tests responded to cache size and could look only at those which didn't. However, I think it's generally well-established that SPEC 2006 suite, overall, benefits rather little from increased cache sizes.
Comment