The presentation event of the new Apple M3, MacBook Pro and iMac has allowed us to learn about the company’s renewed commitment to this segment. The jump to 3 nm photolithography has been confirmed, but the curious thing is that this feature has turned out to be a disappointment. Let’s see why.
unfair comparisons. Apple highlighted how the main improvements of the Apple M3 SoC come to its GPU, which are “up to 2.5 faster than those of the M1 chip family.” The CPU performance and efficiency cores are 30% and 50% faster than the M1 respectively. All very well, but the M1 (the base model) are chips from three years ago.
Not much faster than the M2. According to Apple’s own data published, the M3s are not significantly more powerful than the M2s they replace. Despite the jump to 3nm lithography, the efficiency cores of its CPU are 30% faster but the performance cores are 15% faster than those of the M2 chips. We expected more.
TSMC already warned. In 2020, TSMC already indicated how its chip production processes with 5 nm and 3 nm photolithography were underway, and the gains were going to be notable: the density improved with each jump, and if with the 7 nm node they could ” meter” 91.20 million transistors per square millimeter, with 5 nm the figure would rise to 171.3 million, and with 3 nm the number would reach 291.21 million transistors. Even so, the improvements in efficiency and performance would be more modest: when comparing 3nm with 5nm, they explained, they would achieve between 10 and 15% improvement in performance with the same energy consumption, or a reduction in consumption of between 25 and 30% with the same power.
Where are my transistors? There is, however, one piece of information that Apple has published that may explain these modest improvements in the performance of its CPU: the number of transistors has not changed as much as one would have expected. The move from 5 nm photolithography to 3 nm, as we have seen, allows the density to be multiplied by 1.7, but Apple apparently did not want to take full advantage of that option, and at least in its M3 and M3 Pro the number of transistors is suspiciously similar to that of the M1/M2 and M1 Pro/M2Pro.
Something strange is happening with the M3 Pro. This strange phenomenon is especially evident with the new M3 Pro, SoCs that surprisingly have fewer transistors than the M2 Pro: those had 40 million transistors, and the new M3 Pro have 37 million. How is it possible?
The M3 and M3 Max comply, the M3 Pro does not. Taking into account that the density is multiplied by 1.7, one would expect many more chips in both the M3 and the M3 Pro. Taking the M1 (also manufactured in 5 nm) as a reference, of the 16 million transistors we should have passed to 27.2, so the real 25 million is a fairly consistent figure. The 57 million M1 Max should have jumped to 96.9 million, so again the M3 Max’s 92 million is reasonable. On the other hand, the M1 Pro had 33.7 million, so the M3 Pro should have about 57.3 million. He is actually 37, an alarmingly low number.
Waiting for the M3 Ultra. If this progression and density multiplication factor is maintained, we can estimate the number of transistors in the M3 Ultra. If the M1 Ultra had 114 million transistors, it is expected that the M3 Ultra will end up having just under 193.8 million. That all these chips are somewhat lower, especially in the case of the M3 Pro, may have an explanation.
Smaller chips (and cheaper for Apple). With this decision, the Cupertino firm could have managed to maximize the profit margin of these chips, which do not include as many transistors for the simple reason that they are smaller. If they are smaller, the calculation is easy: more chips would be taken from each wafer, so the benefits grow for Apple, which here would be artificially “layering” the surface of each chip. The fewer transistors, the lower the performance gain across the board.
Either they are reserved, or the 3 nm are green. Here the reasons that could explain this decision are varied, but we would highlight two. The first is that Apple is simply holding back: if they launch somewhat more modest M3s, they will have room for improvement to tighten the screws on these chips with the future launch of the M4s. The second, that perhaps the current TSMC manufacturing node used in these chips is simply taking its first steps and the “yields” – effective production – are not as high as in more mature processes, something common in semiconductor manufacturing. Be that as it may, in the absence of independent performance tests one thing seems clear: the M3s do not seem as brutal as we had anticipated.