Something I’ve been curious about since the first M1 is how the IC design affects peak clock and power consumption. People knowledgeable in this matter have provided a lot of input and ideas, the full picture is still somewhat missing.
There is at least some evidence that Apple CPU cores are meticulously designed to provide peak performance at carefully chosen power levels. For example, power consumption of M4 starts climbing very quickly past 4.05ghz — the system reports 7 watts at 4.04Ghz, 7.5 watts at 4.1Ghz and 8.7 watts at 4.2Ghz. It costs 1 watt to climb the 50Mhz from 4.15 to 4.2 Ghz! Andrei F. from Anandtech described a similar effect starting from at least A12.
The crux of the question is this - are higher clocks actually feasible with the current design or is what we see the practical ceiling? It is clearly not economically viable to design separate core families for mobile and desktop use. In x86 land, it is a tradition trade mobile performance for higher desktop performance potential. Would Apple need to follow a similar route? Right now we get excellent mobile performance at mobile-level power targets, but we can't scale it futher. Would Apple Silicon that can scale further mean reduction is deliverable mobile performance? If that is indeed the case, then I think their strategy is sound. Laptops are the core of the Mac business, and it probably makes sense to prioritize 5% higher laptop performance instead of 5% higher desktop performance.
Anyway, just ruminating aloud 🙂
I guess I would slightly rephrase it as "Right now we get excellent desktop performance at mobile-level power targets, but we can't scale it further to get super-duper desktop performance." 🙃Like I get what
@theorist9 is saying that he wants the max possible SC performance in an M4 Ultra regardless of how fast the base M4 is already compared to its competitors but I feel like that approach is partially what got Intel into so much trouble and Apple creeping up on SC power is ... well it's fine for now, but I do worry a little bit about the trend. Maybe as you say it's for naught and it's just a natural part of how processors evolve if they want to continue "being #1 overall", but I fundamentally agree that if forced to choose Apple would probably rather remain "#1 mobile".
As for the role IC design plays, I do wonder what the effect of 2nm GAA will be. I suspect GAA will actual make the targeting of clock speed and power even more specific as GAA promises to allow an even greater exactitude of balancing power and performance by varying the gate size for each transistor to get the exact balance the designer wants in the exact right paths of the processor. It allows for large amount of customization where widening the gate results in more performance for greater power loss, while narrowing it results in less power loss but less performance and
you can mix and match as you please for every transistor. That's what TSMC, and other fabs, promise with this tech anyway.
=======
I think what some people are missing when they say a workstation CPU is too niche is that if Apple were to create a full workstation "Extreme" chip, to my mind the primary draw would not be the CPU at all, or at least not the CPU by itself - SC speed or the huge MC throughput both of which would be excellent but as mentioned many many times in this thread, of more niche interest than it once was. Rather, it would be the on die accelerators with that CPU and the huge memory pool of high bandwidth LPPDR on a device that might be expensive for an average user but is peanuts for the professional compared to buying anywhere near that level of GDDR/HBM memory. The massive CPU performance would be a bonus except as it aids in the calculations with the accelerators, which it can more easily do because they all share the same pool of RAM. It's that straddling of the consumer/professional divide that Apple likes to play in. Having a capable GPU, NPU, and other accelerators with that huge memory pool on a $10-15K device with a powerful CPU on die would be a unique solution. However, Apple still needs a few more features, especially on the GPU, to get there (and software). This wouldn't really be a competition with big iron, not right now but ... we've seen consistently how good enough consumer hardware almost always eventually replaces big iron in all but the most high powered of situations. So OpenAI won't buy hundreds of thousands of them, but a small firm? or the individual professional? who can't afford >$30K for a H100 or even more for a super chip? (and don't wait months or longer to get theirs). Hmmm ... to say nothing of graphics rendering and other applications like scientific compute which are increasingly moving to accelerators where they can. Plus you have that powerful CPU when you need it! And that's the target market Apple likes already for its professional/prosumer products.
www.notebookcheck.net
