If the N3B node costs more per transistor (wafer costs matter less because they don’t dictate the number of transistors per SoC) Apple has a pretty major problem on their hands. However even if it were the case that they couldn’t make an M3 pro for the same cost they could have done like the M2 Pro and unsold to the 8 HP core variant.
For all transistors that isn't true. But for circuits like I/O and SRAM which are not shrinking. If the wafer costs more and the area consumed by that particular kind of element isn't getting any smaller, then the defacto cost is going up. Yes.
Other parts might be at breakeven where more transistor density better offsets the increase in cost , but that isn't possible across the whole design anymore. Different types of circuits are getting smaller at substantially different rates ( or not at all. N3E SRAM/cache is the exact same size as N5. No shrink whatsoever. It just costs more on N3E. N3E is more expensive also, just not quite as large an increase as N3B. ).
This is why several other implementors are breaking toward chiplets. Some circuits don't make as much economic sense to keep pushing to more expensive fab wafers. But there is a tradeoff in manufacturing complexity and maximum Perf/Watt. Phones/Tabelts/thin laptops are going to try to stick with the monolithic dies for as long as they can, but that likely will lead to price creep.
Jensen Huang is right when he points at wafer costs going up means high end Nvidia GPU prices going up also. ( not only wafers but R&D cost go up because have to plow through experimental wafers to get to a working product and all of those cost more too. the design tools cost more. the simulation is more expensive , etc. etc. )
For a long time a major side effect of Moore's Law lead to cheaper computers over time. That part is pretty much over. Density will go up with ever more complexity tap dancing around the increasing issues, but costs aren't going down.
Bandwidth and cores lopped off go together in my de-contenting point. When NVIDIA does this people rightly complain that they are essentially getting a lower tier chip than the previous generation.
That is rather shallow. Bandwidth to the cores doing the actual work isn't necessary the same as bandwidth to highest level of the memory heirarchy. AMD's RNDA 2 added a big "Inifnity cache" that leverages bigger L3 to offset 'limiting' the width to GDDR5. RNDA3 rebalanced that again to a bit less case and faster GDDR memory (bit still leaning on the cache pretty heavily.)
Apple does the same thing. You can't just myopically look at just the LDDR bandwidth without looking at changes to internal network and the caching being done. Apple explicitly mentioned they have done some dramatic caching changes for the GPU cores. The E cores jumping 50% in performance over the M1 versions probably means have done some very substantive caching adjustments there also. Likewise if the P core cluster is larger then the aggregate local L2 is likely even larger too. ( which when run with a subset of cores active lead to effectively larger L2 hit rat). If can keep more stuff in the on die caches things go faster ( bandwidth to memory is incrementally less significant).
That isn't 'fanboy' bragging rights inference friendly ( bigger/more ... better 'porn' yeah ! ) . It is just betterto wait for real world , real app benchmarks before having a fit. Nvidia also manages to get 'at least as good if not better than last gen' improvements when they tweak things also. Alot of this is "would-a could-a should-a ... it would have been even better if they had done X" from the sidelines.
M3 pro is not inherently more balanced because balance is kind of arbitrary. the file enabled m3 pro no longer represents a middle point between m3 and m3 Max but is now more equivalent to the old harvested m2 pro dies that only had 10 cores.
From a 'win by getting maximum score' viewpoint perhaps can look at it as not balanced. But balance doing a wider variety of workload better is arbitrary at all. It just isn't the tech spec porn viewpoint. For what average users do these machines are more than fast enough. It is why Apple spends most of their time in the event and in the marketing pages primarily emphasizing the speed improvements of the M3 over the M1 (and legacy Intel) products than the M2. Vast majority of folks squat on systems for 3+ years now. Why? because they work just fine for vast majority of that time. Relatively few have to buy the newest shiny every year to get actually work done.
It’s a downgrade, the m3 pro does not represent straightforward evolution from m2 pro but makes many tradeoffs that make it worse than an upgrade or evolution should be.
Extremely few folks are going to turn around and buy a M3 Pro just 9-10 months after just bought a M2 Pro. They haven't even had it a year. M3 Pro far, far , far more about getting folks dragging their feet on Intel models to move. Next in line are getting folks who misbought a M1 model to move. Next in line is trying to get folks to abandon AMD/Intel Windows for Mac. M2 Pro buyers are far back in the priority stack.
Apple relatively just started selling the M2 Pro !!!! Apple isn't trying to move hardly any of those folks.
