i think maybe something we’re forgetting about here is that engineering resources have been diverted to server chips in the wake of the AI boom.
That really shouldn't be having a deep impact. Chip design is a 3-4 year horizon pipeline. That "AI hyper fad" in 2023 would lead to a chip coming in 2026-27 ... which is ...
" ... At least a trio of companies are believed to be involved with the chip. Apple is said to be handling the overall design of the chip, while Broadcom is said to be providing some networking technology for it. TSMC is expected to begin mass production of the chip in 2026, using its third-generation 3nm process, known as N3P ..."
Apple Intelligence servers are currently powered by the M2 Ultra chip, and they are expected to start using M4 series chips next year. In an eventual...
www.macrumors.com
What would have been 'oversubscribe' on design resources 3-4 years ago would have been R1 ; not the server chip.
There is little indication that what Apple is deploy in the near term is just different logic boards for the mainstream Mac based SoC for these server deployments.
The core issue the lack of proper time overlooks is that Apple had to make bets on the M3 before the M1 ever shipped. The M3 was in flight without not knowing at all how the transition was going. ( made even more opaque with layering the pandemic on top of the transition. And TSMC N3 being a substantive leap forward in fab process risk. )
They’re not going to divert from the flagship line (which is really only the pro and max), so the low-sales specialty tier (Ultra) is what suffers.
Again also not quite up to what is occuring at tactical level now. For the M1 and M2 the Pro shared a very high degree of layout/floorplan with the Max. ( The CPU count the same , NPU same , top 'half' GPU same, top 'half' memory channels same )
With M3 that changed. M3 Pro has six E cores. The M3 Max four E course. The Max has more P cores and generally just as substantially more silicon throw at it.
What missing was before Mac transition Apple had basically two A-series chip dies. The A__X designs periodically would skip a generation.
The plain Mn die basically takes the place of the A__X ( iPad Pro and entry Mac same SoC). So the Pro and Max were additional work overhead. As noted before M1 and M2 maximized that shared R&D for those two.
With M1 Ultra Apple likely was shooting to get folks to understand the move from large scren iMac --> Studio. M2 generation saw Mini get a Pro ( move volume to support future diverging Pro design) and M2 Mac Pro ( that was in class to supercede Intel 16 core + W5700X common configuration).
The M1/M2 Max being shared with the laptops is also a risk mitigation. If 14"/16" Max didn't sell well then the Max SoC can be consumed by the Studio ( and later Mac Pro in M2 generation). If 14"/16" Maxs sold extremely well they could just different the Studio and even more of the overhead costs of UltraFusion paid down by customers that couldn't possibly use it. (so below/above Studio Ultra sales expectation would matter less. )
“Not every generation will have an Ultra” then reads a bit like revisionist commentary on product strategy. AKA, it’s a convenient explanation that makes what is happening seem like it was the plan all along.
Probably not. Very good chance the M2 Ultra ( which was necessary to launch a creditable Mac Pro) likely was not targeted until end of 2022. Remeber Apple said two year transition. ( so either June 2022 or December 2022) needed to make that. M3 Ultra was very uniikely. If N3 had gone pefectly to plain in 2H2022 then would only have gotten M3 , Pro , Max out the door in time for that deadline. Ultra wouldn't have made it at all. That N3B had even more hiccups than expected; not a chance.
By M3 generation it would have been time to see if the desktop Max/Ultras could pull their own weight without subsidies from users that cannot possibly use the UltraFusion connector. Subsidize while in risky transition phase with radical 'change' products ( no screen on large iMac performance zone. No 3rd party GPU cards) makes lots of sense. Subsidize forever on every generation ... not so much.
I’d bet that, under the hood of top-level marketing and strategy goings-on, in the wake of making a killer product that redefined the boundaries of necessity, they’re trying to adjust what all the performance breakpoints should be, as well as the corresponding naming.
I wouldn't take that bet. I think the primary objective was to make the Max a Ultra 'killer' all along. To put a "Mac Pro 2019" into a 16" laptop was always the lead goal. The M4 Max is largely that (minus some fringe corner cases at top end of price sale that the MBP is relatively far cheaper than ) .
The AI server thing is likely going to look somewhat like a Google Tensor chip that Broadcomm helps them with.
Cloud Tensor Units.
cloud.google.com
These "sever chips" have HBM ( not RAM DIMMs ) and high connectivity built in. It isn't trying to mimic AMD EPYC or Intel Xeon 6 chips. More likely the Broadcomm connectivity will display all of the Thunderbolt stuff that Apple puts into the laptop/desktop units. It will look more like a data center GPU (next to no or zero display out) ... that really won't be viable for a Mac Pro.
Taking the display controllers and Thunberbolt out of a Max/Pro die really wouldn't be too hard. There are options for attaching the Broadcomm stuff. Either via UCI-e ( standard interconnect that will be quite viable in 2026) . Apple could reuse their "poor man's HBM" memory with LPDDRx if want to shave costs on that also. Change ratio of die space for NPUs versus CPU shouldn't be a huge stretch.
So what we get in the M3 Ultra Studio is just something easy based on an “older” architecture to satisfy that high-end specialty segment. At the end of the day, an M3 Ulta is still going to belch out those exports faster than an M4 max, and run large language models that an M4 Max can’t.
There never any economic sense in Apple throwing Ultra sized Chips in the trashcan every 12 months made any economics sense at all. The plain Mn chip doesn't disappear every 12 months. It gets 'handed down" to the iPad Air. The A-series get handed down to iPads/AppleTV/etc. The Apple Watch SoC gets handed down to the HomePods. The vast majority of Apple's silicon is about using the SoC in
multiple products. Throwing away the Ultra every year doesn't do that.
[ The Mn Pro is on slippery slope also. Somewhat likely why they shrank it a bit by leaning more on E cores than P cores. Apple probably is looking for something else to toss it into. Or may fall back to staggering the Mini updates to extend the lifetime. ]
The M3 Ultra is likely coming now on Studio because Apple needed more time to get returns back on M2 Max ejected from the MBP 14/16" in 12 months. The rapid pace of ejection from the MBP 14/16 is likely contributing to the delay in updating the Studio and Mac Pro.
The M4 Max Studio has to compete with the M4 Pro Mini. It really can't wait too long because the GPU upgrades Apple has made M3,M4 have been substantive. Which is mainly to enhance the laptop segment.