Apple s AI advantage is the chimeric memory dependency. Nvidia seems to artificially stratify the market by memory segregation.

As soon as Intel or AMD say fuck it and shoot the moon on memory Apples distinction evaporates.

I assume Intel is the most likely to do this by 2025.

Yes, but Intel is still mostly vertically integrated and I think a lot of people are looking to see if they can compete without completely changing their business model.

Intel says this is built on their Intel 4 process, which is, as I understand, their own 7nm process: https://www.intel.com/content/www/us/en/newsroom/resources/c...

> Intel says this is built on their Intel 4 process, which is, as I understand, their own 7nm process:

nm are meaningless marketing numbers. In terms of performance/density Intel 4 should be in same ballpark as '4nm' of Samsung/TSMC.

I know they're all marketing numbers. My point stands that everyone is watching to see how they compete, because their fab tech has been playing catch up.

>In terms of performance/density Intel 4 should be in same ballpark as '4nm' of Samsung/TSMC.

Any proofs to back that up?

It is true that Intel was ahead of TSMC for transistor density relative to the "nm" labelling of their nodes, which is exactly why the naming was changed. Intel 10nm was roughly equivalent to TSMC 7nm in density. They aren't beating TSMC in real terms, though. Intel is still lagging after their embarrassing stall developing the 10nm node.

Which reviewers are you talking about?

> Tests by independent reviewers (i.e., those not dependent on access to Apple for their continued existence) have demonstrated that the M series gains are almost entirely due to node advances

That’s a bold claim conspicuously missing data. Can you cite a source?

Well, I'm out of my depth, but it doesn't seem entirely inaccurate to say their their node advantage allowed them to sort of throw lots of hardware (transistors) at the problem.

https://www.anandtech.com/show/16226/apple-silicon-m1-a14-de...

    What really defines Apple’s Firestorm CPU core 
    from other designs in the industry is just the 
    sheer width of the microarchitecture. Featuring 
    an 8-wide decode block, Apple’s Firestorm is by 
    far the current widest commercialized design in 
    the industry.

For a list of relative transistor counts among CPUs:

    https://en.wikipedia.org/wiki/Transistor_count

Perhaps more to the point, I think the onus would be on somebody who claims Apple is doing something uniquely spectacular with their chips, other than simply capitalizing on a massive transistor count and power efficiency advantage made possible by their node advantage.

By the way, I'm not knocking what Apple has achieved. I recently upgraded from an 2018 Intel MBP to a 2021 M1 Max MBP. And, holy smokes. This thing is fast and it is effortlessly fast. I have been running all 10 cores pretty hard for the job I'm currently doing. I am absolutely beyond impressed.

But, from an engineering standpoint, it's definitely worth wrapping our heads around what Apple has and has not achieved here. I don't think they have any magic fairy dust or special sauce.

That quote is kind of the point: nobody would say that Apple didn’t effectively take advantage of the new process nodes but they didn’t _just_ take an off-the-shelf ARM core and run it n% faster, they also did significant work which has made their chips faster than those stock designs even on the same process. Android phones tend to be 4-5 years behind iPhone hardware because even when their chips are made on the same process as the older iPhone, they aren’t the same designs and by all accounts Apple’s design team is outperforming ARM, Qualcomm (prior to the Nuvia acquisition, we’ll see how much that shifts when they ship), and Samsung. Changing something like the execution width of a processor isn’t like adding “-n 8” to a build script, it has huge ripple effects throughout the design and that’s expensive to do.

The underlying challenge here is vertical integration: Apple can invest more in chip design not just on the CPU but also balancing with accelerators for things like AI or media processing, security, etc. because they work closely across the OS and app teams to see what they need & make sure it’s actually used (if they add a new coprocessor to reduce need for certain CPU features, it’s priority 1 for the next OS release). Microsoft and Intel or Samsung/Qualcomm and Google work together, of course, but they have different goals and budgets, they need to support multiple companies worth of corporate overhead on the same sale, and they can’t focus too much on any one deal because they need to hedge their bets.

This cuts the other way, of course: Apple’s stuff is best at the mobile and up to mid-range, but beyond that if you need more than a certain amount of RAM or features like advanced virtualization, they don’t have an alternative the way Google can choose between AMD and Intel for GCP with no concern about affecting the Pixel lineup. We’ll see whether Apple stumbles badly enough for this to matter.

Citations required.

More importantly, Apple doesn't fab its own chips. TSMC does. Everyone has access to the same vendor. Similarly everyone has access to ARM cores. Yet somehow Apple managed to build an Intel/AMD competitor and others didn't? What am I missing?

Yes. Absolutely, this. For those unwilling to click the link I think the title says it all:

"Apple Bought All of TSMC's 3nm Capacity for an Entire Year"

During the M2 series Apple had a node disadvantage compared to AMD and Apple was still ahead.

Being the first to buy 3nm doesn't prove what was initially claimed.

Only M3 was made with this latest 3nm process. However, M1 and M2 were produced using N5 and N5P respectively. Samsung, Qualcomm, AMD all have access to 5nm processes. In fact Samsung fabs their own 5nm chips while AMD's 7000 series / Zen4 chips were fabbed using the N5 process. However, they are not nearly as competitive as Apple's chips. The whole premise of "Apple's M series chips are only fast because of the process is incorrect".

> Everyone has access to the same vendor.

Unlikely.

Who can pay as much as Apple - likely in advance and for exclusivity.

My guess is that Apple is financing some of the equipment (TSMC is a high capital business, and Apple has overseas retained profits it doesn't want to repatriate) and there will be contracts for the exclusive use of the new equipment nodes. With everything designed for taxation efficiency.

AMD, Intel, Google, Samsung, and Qualcomm aren’t as big as Apple (a recent development which still sounds odd to say) but they’re all big enough to get first class support from TSMC even if several of them didn’t have the resources to compete directly.

Using the latest process certainly helps but look at the older ones as well - it’s not like the performance gap disappeared when competing AMD processors were launched on the TSMC 5nm process, but that really highlighted the different trade offs those teams make: Zen4 CPUs certainly dusted the Apple chips, but the ones which did were desktop / server designs using far more power, too, since that’s where the money is in the PC side.

Right. Anyone can beat Apple if you’re willing to burn electricity. Intel still makes the absolute top chips the public can buy. There is no contest in GPUs, the discrete AMD/nVidia cards stomp on Apple for the same reason.

But for similar performance, like in the MacBook Air or the MacBook Pros, Intel was embarrassed. The M1s were so much faster and cooler than the Intel chips they replaced it was hilarious.

The Mac Pro is absolutely the weak spot. Apple doesn’t sell enough so it seems unlikely they will spend the money to try and keep up with Intel there. The first Apple Silicon Mac Pro is not what people wanted. And I don’t know if that machine will be coming.

The old "anyone who disagrees with my world view is being bought" argument.

M series gains aren't just to do with the node. It's also because of the unified memory architecture and most importantly the fact that custom silicon e.g. Neural Engine is taken advantage of through the OS and SDKs.

That's true, given that the M-series gains are exclusively against Intel chips and are less performant than comparable AMD chips, upon which Apple's memory architecture is based.