Talk:Apple M3

M3 Ultra

Has anyone found confirming information on the rumours of the M3 Ultra? "... the highest-end M3 Ultra will feature 32 CPU cores and an 80-cluster GPU ...": https://www.tomshardware.com/news/apples-high-end-m3-ultra-m3-max-and-m3-pro-expected-to-get-major-upgrades Jari Pyyluoma (talk) 09:47, 31 October 2023 (UTC)

Please do not discuss rumours. Only add verified information once it has launched. Ng.j (talk) 09:50, 31 October 2023 (UTC)

N3B or N3E

It looks like the linked source (https://asia.nikkei.com/Business/Technology/Apple-unveils-new-M3-processors-as-Arm-PC-chips-gain-traction) says:

"The M3 silicon will be manufactured by Taiwan Semiconductor Manufacturing Co., the world's largest contract chipmaker, using the latest semiconductor production technology, known as N3E, Nikkei Asia previously reported."

While the wikipedia article linking it says:

"It is built using the N3B process from TSMC."

My understanding is that the source is wrong and N3E isn't available yet, but the inconsistency with the presented source seems weird. 2601:601:200:1CAD:5C36:4E2B:70C7:9792 (talk) 19:59, 31 October 2023 (UTC)

I have provided multiple reliable industry sources that state it is made with N3E. The only dissenting source is Anandtech, which states it is probably N3B instead, but not definitive as they did not receive input from Apple or TSMC. Ng.j (talk) 19:03, 1 November 2023 (UTC)

I can't help but be skeptical that it's N3E. N3E entered volume production in 2H23; assuming (generously) that 2H23 means 1 July 2023, that's 4 months from the start of production to computers stocked in stores awaiting release. That seems way too fast. And according to this rumor from 29 August, Apple is the only customer of 3nm chips right now so if M3 is N3E then TSMC's entire efforts on N3B would have yielded just one chip (A17 Pro). To my knowledge this would be unprecedented and I highly doubt it would go unreported for this long; someone somewhere would have said something about misplaced priorities or bad assumptions on TSMC's part.

And then there's the fact that Nikkei previously said A17 Pro "will be mass-produced using TSMC's N3E" and "will be used in the premium entry in the iPhone lineup slated for release in 2023". Even last year that time frame was obviously impossible (assuming 2H23 N3E production) unless Apple pushed back the iPhone release significantly, which the report says nothing about. (To no one's surprise, Apple released the iPhone 15 Pro with N3B at the usual time of year.) Makes you wonder how reliable Nikkei really is.

All in all, it's too early to say for sure what's going on. I'm adding a disputed inline tag for now. Hopefully this will be resolved before too long. ErrorDestroyer (talk) 09:45, 4 November 2023 (UTC)

N3B is extremely expensive and has low yields. The rest of the industry is looking to N3E and skipping N3B because it's not viable. N3E and N3P have compatible design rules, but N3B is not. It is more of a developmental step. Ng.j (talk) 22:27, 6 November 2023 (UTC)

Precedent with disputed topics has been to focus on commonly agreed upon points. To that end, let's just put it as "3nm process" until there is more clarification from Apple and TSMC. Ng.j (talk) 22:29, 6 November 2023 (UTC)

Memory

I disagree with the Memory sizes that are mentioned. This is a page on the M3 chips. Each of the SoCs themselves do not support a certain quanta of memory such as the specified 36-192Gi, the new Apple Macbooks they come. I'll correct this. -- KelleyCook (talk) 15:48, 4 December 2023 (UTC)

Apple A16 or 17?

Has there been any verifiable source that has said what CPU cores are in the Apple M3? Is it truly a variant of the A17 or would it be closer to the A16? 2601:602:9000:8F40:4149:737E:9D46:B62B (talk) 22:56, 28 December 2023 (UTC)

If the latter is true, then the table of the variants is incorrect. 2601:602:9000:8F40:4149:737E:9D46:B62B (talk) 22:57, 28 December 2023 (UTC)

A17 is wrong, A16 is correct. You can see this by downloading IPSW files from Apple for A16, A17 and M3/Pro/Max devices, and looking at the DeviceTree files (extract the im4p with img4lib and then use my dt utility to parse it). There's a node called arm-io with a soc-generation property, which says H15 for A16 and M3/Pro/Max, and H16 for A17. If you want an external link of me saying that: here. Can't find any results on Google for "soc-generation" "H15" unfortunately, so I don't think any blog or tech outlet has picked up on this.

— Siguza (talk) 02:13, 7 January 2024 (UTC)

Well then that means the Variants table is incorrect; can someone fix it? 2601:602:9000:8F40:A51A:374D:E9D7:5FE5 (talk) 22:09, 8 January 2024 (UTC)

The issue I have with this is, the GPU architecture of the M3 is closer to the A17 Pro with Mesh Shading support and hardware accelerated ray tracing. Things the A16’s GPU doesn’t have. The M3 even has a new dynamic caching feature that atleast from what I can tell, neither the A16 or A17 Pro have.

In addition, the media engine on the M3 has AV1 support, which is in line with the A17 Pro. The A16 does not have such support.

The CPU arcitectures of both the big and small cores is where things get a little interesting. On the P core side of things, both press diagrams of improvements for the A17 Pro and M3 call out the exact same improvements. Improved branch prediction, wider decode and execution engines. I don’t think the A16 got such improvements over the A15, which the M2 was a variant of. The A17 pro’s E cores apple does not call out direct improvements but the M3 does. Maybe they share similar CPU architectures, maybe they are different.

The Neural engine is the biggest discrepancy. It would be nice to have comparable independent performance numbers of both to confirm, but apple does has the performance figures for the A17 pro are much higher than the M3’s neural engine. This is the only blatant example outside of the SoC-generation in the IPSW files that the M3 is not a direct variant of the A17 Pro.

In reality it is probably best to not list either mobile SoC as a variant, as without more concrete architecture diagrams it may be hard to say. It may also be fair to say that Apple is diverging its M-line development from its A-line a bit more than previous generations. TheMaxXHD (talk) 14:57, 12 January 2024 (UTC)

performance comparison according to table in article

Can it really be true that the A17 Pro has almost twice the performance (in TOPS) of every single M3 chip? As far as I understand the table in the article, this is the case despite A17 Pro sporting far less hardware units than the M3 variants. And all the M3 variants have the same performance (in TOPS) while sporting different extents of the same chip. To me it smells like at least one severe error in that table.

edit/addendum: Aah, now I see that perfomance column is supposed to be true for the "NPU" alone. I get it now. -- 91.137.24.149 (talk) 21:56, 4 August 2024 (UTC)

mention architecture acronyms

The M3 cpu gets referred to in multiple ways by e.g. "aarch64" (e.g. on many download pages for open source software and linux distros), "ARM64" (how the `arch` command on a macbook identifies itself, "RISC64", "A64", "ARMv8", "ARMv8-A", "ARMv8.6-A" (instruction), "ARM-A" (application architecture), armel, armhf, ...

Because the M3 cpu is so common and also because of all of the "can" and "may" within the relevant specifications see e.g. https://en.wikipedia.org/wiki/AArch64#ARM-A_(application_architecture) adding a section that list only what applies to the M3 CPU probably would be quite helpful to users for answering the seemingly simple question of "what is my architecture and which software version do I need? Or between which of them can I 'pick' and why?".

E.g. when looking at a download page like this: https://alpinelinux.org/downloads/ is the "x86_64", "ARMv7" or the "aarch64" the best choice? Or would it even run at all? Agowa (talk) 15:27, 29 August 2024 (UTC)