Discussion Apple Silicon SoC thread

[senttoschool]

We are closing the circle of this useless discussion (and I'll stop replying after this one, no sense in discussing tastes), but I expected a redesign to eventually happen because the M1 Mac mini contained enormous unused space:

I'm not sure if Apple thinks the Mini is important enough to do a redesign. There's not much value in spending resources redesigning it and changing the supply chain. Desktops don't sell well and space doesn't matter for desktops.

 

[GC2:CS]

It looks to me like M2 Pros might use the A16 blocks. The CPUS have 16 MB caches(x2 for each cluster) they also look much larger than the A14/15 cores on the die shot. The GPU was mentioned to have larger caches as well (i think A16 mentined that too).
It is interesting they use 4x ram packages on both the Pro and Max. Why that might be.

 

[richardskrad]

.

 

[Eug]

Is there any more doubt that Apple plans to update their chips once a year?

But Apple didn't actually update their chips once a year.

And much faster than the 18 or 24 month update cadence that most people here speculated.

I was expecting 16 months, but they beat that by 2 months, at 14 months.

I don't understand why some people were expecting M2 Pro and M2 Max to be on N4 or even N3.

Base M1 was on N5, and so was the rest of the M1 family.

Base M2 was on N5P, so I full expected the M2 Pro and M2 Max to be on N5P.

Well, people were not expecting a January 2023 release. They were expecting a spring 2023 release (or later).

 

[senttoschool]

But Apple didn't actually update their chips once a year.

I was expecting 16 months, but they beat that by 2 months, at 14 months.

*sigh*

 

[guidryp]

Regarding the empty space, I wonder how the M2 Pro changes the equation. I look forward to seeing the inevitable teardown next week.

Good point.

Also, Apple Design is more practically oriented since Jony Ive left.

If Ive was still in control, I'm sure he would have looked at that space and redesigned it to make it 7mm thinner, while making the HSF cooler design and PSU design more difficult.

The more practical Apple can just keep using the same case and PSU, and have a bit of breathing room on HSF cooler design.

 

[Eug]

The Mini form factor was designed to handle an Intel CPU with 65W TDP. Which is enough for a 'Max' chip. So maybe next year you'll be able to get it with M3, M3 Pro or M3 Max.

According to various users, the higher spec Intel Mac minis were unbearably loud under heavy load.

The M1 Max MacBook Pros also can get audibly noisy under heavy load, even though it's nowhere near as bad as the Intel MacBook Pros.

Given the noise issues of the M1 Max Mac Studio (which may be more related to fan and enclosure design than heat output), perhaps the M2 Pro Mac mini will get a lot of sales from people like those in music production.

 

[Ajay]

Apple has previously taken standard TSMC processes because they help design them and use them when they are brand new - there is no time for further customization. I think it is highly likely they use some custom N3 variant that includes some/most of the changes N3E is getting rather than bog standard N3B. Why should they use a cookie cutter process when they are TSMC's biggest customer, and work with them from the early stages of process definition? Especially when they have plenty of time before they need to start shipping A17 SoCs to Foxconn assembly plants.

N3B is going to see almost no use, the rumors that Apple would use it for M2 Pro/Max proved untrue so I don't think Apple will ever use it. It was probably going to be used by companies making bitcoin ASICs but that business has dried up so even hitting the 1000 wpm they thought they would ship is probably not going to happen.

Well, if that is the case, Apple. could start buying risk wafers of their customized N3 process and have enough good A17 SoCs for at least the iPhone Pro models - they've done this before (essentially funding the ramp to HVM). If yields improve enough - that process could be used in M3 SoCs. The geometric and electrostatic gains for that node are good enough.

 

[Mopetar]

According to various users, the higher spec Intel Mac minis were unbearably loud under heavy load.

The M1 Max MacBook Pros also can get audibly noisy under heavy load, even though it's nowhere near as bad as the Intel MacBook Pros.

I have an M1 Max MBP and can confirm that if you can log it enough you can hear it, but it's not as load as my older Intel MBP could get and it's far less likely to get to that state so you rarely hear the fans fully spin up unless you're running some workload that will keep the CPU/GPU working hard for several minutes straight.

 

[q52]

I didn't know Apple was in the business of stealing in-progress production projects and selling them on the black market.

IME the real issue with using cloud storages when it comes to production work is sharing

too easy to mark a cloud folder as "public" by accident or share a link to a folder of content that was supposed to be kept private

a lot of horror stories and lawsuits have come down to someone accidentally sharing e.g. a Dropbox link inappropriately so its not unexpected for such a company to want to crack down on cloud-anything in response

 

[q52]

I have an M1 Max MBP and can confirm that if you can log it enough you can hear it, but it's not as load as my older Intel MBP could get and it's far less likely to get to that state so you rarely hear the fans fully spin up unless you're running some workload that will keep the CPU/GPU working hard for several minutes straight.

its hard to put into words just how happy I am that we are finally leaving this era of laptops with crazy super loud fans spinning up all the time

it was always SO embarassing to be in a meeting with a client and accidentally click something and your laptop goes beserk and the fans start spinnning up soooo loudly and there's nothing you can do to stop it

or when you are in the office trying to work and one of your colleague's laptop starts spinning up so loud it sounds like a plane about to take off

using a fanless MBA for daily use is such a game changer, and the M1/M2 MBP's are starting to trickle into the offices too and fan noise is slowly fading into the past. Cant wait to finally put this behind us

 

[MadRat]

I don't remember being unable to set maximum speeds in profiles to prevent fan noise. Even my old backpack 386 had the option.

 

[repoman27]

I don't understand why some people were expecting M2 Pro and M2 Max to be on N4 or even N3.

Base M1 was on N5, and so was the rest of the M1 family.

Base M2 was on N5P, so I full expected the M2 Pro and M2 Max to be on N5P.

M1 Pro/Max/Ultra were actually on N5P, same as A15 and M2. I originally thought the process would be dictated by whatever node the cores were originally designed for, but then it started to look more like Apple was going with the best available node within the process family at the time the SoC was released.

SoC	Core Microarchitecture	Process Node	Release Date
A14 Bionic	Firestorm / Icestorm	N5	Sep 2020
M1	Firestorm / Icestorm	N5	Nov 2020
A15 Bionic	Avalanche / Blizzard	N5P	Sep 2021
M1 Pro, M1 Max	Firestorm / Icestorm	N5P	Oct 2021
M2	Avalanche / Blizzard	N5P	Jun 2022
A16 Bionic	Everest / Sawtooth	N4	Sep 2022
M2 Pro, M2 Max	Avalanche / Blizzard	N5P?	Jan 2023

N3 was never in the cards, but N4 made sense, at least to me.

 

[repoman27]

Apple has previously taken standard TSMC processes because they help design them and use them when they are brand new - there is no time for further customization. I think it is highly likely they use some custom N3 variant that includes some/most of the changes N3E is getting rather than bog standard N3B. Why should they use a cookie cutter process when they are TSMC's biggest customer, and work with them from the early stages of process definition? Especially when they have plenty of time before they need to start shipping A17 SoCs to Foxconn assembly plants.

N3B is going to see almost no use, the rumors that Apple would use it for M2 Pro/Max proved untrue so I don't think Apple will ever use it. It was probably going to be used by companies making bitcoin ASICs but that business has dried up so even hitting the 1000 wpm they thought they would ship is probably not going to happen.

According to TSMC's recent earnings call:

"Our N3 has successfully entered volume production in late fourth quarter last year as planned, with good yield. We expect a smooth ramp in 2023 driven by both HPC and smartphone applications. As our customers' demand for N3 exceeds our ability to supply, we expect the N3 to be fully utilized in 2023. We expect [sizable N3 revenue contribution] to start in third quarter '23, and N3 will contribute [a] mid-single-digit percentage of our total wafer revenue in 2023. We expect the N3 revenue in 2023 to be higher than N5 revenue in its first year in 2020.

N3E will further extend our N3 family with enhanced performance, power, and yield and offer complete platform support for both smartphone and HPC applications. Volume production is scheduled for second half '23.

Despite the ongoing inventory correction, we continue to observe a high level of customer engagement at both N3 and N3E with [the] number of tape-outs more than 2x that of N5 in its first and second year."

Bear in mind that TSMC recognized N5 revenue for both the A14 and M1 in 2020, a year when Apple posted an all-time record calendar Q4. TSMC's net revenue from N5 in 2020 was around $3.64 billion USD.

N3E was planned to follow N3 by one year. Even if TSMC can pull it in by 3 months, they'd only get a single quarter of revenue from it in 2023, and products based on N3E chips would not make it to shelves this year. N3 is higher complexity, higher cost, and higher risk than N3E, but it's also ready now.

And for reference:

	N5P vs N5	N4 vs N5P	N3 vs N4	N4P vs N4	N3E vs N3
Logic Density	same​	1.06x​	1.6x​	same​	0.94x​
SRAM Density	same​	same​	1.05~1.2x​	same​	0.83~0.95x​
Speed at Same Power	+5%​	same​	+5~10%​	+6%​	+5%​
Power at Same Speed	-15%​	same​	-13~18%​	-9%​	-7%​
Product Tape-out				H2'22​
Risk Production	Q2'20​	Q3'21​	Q4'21​		Q4'22​
Volume Production	H1'21​	H1'22​	H2'22 (Dec)​	H1'23?​	H2'23 (Sep~Oct)​

edit: Fixed table to correct minor errors and add additional data; confirmed all data to the best of my ability against most recent public statements issued by TSMC.

 

[Eug]

M2 Pro gets 1950/15000 in Geekbench 5, at 3.5 GHz:

Mac14,12 - Geekbench Browser

Benchmark results for a Mac14,12 with an Apple M2 Pro processor.

browser.geekbench.com

Not sure if this is a MacBook Pro or a Mac mini, but regardless, it's impressive we can get this level of performance in a Mac mini now. This is literally twice as fast multi-core as the M1 Mac mini I bought last year, and 4 times as fast multi-core as my 27" Core i5-7600 iMac 2017 sitting beside it.

I'm still trying to figure out what to do with the iMac. I was planning on giving it to my kid, but she just wants to stick with the 2015 13" MacBook Pro I got for her a year ago. The decline in popularity of the desktop for mainstream users (and many pro users too) continues...

 

[IntelUser2000]

Is that 200/400GB for RAM number the same as AIDA numbers for Zen4 people posted in the MaxxMem2 thread, which were about 60+GB? In other words is Apple architecture this superior to x86?

That has nothing to do with the ISA. The Pro has 256-bit memory interface while the Max has a 512-bit memory interface. In contrast "dual-channel" setups typically used in PCs is 128-bit.

Also, based on Anandtech's tests of the M1 devices, it doesn't always reach the quoted bandwidth. Likely because that's the theoretical figure.

512 bits x 8 bits/byte x 6400MT/s = 409.6GB/s which corresponds to the 400GB/s+ claim.

And raising width isn't free. Each bit requires extra wire routing on the board.

 

[name99]

M2 Pro gets 1950/15000 in Geekbench 5, at 3.5 GHz:

Mac14,12 - Geekbench Browser

Benchmark results for a Mac14,12 with an Apple M2 Pro processor.

browser.geekbench.com

Not sure if this is a MacBook Pro or a Mac mini, but regardless, it's impressive we can get this level of performance in a Mac mini now. This is literally twice as fast multi-core as the M1 Mac mini I bought last year, and 4 times as fast multi-core as my 27" Core i5-7600 iMac 2017 sitting beside it.

I'm still trying to figure out what to do with the iMac. I was planning on giving it to my kid, but she just wants to stick with the 2015 13" MacBook Pro I got for her a year ago. The decline in popularity of the desktop for mainstream users (and many pro users too) continues...

Here's a nicer (or at least richer context) version of the same GB5 result:

Mac14,12 vs Mac Studio - Geekbench Browser

browser.geekbench.com

 

[Timmah!]

That has nothing to do with the ISA. The Pro has 256-bit memory interface while the Max has a 512-bit memory interface. In contrast "dual-channel" setups typically used in PCs is 128-bit.

Also, based on Anandtech's tests of the M1 devices, it doesn't always reach the quoted bandwidth. Likely because that's the theoretical figure.

512 bits x 8 bits/byte x 6400MT/s = 409.6GB/s which corresponds to the 400GB/s+ claim.

And raising width isn't free. Each bit requires extra wire routing on the board.

Great. Thanks for explanation!

 

[repoman27]

Well, if that is the case, Apple. could start buying risk wafers of their customized N3 process and have enough good A17 SoCs for at least the iPhone Pro models - they've done this before (essentially funding the ramp to HVM). If yields improve enough - that process could be used in M3 SoCs. The geometric and electrostatic gains for that node are good enough.

I'm not singling you out in any way, your post is just the most recent one to mention the potential use of risk starts. When it comes to their official communications, TSMC tries very hard to reference certain milestones in a consistent manner, because they tend to have a material impact on financial results. A few of the commonly mentioned milestones include:

• design solution development
• customer design start
• customer product tape-out
• risk production
• volume production

Volume production in TSMC parlance is critical because it indicates the beginning of revenue recognition for that manufacturing process. TSMC does not recognize revenue for a node until after reaching volume production (i.e. they have fully qualified customer designs, good wafer outs that they can get paid for, and are beginning the volume ramp in earnest). This milestone is similar in some ways to PRQ for an Intel chip design.

Ultimately, Apple doesn't pay TSMC for wafer starts, they pay them for wafer outs. TSMC is obligated to report that revenue, which in turn pushes the node past the volume production milestone. Due to the way TSMC defines risk and volume production, it is impossible for TSMC to meet Apple's iPhone SoC requirements using risk starts alone. Furthermore, all of this is carefully coordinated and planned years in advance. That's why TSMC was able to tell investors three years ago that N3 would reach volume manufacturing in H2'22, but not to expect revenue from N3 until 2023 due to the long cycle times.

For their size and complexity, Apple's A-series chips are some of the highest volume SoCs manufactured. Even with the A16 only being in the iPhone 14 Pro and 14 Pro Max, Apple still required tens of millions of finished chips well in advance of the September launch. Apple is TSMC's largest customer, potentially representing around 25% of their total revenue. iPhone is Apple's largest product category, potentially representing around 50% of their total revenue. From a financial perspective, it's no exaggeration to say that the SoC for a given year's halo iPhone is the most important bit of silicon that TSMC produces.

edit: grammah

 

[Eug]

I'm still surprised but for my edu pricing, they dropped prices even more.

My M1 16 GB / 1 TB was CA$1529 (US$1135) edu, although I paid CA$1099 (US$816) used.
The M2 16 GB / 1 TB is CA$1369 (US$1017) edu. This is US$1059 on the US edu store.

That's why I bought my M1 last year. I had assumed it would still be at least $1529 for that configuration, and it wouldn't come out until spring.

People had been predicting price increases, and even keeping the M1 as the base model with the M2 as the higher end model at a higher price, but Apple did the opposite. I thought I was being optimistic predicting pricing in Canada would stay the same.

 

[poke01]

I'm still surprised but for my edu pricing, they dropped prices even more.

My M1 16 GB / 1 TB was CA$1529 (US$1135) edu, although I paid CA$1099 (US$816) used.
The M2 16 GB / 1 TB is CA$1369 (US$1017) edu. This is US$1059 on the US edu store.

That's why I bought my M1 last year. I had assumed it would still be at least $1529 for that configuration, and it wouldn't come out until spring.

People had been predicting price increases, and even keeping the M1 as the base model with the M2 as the higher end model at a higher price, but Apple did the opposite. I thought I was being optimistic predicting pricing in Canada would stay the same.

Honestly, M2 Mac Mini is good. HDMI 2.1 and 6E is even on the base $599 model.

 

[poke01]

and many pro users too) continues...

Not Gaming and the DIY desktop sector. That's on the rise.

 

[Ajay]

Not Gaming and the DIY desktop sector. That's on the rise.

Neither of which apply to to Macs.

 

[Eug]

Honestly, M2 Mac Mini is good. HDMI 2.1 and 6E is even on the base $599 model.

HDMI 2.1 doesn’t mean much in itself because any new product that is HDMI 2.0 complaint is now officially called HDMI 2.1. HDMI 2.0 no longer exists.

When HDMI 2.1 Isn't HDMI 2.1 - The Confusing World of the Standard, "Fake HDMI 2.1" and Likely Future Abuse - TFTCentral

When HDMI 2.1 Isn't HDMI 2.1 - The Confusing World of the Standard, "Fake HDMI 2.1" and Likely Future Abuse

tftcentral.co.uk

For example, the M2 Mac mini only supports 4K 60 Hz over HDMI. In fact, neither my M1 Mac mini nor the new M2 Mac mini can support a 3840x2560 monitor at 60 Hz using HDMI.

In contrast, the M2 Pro Mac mini supports 4K 240 Hz or 8K 60 Hz over HDMI.

 

[poke01]

In contrast, the M2 Pro Mac mini supports 4K 240 Hz or 8K 60 Hz over HDMI.

Ahh yes. I meant the HDMI version that can do 8k 60 or 4k 120hz. IF Apple put fake HDMI version on a $1999 Mac then heads will roll.