Discussion Apple Silicon SoC thread

Eug · Nov 10, 2020

M1
5 nm
Unified memory architecture - LP-DDR4
16 billion transistors

8-core CPU

4 high-performance cores
192 KB instruction cache
128 KB data cache
Shared 12 MB L2 cache

4 high-efficiency cores
128 KB instruction cache
64 KB data cache
Shared 4 MB L2 cache
(Apple claims the 4 high-effiency cores alone perform like a dual-core Intel MacBook Air)

8-core iGPU (but there is a 7-core variant, likely with one inactive core)
128 execution units
Up to 24576 concurrent threads
2.6 Teraflops
82 Gigatexels/s
41 gigapixels/s

16-core neural engine
Secure Enclave
USB 4

Products:
$999 ($899 edu) 13" MacBook Air (fanless) - 18 hour video playback battery life
$699 Mac mini (with fan)
$1299 ($1199 edu) 13" MacBook Pro (with fan) - 20 hour video playback battery life

Memory options 8 GB and 16 GB. No 32 GB option (unless you go Intel).

It should be noted that the M1 chip in these three Macs is the same (aside from GPU core number). Basically, Apple is taking the same approach which these chips as they do the iPhones and iPads. Just one SKU (excluding the X variants), which is the same across all iDevices (aside from maybe slight clock speed differences occasionally).

EDIT:

M1 Pro 8-core CPU (6+2), 14-core GPU
M1 Pro 10-core CPU (8+2), 14-core GPU
M1 Pro 10-core CPU (8+2), 16-core GPU
M1 Max 10-core CPU (8+2), 24-core GPU
M1 Max 10-core CPU (8+2), 32-core GPU

M1 Pro and M1 Max discussion here:

Page 78 - Discussion - Apple Silicon SoC thread

Page 78 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

M1 Ultra discussion here:

Page 109 - Discussion - Apple Silicon SoC thread

Page 109 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

M2 discussion here:

Page 127 - Discussion - Apple Silicon SoC thread

Page 127 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

Second Generation 5 nm
Unified memory architecture - LPDDR5, up to 24 GB and 100 GB/s
20 billion transistors

8-core CPU

4 high-performance cores
192 KB instruction cache
128 KB data cache
Shared 16 MB L2 cache

4 high-efficiency cores
128 KB instruction cache
64 KB data cache
Shared 4 MB L2 cache

10-core iGPU (but there is an 8-core variant)
3.6 Teraflops

16-core neural engine
Secure Enclave
USB 4

Hardware acceleration for 8K h.264, h.264, ProRes

M3 Family discussion here:

Page 215 - Discussion - Apple Silicon SoC thread

Page 215 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

M4 Family discussion here:

Page 263 - Discussion - Apple Silicon SoC thread

Page 263 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

repoman27 · Tuesday at 12:08 AM

dr1337 said:
Because all the probably dozens of usb c cables I have are type A to type C except for one really nice type C cable that I always leave at home. And I only even have it for the couple times I wanted to try display output. I had a USB-C android phone for a bit and it legit came with type A to type C only and I cant for the life of me see why it would be worth buying new cables instead of dongles or hubs.

Meanwhile, I only recently acquired an A to C cable, because I wanted to back up some older computers that didn't have Type-C ports to a portable SSD and didn't trust the dodgy, not even remotely to spec adapter that came packed in with the SSD. How did you end up with a dozen A to C cables and only one C to C?

If you're just charging a phone, you can use a very inexpensive USB 2.0 Type-C cable. Next time you're at a hotel, ask if they have a charging cable you can borrow, and they'll probably let you pick from a box of hundreds of cables that were left behind in rooms. Some places will let you just take a handful and keep them.

Good cables are far more reliable and often a lot cheaper than dongles. We're not talking about active dongles that provide additional functionality here. USB Type-C to Type-A legacy adapters are simple electro-mechanical adapters and are primarily intended for situations where such an adapter is actually necessary. Combining adapters and cables to make frankencables instead of just using the right cable for the job is really not the way to go.

name99 · Tuesday at 4:52 PM

Eug said:
Would you hazard to guess the I/O on an M4 (non-Pro) Mac mini from the die shot?

This is par for the course for Apple. It has often done this with the Mac minis and with other Macs, for the upsell. The only way the M2 Pro Mac mini makes any sense is if you don't max it out.

The main drawback with the M2 Pro Mac mini is the fact the only memory upgrade option available from its 16 GB base is straight to 32 GB. I'm sure a lot of people would like to have a 24 GB option on the M2 Pro, like Apple offers with the M2 Mac mini.

M2 Mac mini
24 GB RAM
512 GB storage
Thunderbolt 4 ports x 2, USB-A ports x 2 (Total 4 USB-C ports)
US$1199

M2 Pro Mac mini (10/16-core)
32 GB RAM
512 GB storage
Thunderbolt 4 ports x 4, USB-A x 2 (Total 6 USB-C ports)
US$1699

M2 Max Mac Studio (12/30-core)
32 GB RAM
512 GB storage
Thunderbolt 4 ports x 4, USB-C ports x 2, USB-A ports x 2 (Total 8 USB-C ports)
US$1999

We'll have to see how the pricing plays out with the M4 series, with 12 GB RAM base (my prediction), and four USB4/Thunderbolt controllers.

And 10Gbps ethernet... That reduces the difference by another $100.

name99 · Tuesday at 5:02 PM

repoman27 said:
I will.

It'll probably look very similar to the current M2 Pro Mac mini. The only difference will be that the SSD interface will be limited to the equivalent of two PCIe Gen4 lanes and probably 2GB capacity due to the lower number of channels.

I don't see any reason why Apple would omit the USB Type-A ports, which have always been implemented via a discrete controller. Or why they wouldn't expose all four Thunderbolt 4 ports.

So general belief is still PCIe4 for M4?

Is there any advantage for APPLE in particular (not <insert random other use case>) of moving to PCIe5?
Obviously we can then pump twice as much data over each wire (or more specifically, can halve the wiring devoted to PCIe).
On the other hand, we will probably have to increase the area devoted to PCIe logic to get that doubled speed.

Is there an easy rule of thumb for how this plays out in terms of overall area and power? And is there any other functionality in PCIe5 apart from the speed that would make it worth transitioning?
Given that Apple doesn't do slots (apart from the Mac Pro, and who knows long term how that will play out) I have very little intuition for what makes sense (business and technical) for Apple's IO path going forward. Presumably what the user will see is a succession of TB5 then USB5, and so on; but internally I have no idea what the optimal protocols are to drive these given Apple's concerns [presumably power first, area second].

name99 · Tuesday at 5:04 PM

dr1337 said:
Because all the probably dozens of usb c cables I have are type A to type C except for one really nice type C cable that I always leave at home. And I only even have it for the couple times I wanted to try display output. I had a USB-C android phone for a bit and it legit came with type A to type C only and I cant for the life of me see why it would be worth buying new cables instead of dongles or hubs.

I mean maybe not on the mac mini but type A has been long dead in laptops, and not a thing in the imac anymore. Also anyone trying to use a brand new ipad pro with peripherals implicitly has to use a dock or adapter of some kind to use more than one device, or use wall power. Its actually kinda crazy to me thinking now that the M4 with all of the connectivity it has gets put into an ipad with a single USB-C port.

Converters cost like $2.50. AMZ is a thing.
Complaining about these is the most boring thing in the world.

poke01 · Tuesday at 7:50 PM

Eug said:
Remember, as of next year, when I expect the M4 Mac mini will be released, USB-A will be 29 years old

The headphone jack is even older, will Apple drop that on Mac’s?

name99 said:
So general belief is still PCIe4 for M4?

Is there any advantage for APPLE in particular (not <insert random other use case>) of moving to PCIe5?
Obviously we can then pump twice as much data over each wire (or more specifically, can halve the wiring devoted to PCIe).
On the other hand, we will probably have to increase the area devoted to PCIe logic to get that doubled speed.

Is there an easy rule of thumb for how this plays out in terms of overall area and power? And is there any other functionality in PCIe5 apart from the speed that would make it worth transitioning?
Given that Apple doesn't do slots (apart from the Mac Pro, and who knows long term how that will play out) I have very little intuition for what makes sense (business and technical) for Apple's IO path going forward. Presumably what the user will see is a succession of TB5 then USB5, and so on; but internally I have no idea what the optimal protocols are to drive these given Apple's concerns [presumably power first, area second].

unless Apple plans support for the RTX 5090 for the Mac Pro or enable eGPU support for MacBooks , I don’t see what PCIe 5 will do but increase cost for Apple.

If they wanted PCIe5 support we would have with M4 because we are actually getting m4 ultras this generation. Maybe we will see pcie5 with pro/max.

Eug · Tuesday at 8:56 PM

poke01 said:
The headphone jack is even older, will Apple drop that on Mac’s?

Not really an equivalent example. USB-A is an old standard that has been superseded by USB-C.

The headphone jack is an even older standard, but nothing has superseded it. But even then, some devices like iPads and iPhones no longer have headphone jacks as you know.

Fortunately, both USB-A and headphone jacks are available as adapters from USB-C. Furthermore, while you can use USB-C to create full spec USB-A and a full spec headphone jack, you cannot use USB-A or a headphone jack to create full spec USB4. My favourite Mac form factor is the 12" MacBook, which I still own. However, it only has two ports, with one of them being a headphone jack and the other being USB-C. I would have MUCH rather have had two USB-C ports, which would have allowed me much more flexibility, including using one as a headphone jack via a <$10 adapter if I wished.

Anyhow, this is not really about what I want to happen, but about what I expect to happen. I often wondered when and how Apple would bring four USB-C ports to the Mac mini. The info that M4 has four Thunderbolt controllers with a not insignificant use of die area, leads one to think the Mac mini will get 4 Thunderbolt ports, and my expectation also was that Apple would just kill off USB-A when it increased the number of USB-C ports. (However, my previous expectation before we saw the die shot was not four Thunderbolt ports, but instead two Thunderbolt ports and two USB-C ports utilizing a separate USB controller.) Ultimately though, we're both just guessing.

However, if Apple were to decide that the base Mac mini would get four USB-C and two USB-A, then even better. The only reason I would consider the Mx Pro Mac mini in the first place is because it has four Thunderbolt and two USB-A.

Doug S · Tuesday at 8:59 PM

poke01 said:
unless Apple plans support for the RTX 5090 for the Mac Pro or enable eGPU support for MacBooks , I don’t see what PCIe 5 will do but increase cost for Apple.

If they wanted PCIe5 support we would have with M4 because we are actually getting m4 ultras this generation. Maybe we will see pcie5 with pro/max.

I agree. Apple could implement PCIe 5.0, but they have to ask themselves, what is the use case. Who is asking for this? Most of the people using add-in cards on Mac Pro are using video stuff that's still PCIe 3.0. If they aren't even using 4.0, why go 5.0?

I guess there are SSDs, but you're in some pretty rarified air for an SSD to actually NEED PCIe 5.0 vs 4.0, and even then that's only if your load is lots of multi multi gigabyte sequential reads and writes. Something like GPGPU/AI, there aren't any Mac drivers even for the stuff that uses 4.0 so that's hardly a case for 5.0.

Honestly they'd be better off pushing implementations of new TB standards as quickly as they can, as there are some tangible benefits there since that upgrade is applicable to a much wider range of Macs instead of just the Mac Pro. The faster that gets the more it pushes off things that used to be PCIe cards, like 40/100GbE or fibre channel HBAs.

poke01 · Tuesday at 9:28 PM

Eug said:
Not really an equivalent example. USB-A is an old standard that has been superseded by USB-C.

The headphone jack is an even older standard, but nothing has superseded it. But even then, some devices like iPads and iPhones no longer have headphone jacks as you know.

Fortunately, both USB-A and headphone jacks are available as adapters from USB-C. Furthermore, while you can use USB-C to create full spec USB-A and a full spec headphone jack, you cannot use USB-A or a headphone jack to create full spec USB4. My favourite Mac form factor is the 12" MacBook, which I still own. However, it only has two ports, with one of them being a headphone jack and the other being USB-C. I would have MUCH rather have had two USB-C ports, which would have allowed me much more flexibility, including using one as a headphone jack via a <$10 adapter if I wished.

Anyhow, this is not really about what I want to happen, but about what I expect to happen. I often wondered when and how Apple would bring four USB-C ports to the Mac mini. The info that M4 has four Thunderbolt controllers with a not insignificant use of die area, leads one to think the Mac mini will get 4 Thunderbolt ports, and my expectation also was that Apple would just kill off USB-A when it increased the number of USB-C ports. (However, my previous expectation before we saw the die shot was not four Thunderbolt ports, but instead two Thunderbolt ports and two USB-C ports utilizing a separate USB controller.) Ultimately though, we're both just guessing.

However, if Apple were to decide that the base Mac mini would get four USB-C and two USB-A, then even better. The only reason I would consider the Mx Pro Mac mini in the first place is because it has four Thunderbolt and two USB-A.

Unlike USB-A, a jack is a must for audio pros and this is why I think it’s never leaving the Mac’s. If it survived the 2016 MacBook era then it’s staying for the long run.

Good analysis, let’s see what the M4 mini looks like.

LightningZ71 · Tuesday at 11:32 PM

It sems that apple likes to keep their external lanes to a minimum onbtheir processors. Going with PCIe5 allows them essentially full data rate out of an SSD chip for only a single physical lane. I/O real estate on modern process tech SOCs is expensive as it doesn't shrink well.

Doug S · Wednesday at 1:02 AM

LightningZ71 said:
It sems that apple likes to keep their external lanes to a minimum onbtheir processors. Going with PCIe5 allows them essentially full data rate out of an SSD chip for only a single physical lane. I/O real estate on modern process tech SOCs is expensive as it doesn't shrink well.

Apple, the company that produced the M1 Max that has 10,000 I/Os to connect to another M1 Max, that Apple tries to keep external lanes to a minimum??

I think the fact they used 10,000 I/Os for 2.5 TB/sec is instructive. Obviously if you look at the interconnect that others are using they could have realized the same bandwidth with at least an order of magnitude fewer I/Os. They used a lot because they weren't pushing them that hard, so they didn't have to worry about signal integrity as much - and more importantly power draw per bit was reduced.

Signal integrity at 32 Gbps was enough of a problem with PCIe 5.0 that in order to double the speed again with PCIe 6.0 that the standards body had to switch from NRZ to PAM4 encoding. So while I can buy that Apple might possibly have reason to upgrade to PCIe 5.0 on its own merits, I don't buy that they'd upgrade just so they could save a few pins on the SoC.

igor_kavinski · Wednesday at 5:57 AM

Offtopic but posting this here because most Apple proponents visit this thread: http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=threads/apple-store-apps-dont-respect-your-privacy.2620241/

Feel free to leave your comments in the linked thread.

LightningZ71 · Wednesday at 8:49 AM

The I/O links between the two die on the M2 MAX are significantly smaller than the ones on a PCIe link. The distance that they are traveling, the likely current that they are pushing, and the lower signaling rate per pin allows them to be far denser with less circuitry dedicated to supporting each connection.

The other consideration is that, with a quarter of the die edge conserved for that connection, and half in use for the many DDR channels, there isn't a whole lot of border space left over for I/O! Of course they aren't going to go ham on PCIe lanes.

name99 · Wednesday at 2:53 PM

Doug S said:
I agree. Apple could implement PCIe 5.0, but they have to ask themselves, what is the use case. Who is asking for this? Most of the people using add-in cards on Mac Pro are using video stuff that's still PCIe 3.0. If they aren't even using 4.0, why go 5.0?

I guess there are SSDs, but you're in some pretty rarified air for an SSD to actually NEED PCIe 5.0 vs 4.0, and even then that's only if your load is lots of multi multi gigabyte sequential reads and writes. Something like GPGPU/AI, there aren't any Mac drivers even for the stuff that uses 4.0 so that's hardly a case for 5.0.

Honestly they'd be better off pushing implementations of new TB standards as quickly as they can, as there are some tangible benefits there since that upgrade is applicable to a much wider range of Macs instead of just the Mac Pro. The faster that gets the more it pushes off things that used to be PCIe cards, like 40/100GbE or fibre channel HBAs.

I thought I was very clear about this, but the issue is not PCIe CARDS! No-one gives a damn about those.
The issue is PCIe as an INTERNAL IO interconnect. And not primarily bandwidth (that's just replication, ie lane count), but the tradeoffs I mentioned in my post.

Doug S · Wednesday at 3:00 PM

igor_kavinski said:
Offtopic but posting this here because most Apple proponents visit this thread: http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=threads/apple-store-apps-dont-respect-your-privacy.2620241/

Feel free to leave your comments in the linked thread.

Is your entire life all about sh**ing on Apple? What a sad life that is.

igor_kavinski · Wednesday at 3:07 PM

Doug S said:
Is your entire life all about sh**ing on Apple? What a sad life that is.

I just report the facts. Don't discriminate. Wouldn't shy about posting something bad about AMD or Intel or any other company.

Eug · Wednesday at 4:11 PM

Eug said:
I know this is late to the game, but just for fun I ran Geekbench 6.3 (CPU) exactly once on my new 11" iPad Pro (M4), and got 3726 / 13289.

I tried the one-run bench on my binned 9-core M4 with Geekbench 3,4,5.

Geekbench 5: 2599 / 11396

Geekbench 4: 11486 / 43564

Geekbench 3: Cannot complete the install on this OS.

---

BTW, Bloomberg is sticking with the predicted M4 series Mac release dates:

Late 2024 to early 2025
iMac
MacBook Pro (low and high end)
Mac mini

2025 Q2
MacBook Air

2025 H2
Mac Studio
Mac Pro

It still just seems a bit odd to me that they've released M4 last month in the iPad Pro and then would wait another 2/3 of a year or so to release it in Macs.

For reference, here are the dates of the last Mac releases:

2024-03: M3 MacBook Air
2023-10: M3 iMac
2023-10: M3, M3 Pro, M3 Max MacBook Pro
2023-06: M2 Ultra Mac Pro
2023-06: M2 Ultra Mac Studio
2023-01: M2, M2 Pro Mac mini

FlameTail · Wednesday at 5:01 PM

If M4 Mac Studio/Pro is going to come in 2025H2...

Then will M5 even come in 2025H2 at all?

The Hardcard · Wednesday at 7:37 PM

Eug said:
I tried the one-run bench on my binned 9-core M4 with Geekbench 3,4,5.

Geekbench 5: 2599 / 11396

View attachment 101942

Geekbench 4: 11486 / 43564

View attachment 101941

Geekbench 3: Cannot complete the install on this OS.

---

BTW, Bloomberg is sticking with the predicted M4 series Mac release dates:

Late 2024 to early 2025
iMac
MacBook Pro (low and high end)
Mac mini

2025 Q2
MacBook Air

2025 H2
Mac Studio
Mac Pro

It still just seems a bit odd to me that they've released M4 last month in the iPad Pro and then would wait another 2/3 of a year or so to release it in Macs.

For reference, here are the dates of the last Mac releases:

2024-03: M3 MacBook Air
2023-10: M3 iMac
2023-10: M3, M3 Pro, M3 Max MacBook Pro
2023-06: M2 Ultra Mac Pro
2023-06: M2 Ultra Mac Studio
2023-01: M2, M2 Pro Mac mini

They are close to having ideal machines for generative AI for basic consumers up through creative professionals. Recently people have been able to cluster together up to Mac Studios via Thunderbolt with linear scaling. Even with the current Max at 128 GB, that would give a 4 Studio cluster with Ultras 1 TB of GPU-accelerated, 800 GB/s RAM.

While slower than NVidia or AMD, it would cost less than $25K compared to nearly $350K for NVidia or $150K for AMD. Not to mention needing less than 600 watts for the Studios, versus 5000 to 8000 watts for the big dogs.

The guy who designed Google’s inference infrastructure now works for Apple. I’m hoping the delay in M4 Macs is for making the maximum available tweaking to mitigate some remaining weaknesses.

Apple can secure a sizable niche of the AI market, including datacenter use if they leverage their architecture correctly. As well, Apple laptops and desktops can be an order or two of magnitude more capable of running neural networks than anything Qualcomm, AMD, or Intel can provide.

I’m really hoping they are making the moves to lean into this.

Doug S · Wednesday at 9:40 PM

Eug said:
It still just seems a bit odd to me that they've released M4 last month in the iPad Pro and then would wait another 2/3 of a year or so to release it in Macs.

The only reason I can think of is that only the M4 is out, but the M4 Pro and Max aren't coming until the fall. They don't want to release only M4 versions of e.g. Mac Mini and MBP, while sticking with the M3 Pro version for the higher end, and MBA has just been updated (and they don't want it to be faster than MBP)

Otherwise it makes no sense at all.

TwistedAndy · Thursday at 3:43 AM

The Hardcard said:
They are close to having ideal machines for generative AI for basic consumers up through creative professionals. Recently people have been able to cluster together up to Mac Studios via Thunderbolt with linear scaling. Even with the current Max at 128 GB, that would give a 4 Studio cluster with Ultras 1 TB of GPU-accelerated, 800 GB/s RAM.

While slower than NVidia or AMD, it would cost less than $25K compared to nearly $350K for NVidia or $150K for AMD. Not to mention needing less than 600 watts for the Studios, versus 5000 to 8000 watts for the big dogs.

The guy who designed Google’s inference infrastructure now works for Apple. I’m hoping the delay in M4 Macs is for making the maximum available tweaking to mitigate some remaining weaknesses.

Apple can secure a sizable niche of the AI market, including datacenter use if they leverage their architecture correctly. As well, Apple laptops and desktops can be an order or two of magnitude more capable of running neural networks than anything Qualcomm, AMD, or Intel can provide.

I’m really hoping they are making the moves to lean into this.

Unfortunately, things are much more complicated. Apple Silicon devices with a sufficient amount of memory can run relatively big inference models (70-180B), but the performance will not be so great. It's much more efficient to use smaller and much faster 3-7B models.

If we consider training and other more complex ML tasks, nVidia cards are much better in terms of software, performance, and scalability. There are reasons why nVidia develops NVLink, dedicated networking hardware, and systems like NVL72. For example, the latest NVLink version offers 1,800 GB/s GPU-to-GPU bandwidth. Thunderbolt 4 offers only 4 GB/s.

In terms of software, nVidia has been developing CUDA for 17 years. There is a huge amount of documentation, examples, libraries, integrations, developer programs, etc. And a lot of stuff is available for free and can run on regular nVidia cards. AMD and Intel are trying to implement their own CUDA alternatives (ROCm and openAPI), but they are far behind nVidia in many aspects.

From a company perspective, nVidia is a much safer choice. Yes, NVL72 costs a few millions of dollars, but it's an investment. AMD and Intel platforms look worse in general, but in some cases, they can fit well.

Apple is not an option here at all. It offers an even more restricted ecosystem than nVidia, but without its benefits.

igor_kavinski · Thursday at 5:35 AM

Doug S said:
Otherwise it makes no sense at all.

They can't release new stuff when they have tons of unsold chips and possibly millions of base model laptops in their warehouses. Releasing M4 will make it harder for them to offload their inventory of M3 chips (and maybe even M2 too!). The only way they can avoid this is if they made some sort of deal with TSMC where they can signal to TSMC to immediately stop production and move to producing new chips, regardless of the impact that has on TSMC's bottom line because obviously stopping M3 chip production would mean nothing owed to TSMC for chips not made and then TSMC is on their own to find someone else to fill the fabs that were previously making the M3 chips. If Apple has a deal to make 50 million M3 chips, that's what they get or they have to pay some early termination fee to stop production and move to newer chip production on better process.

Eug · Thursday at 5:40 AM

igor_kavinski said:
They can't release new stuff when they have tons of unsold chips and possibly millions of base model laptops in their warehouses. Releasing M4 will make it harder for them to offload their inventory of M3 chips (and maybe even M2 too!). The only way they can avoid this is if they made some sort of deal with TSMC where they can signal to TSMC to immediately stop production and move to producing new chips, regardless of the impact that has on TSMC's bottom line because obviously stopping M3 chip production would mean nothing owed to TSMC for chips not made and then TSMC is on their own to find someone else to fill the fabs that were previously making the M3 chips. If Apple has a deal to make 50 million M3 chips, that's what they get or they have to pay some early termination fee to stop production and move to newer chip production on better process.

I don’t understand the reasons for Apple’s M4 release schedule delay after the iPad Pro release, but M3 supply chain management definitely isn’t the reason.

igor_kavinski · Thursday at 6:44 AM

Eug said:
but M3 supply chain management definitely isn’t the reason.

Kindly elaborate.

Eug · Thursday at 7:38 AM

igor_kavinski said:
Kindly elaborate.

Millions of MacBook Airs in inventory is not how they operate. Their model is largely just in time manufacturing (outside initial release). In fact, JIT is why Tim Cook is there in the first place.

As for the chips, the rough schedule is planned out far, far in advance, but they have the option to adjust up and down as needed. It’s not as if they released M3 and then all of a sudden at the last minute decided they couldn’t release M4 in Macs because they needed to protect M3 shipments due to them screwing themselves over in a bad TSMC deal.

Furthermore, even though Bloomberg says an M4 MacBook Air is coming, there is no obligation for them to release one. Apple could just as easily skip M4 for the MacBook Air, while releasing M4 series in other Macs. They’ve done this before - skip MBA chip generations even when suitable drop-in replacement chips were widely available.

The Hardcard · Thursday at 8:22 AM

TwistedAndy said:
Unfortunately, things are much more complicated. Apple Silicon devices with a sufficient amount of memory can run relatively big inference models (70-180B), but the performance will not be so great. It's much more efficient to use smaller and much faster 3-7B models.

If we consider training and other more complex ML tasks, nVidia cards are much better in terms of software, performance, and scalability. There are reasons why nVidia develops NVLink, dedicated networking hardware, and systems like NVL72. For example, the latest NVLink version offers 1,800 GB/s GPU-to-GPU bandwidth. Thunderbolt 4 offers only 4 GB/s.

In terms of software, nVidia has been developing CUDA for 17 years. There is a huge amount of documentation, examples, libraries, integrations, developer programs, etc. And a lot of stuff is available for free and can run on regular nVidia cards. AMD and Intel are trying to implement their own CUDA alternatives (ROCm and openAPI), but they are far behind nVidia in many aspects.

From a company perspective, nVidia is a much safer choice. Yes, NVL72 costs a few millions of dollars, but it's an investment. AMD and Intel platforms look worse in general, but in some cases, they can fit well.

Apple is not an option here at all. It offers an even more restricted ecosystem than nVidia, but without its benefits.

I’m specifically talking about inference for a reason, and it is because Macs can inference with large models at a decent pace and Thunderbolt is proving sufficient in some cases for linear scaling, so not so much of a bottleneck.

I did just notice that I Ieft out the point of my post, which was a hope and perspective on why M4-based Macs may be coming so many months later than M4 iPads. And not any knowledge on my part, just a hope that the potential market opportunities of reworking some aspects of the silicon to lean into neural network capabilities is part of why there is this abnormal delay.

For me the delay otherwise also makes no sense. You could argue that they are making sure they have enough N3E capacity for iPhones. But then why release the M4 iPads, knowing that they will lose some Mac sales given some people are going to try to hold out for M4 generation Macs.

Sumit Gupta, the Apple guy formerly on Google’s inference infrastructure team, found there benefits to using swarms of smaller chips for cloud inferencing and guided Google’s infrastructure in that direction. Now he is heading the Apple Intelligence Private Cloud Compute run through M2 Ultras which has TSMC banging out large numbers of these chips.

I can’t help but wonder and hope that the experience Gupta’s team with the M2 Ultras is influencing and possibly delaying M4 series Mac designs. To be sure, Apple can also make non Mac datacenter Apple Silicon, but such adjustments could also affect future Mac chips and I’m hoping it is a key reason why new Macs are still so far off.

Discussion Apple Silicon SoC thread

Lifer

Senior member

Senior member

Senior member

Senior member

Golden Member

Lifer

Platinum Member

Golden Member

Golden Member

Platinum Member

Lifer

Golden Member

Senior member

Platinum Member

Lifer

Lifer

Diamond Member

Member

Platinum Member

Member

Lifer

Lifer

Lifer

Lifer

Member