Discussion Apple Silicon SoC thread

[Eug]

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

 

[Doug S]

That makes sense. But where has Apple said it? I've only seen rumors from Taiwan saying that.

Yes so far it is only rumors, and the rumors are probably based on timing seen from M2 rumors. Kind of hard to extrapolate with a sample size of 2, especially when one of those is just a rumor itself! The startup of Apple Silicon designs may be at a different cadence than things go after they've got the first couple generations under their belt. Plus we can only see when things are released, we can't see when they were PLANNED to have been released.

Since Apple isn't announcing dates for this stuff if something slips by 3-6 months because they need to do a few mask revisions or whatever we'll see the timing and can say "they followed M1 with M1P/M1M by x months, and M2 by y months" but guesses on cadence based on that would be incorrect if the release dates had slipped due to mask revisions, covid issues, etc. or they were planned to be different at first as they got the Apple Silicon design pipeline rolling. With iPhone it is pretty obvious when something slips, because we know when to expect new models.

It will take a while to gain the same comfort level when new MBA or MBP or Mac Pro or whatever comes out - and it is possible they never set things in stone to the level they have with iPhone. And like iPhone, we very likely won't know (except via rumor) if the reason it is late has to do with the SoC or one of many other components. I mean, Steve Jobs famously delayed the white iPhone 4 by almost six months because the white in the home button didn't match the white on the case (different materials) to his level of satisfaction, so a delay could be for not only any reason you can think of but one you might never consider.

An 18 month cadence doesn't seem to make a lot of sense when the cores are on a 12 month design cycle. Though A15 doesn't seem have to got much than a couple minor tweaks and a process based clock bump compared to A14 so it remains to be seen if that's an aberration or they are going to move to a two year cadence.

Regardless of whether they keep the yearly cadence for iPhone SoCs or go to two years, splitting the difference at 18 months for Apple Silicon would be a really odd choice unless there were some parts of it that will be on a three year cadence. If they were basically saying "OK, M2 will be the first one to support 4x SoCs in a Mac Pro, and M3 will keep all the Apple Silicon only stuff like the interchip fabric pretty much unchanged and then do a ground up redesign with M4" then you could have phone SoCs at a one year cadence and major upgrades to the Mac SoCs every three, with an incremental bump with new cores and a process based clock bump at the halfway point so customers don't have to wait three years for a new model.

 

[Heartbreaker]

An 18 month cadence doesn't seem to make a lot of sense when the cores are on a 12 month design cycle. Though A15 doesn't seem have to got much than a couple minor tweaks and a process based clock bump compared to A14 so it remains to be seen if that's an aberration or they are going to move to a two year cadence.

I expect there won't be any kind of even set cadence. They will update in spurts, when they feel it necessary.

You can't blame all the Mac neglect on Intel in the x86 period. I expect Mini and Desktop Pro in particular to fall back to a fairly slow update schedule.

 

[Doug S]

I expect there won't be any kind of even set cadence. They will update in spurts, when they feel it necessary.

You can't blame all the Mac neglect on Intel in the x86 period. I expect Mini and Desktop Pro in particular to fall back to a fairly slow update schedule.

I doubt it. They are already getting the CPU cores, GPU cores, NPU cores, etc. etc. redesigned every year for "free", the cost of doing the extra that the single SoC Macs need is quite low. I wouldn't be shocked if they ended up with a yearly cadence. Why not, when all they have to do is cut and paste parts from the A* into the larger M*? The only complicated thing part of it is the interchip fabric, and which is where I could see them going slower (just like Intel, AMD etc. are not redesigning theirs from the ground up with every new core)

If it doesn't cost them much to roll a new Mn+1 SoC and stick it in the exact same MBA or Mini, why not? Like with the iPhone, the ground up redesigns of the Mac hardware aside from the SoC will proceed slowly but what exactly would be the benefit of say rolling a new Mini or even MBA every year? What exactly are they going to change in it on that timescale that would make it worth bothering?

 

[Heartbreaker]

I doubt it. They are already getting the CPU cores, GPU cores, NPU cores, etc. etc. redesigned every year for "free", the cost of doing the extra that the single SoC Macs need is quite low.

It's not low at all. They still need to all new tapeout, all new set of masks. It's easily over $100 Million these days.

Also CPU core updates aren't really happening every year. A14 and A15 main CPU cores essentially have the same performance. So it would be a waste to do a new SoC for that core.

Also even if the build new M chips, I could see them just ignore the Mini for a couple of years, even if there is a new M2 going into laptops.

They definitely aren't going to update all Macs, every year. That was never going to happen.

 

[Doug S]

It's not low at all. They still need to all new tapeout, all new set of masks. It's easily over $100 Million these days.

Also CPU core updates aren't really happening every year. A14 and A15 main CPU cores essentially have the same performance. So it would be a waste to do a new SoC for that core.

Also even if the build new M chips, I could see them just ignore the Mini for a couple of years, even if there is a new M2 going into laptops.

They definitely aren't going to update all Macs, every year. That was never going to happen.

They used the same design for the M1 Pro and M1 Max, with the former just a cut down version of the latter. From what I understand that allows them to use the same mask set. So they are amortizing that $100 million across all units that use the M1 Pro or M1 Max, so the fixed cost allocation of that $100 million per Mac containing a M1P/M1M is like $40, and all the Macs containing them sell for over $1000. Cost is not an issue.

They won't update all Macs every year, but it won't be due to cost reasons. The Apple Silicon SoCs cost them less than the Intel CPUs they are replacing - and they are getting the GPU thrown in for free. Even with the LPDDR5 costing more they are paying far less for the CPU, GPU and DRAM in Apple Silicon Macs versus Intel Macs. And its not even close.

The A15 is the first SoC that didn't get a core update (on the big core, the little core was updated) It remains to be seen if that's the new normal or if it was just a matter of engineering resource/timing as they were busy launching the Apple Silicon SoCs.

 

[Eug]

M1 Ultra is here!

5 nm
M1 Max die x 2 + UltraFusion die-to-die interconnect -
2.5 TB/s.
128 GB unified RAM
800 GB/s memory bandwidth
114 billion transistors
16 high performance cores
4 efficiency cores
64 GPU cores

 

[Oyeve]

M1 Ultra is here!

5 nm
M1 Max die x 2 + UltraFusion die-to-die interconnect -
2.5 TB/s.
128 GB unified RAM
800 GB/s memory bandwidth
114 billion transistors
16 high performance cores
4 efficiency cores
64 GPU cores

yawn.

 

[Panino Manino]

"Glue" is the biggest invention in tech in this century.

What Apple is doing is "smart"?
Don't they end with chips with duplicated hardware that ends wasted?

EDIT: about the GPU, apparently is vs 3070ti. Both are 21TF GPUs, but 330 gigapixels + 660 gigatextels vs 170 gigapixels + 340 gigatextels.

 

[jpiniero]

yawn.

What, not enough for ya? I wonder if this is the same die as the M1 Max or just altered to include the interconnect. The rumor did say they were going to do a 4 die version.

Don't they end with chips with duplicated hardware that ends wasted?

For now yeah. That's something that will need to be worked on.

Edit: At $3999 for the M1 Ultra model Apple can afford it.

 

[Eug]

Mac Studio
Aluminum - Looks like a tall Mac mini
M1 Max, M1 Ultra
7.7" x 3.7"
128 GB RAM (with M1 Ultra)
8 TB SSD
6 Thunderbolt 4 (2 in front), 2 USB-A
HDMI
10 Gigabit Ethernet (standard)
Audio jack
Front SD card reader
Built in power supply
Starts at US$1999

Studio Display
27" 5K Retina
A13 SoC
All aluminum
12 MP camera with Centre Stage (no Face ID)
3-array mics
Lots of speakers with Spatial Audio
Tilting, or optional tilt plus height
600 nits
P3 wide colour
1 billion colours
True Tone
Anti-reflective coating, or nano-texture option
3 USB-C ports, and 1 ThunderBolt 4
Starts at US$1599

Mac Pro is still in the pipeline

 

[insertcarehere]

What, not enough for ya? I wonder if this is the same die as the M1 Max or just altered to include the interconnect. The rumor did say they were going to do a 4 die version.

The speculation when the M1 Max was launched was that the die layout Apple presented had unnecessary duplicated function blocks for what's being touted, and was a modified image to hide the chip interconnect that was on the die all along. It seems that speculation is being proven correct.

From the perf-power curves being presented (35w for M1 Max, 60w for M1 Ultra) , it seems like the chips are not being clocked higher than their counterparts in the MBPs despite being (what looks like) a beefier chassis with much more cooling potential. Perhaps there are limitations elsewhere.

 

[soresu]

M1 Ultra is here!

5 nm
M1 Max die x 2 + UltraFusion die-to-die interconnect -
2.5 TB/s.
128 GB unified RAM
800 GB/s memory bandwidth
114 billion transistors
16 high performance cores
4 efficiency cores
64 GPU cores

..... Why must they all list GPU core count instead of TFLOPS?

It's really hard to compare mobile GPUs from these arbitrary core count specs.

Even when the FPS p/TFLOP value differs between GPU µArchs it is still a good rule of thumb to have.

 

[Saylick]

The speculation when the M1 Max was launched was that the die layout Apple presented had unnecessary duplicated function blocks for what's being touted, and was a modified image to hide the chip interconnect that was on the die all along. It seems that speculation is being proven correct.

Right...

https://twitter.com/x/status/1466526403331952644

 

[Mopetar]

Don't they end with chips with duplicated hardware that ends wasted?

Not necessarily. It does let Apple use any chips that would have had defects in any of those areas that would be duplicated.

If any of that stuff couldn't be disabled due to being redundant then it means Apple can get use out of a die that would otherwise be useless.

 

[ashFTW]

Ordered the M1 Ultra with 48 GPU cores, 4TB SSD and 128GB memory. And the new 5K monitor with height adjustable stand. 😊

 

[The Hardcard]

The speculation when the M1 Max was launched was that the die layout Apple presented had unnecessary duplicated function blocks for what's being touted, and was a modified image to hide the chip interconnect that was on the die all along. It seems that speculation is being proven correct.

From the perf-power curves being presented (35w for M1 Max, 60w for M1 Ultra) , it seems like the chips are not being clocked higher than their counterparts in the MBPs despite being (what looks like) a beefier chassis with much more cooling potential. Perhaps there are limitations elsewhere.

Right...

https://twitter.com/x/status/1466526403331952644

Actually, pictures I’ve seen show the redundant blocks AND the interconnect.

 

[Paratus]

Am I wrong or does a 128GB pool of unified memory running at 800 GB/s have to be using a GDDR6 / 512bit bus or an HBM2 / 4096bit bus?

 

[Ajay]

Studio Display
27" 5K Retina
A13 SoC
All aluminum
12 MP camera with Centre Stage (no Face ID)
3-array mics
Lots of speakers with Spatial Audio
Tilting, or optional tilt plus height
600 nits
P3 wide colour
1 billion colours
True Tone
Anti-reflective coating, or nano-texture option
3 USB-C ports, and 1 ThunderBolt 4
Starts at US$1599

Ok, that's strange. A13 SOC???

 

[Paratus]

Ok, that's strange. A13 SOC???

It apparently handles the cameras and other features of the monitor

 

[Saylick]

Am I wrong or does a 128GB pool of unified memory running at 800 GB/s have to be using a GDDR6 / 512bit bus or an HBM2 / 4096bit bus?

If I recall correctly, M1 Pro and M1 Max have a 256-bit LPDDR5 bus, so you would be right that the M1 Ultra, which is two M1 Maxes, would have a 512-bit bus.

Edit: @Mopetar corrected me. Only the M1 Pro has a 256-bit bus. The M1 Max doubles that again with two more memory controllers, so the M1 Ultra has 8 128-bit buses for a whopping 1024-bit interface.

 

[Mopetar]

Am I wrong or does a 128GB pool of unified memory running at 800 GB/s have to be using a GDDR6 / 512bit bus or an HBM2 / 4096bit bus?

It's the same memory as used in the MacBook Pro models with the M1 Pro/Max SoCs. So LPDDR5 memory if I recall correctly.

 

[jpiniero]

BTW it sounds too much like "MK Ultra"

 

[Paratus]

If I recall correctly, M1 Pro and M1 Max have a 256-bit LPDDR5 bus, so you would be right that the M1 Ultra, which is two M1 Maxes, would have a 512-bit bus.

It's the same memory as used in the MacBook Pro models with the M1 Pro/Max SoCs. So LPDDR5 memory if I recall correctly.

Remind me. Does LPDDR5 double the bandwidth over the same speed (LP)DDR4?

I get 102GB/s with a 256bit bus and 3200mhz DDR4. The M1 Ultra would need to double the bandwidth 3 times - going to 512bit is only a single doubling. To get to 800GB/s would take a 512bit bus and either 12,500mhz memory or double the bandwidth memory running at 6250mhz

 

[Eug]

So, this is my thinking for their final headless lineup:

M1 Mac mini
M1 Pro/Max Mac mini (to be released in fall)
M1 Max/Ultra Mac Studio
M1 Ultra/Ultra Duo Mac Pro (to be released at WWDC)

Ok, that's strange. A13 SOC???

Well, all monitors have an SoC, so I guess they decided to use their own (much faster and much more capable) SoC.

 

[jpiniero]

Remind me. Does LPDDR5 double the bandwidth over the same speed (LP)DDR4?

I get 102GB/s with a 256bit bus and 3200mhz DDR4. The M1 Ultra would need to double the bandwidth 3 times - going to 512bit is only a single doubling. To get to 800GB/s would take a 512bit bus and either 12,500mhz memory or double the bandwidth memory running at 6250mhz

It's 1024 bit @ 6200.