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

 

[Ajay]

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)

Hmm, seems like an M1 Max Mac mini would be pushing close to the price of the base M1 Max Mac Studio.

 

[Ajay]

Mac Pros with M1 Ultra and 2 x M1 Ultra?

Sure, so long M1 Ultras can work as a dual "socket" system.

 

[Eug]

Hmm, seems like an M1 Max Mac mini would be pushing close to the price of the base M1 Max Mac Studio.

Yes, but it would fill in the price gap. Plus, such a product lineup encourages the upsell: "If I spend only $300 more than a spec'd out high end Mac mini, I can get a Mac Studio, so maybe I'll just do that."

But then again, many people seem to think the high end Mac mini will simply disappear completely.

 

[ashFTW]

Sure, so long M1 Ultras can work as a dual "socket" system.

Maybe it will be dual M2 Ultra; it’s the end of the line for M1 based chips.

I'm thinking:

M2: 4P+4E, 10GPU
M2 Pro: 12P+2E, 20 GPU
M2 Max: 12P+2E, 40 GPU
M2 Ultra: 24P+4E, 80 GPU
M2 Grande: 48P+8E, 160 GPU

 

[Ajay]

But then again, many people seem to think the high end Mac mini will simply disappear completely.

That would be my guess - we'll see. The base Mac Studio Max is nicely equipped - except for the standard 512GB NVMe drive (and the $200 1TB upgrade). Sooo close to not being penny pinching jerks for once .

 

[Eug]

That would be my guess - we'll see. The base Mac Studio Max is nicely equipped - except for the standard 512GB NVMe drive (and the $200 1TB upgrade). Sooo close to not being penny pinching jerks for once .

Well, I'd agree it is reasonably spec'd for its price point, but my main point is that it's way too expensive for the majority of the population. Plus it's huge in comparison to the Mac mini. It's got the same footprint, but it's actually fairly tall.

BTW, it turns out the Mac Studio won't actually fit under my monitor. I have a 30" Apple Cinema HD Display, and it only has about 3.4" of clearance under its bottom bezel, or about 3ish inches if you measure the clearance above the base platform. Not that Apple cares though, since that is an ancient monitor.

The Mac mini is nicely priced too, but it's underspec'd for some people, especially because of its 16 GB RAM limitation. Hopefully that's fixed with M2 (although I personally would be OK with 16 GB).

 

[Doug S]

Apple said the M1 Ultra is the last M1 chip. But Apple said Mac Pro is for another day, so who knows what they got planned. They cannot reach a bigger die because they reached the limits of TSMC 5nm.

TSMC's reticle limit is 858 mm^2. Apple could have gone a LOT bigger if they wanted. It makes sense to reuse the M1 Max die in this way rather than create a special die just for the Mac Studio.

Another reason to shy away from larger dies is that it is a dead end for everyone soon. With the high NA EUV machines coming along in a couple years that reticle limit is halved. For everyone, not just TSMC.

 

[Doug S]

Maybe it will be dual M2 Ultra; it’s the end of the line for M1 based chips.

I'm thinking:

M2: 4P+4E, 10GPU
M2 Pro: 12P+2E, 20 GPU
M2 Max: 12P+2E, 40 GPU
M2 Ultra: 24P+4E, 80 GPU
M2 Grande: 48P+8E, 160 GPU

I'm thinking along those lines but they might have 12/24/48 GPU cores, so it is a 50% increase on both CPU and GPU. The die size would grow a bit assuming it is based on N4/N4P. I'm assuming the M2 will have A16's cores, which may get a decent performance bump on their making a 50% increase in their number less necessary (particularly if they get bigger)

 

[Eug]

Hmmm...

https://twitter.com/x/status/1501352016026619906

 

[Ajay]

Well, I'd agree it is reasonably spec'd for its price point, but my main point is that it's way too expensive for the majority of the population. Plus it's huge in comparison to the Mac mini. It's got the same footprint, but it's actually fairly tall.

The Mac mini is nicely priced too, but it's underspec'd for some people, especially because of its 16 GB RAM limitation. Hopefully that's fixed with M2 (although I personally would be OK with 16 GB).

Yeah, configuring a Mac Studio w/1TB is $2199 - is kind of high. It would be cool if the next Mac Mini had an M2/M2 Pro option. M2 Pro would be a nice jump over the current M1 SoC. I would think a Mac Mini M2 Pro 16GB/1TB at ~$1500 would be more enticing - especially with small bumps in the # of CPU and GPU cores. It would certainly entice me (as I’m getting so sick and tired of Windows). I’d probably get hit with the Apple $200 tax for doubling memory

 

[nickmania]

I think is not going to be any Mac Pro or Mac Mini, it does not make sense with this Mac Studio, apple is not going to let you to use its M1 chip for less than $2000 in a Mac Mini and a Mac Pro is unnecessary with this new Mac Studio. Probably the only new computer will be the Imac with M1.

 

[Eug]

Yeah, configuring a Mac Studio w/1TB is $2199 - is kind of high. It would be cool if the next Mac Mini had an M2/M2 Pro option. M2 Pro would be a nice jump over the current M1 SoC. I would think a Mac Mini M2 Pro 16GB/1TB at ~$1500 would be more enticing - especially with small bumps in the # of CPU and GPU cores. It would certainly entice me (as I’m getting so sick and tired of Windows). I’d probably get hit with the Apple $200 tax for doubling memory

16 --> 32 GB is usually $400 at Apple.

 

[Ajay]

16 --> 32 GB is usually $400 at Apple.

lol. Silly me. Then the 'tax' aspect is real

 

[Heartbreaker]

I think is not going to be any Mac Pro or Mac Mini, it does not make sense with this Mac Studio, apple is not going to let you to use its M1 chip for less than $2000 in a Mac Mini and a Mac Pro is unnecessary with this new Mac Studio. Probably the only new computer will be the Imac with M1.

They should have the M1 Pro somewhere in the lineup. I don't really care if it's in the Mini M1 Pro or a Studio M1 Pro.

Not having an M1 Pro model at all is a gap in the lineup, that Apple seems to recognize, since they are keeping the Intel Mini on sale for now. If they felt that base was covered they could discontinue the Intel Mini.

Or they might just be waiting for the M2 Mini as powerful enough to kill off the Intel Mini.

 

[Eug]

M1 Ultra Mac Studio is 2 lbs / 0.9 kg heavier than the M1 Max model.

Some have said it could be due to a beefier power supply, but if you look at the power specs, Apple says the max power output is the same between the two models. So that makes me wonder if the M1 Ultra has a different cooling setup.

 

[gdansk]

With the different package shapes it seems almost mandatory to have different cooling solutions.

 

[Doug S]

They should have the M1 Pro somewhere in the lineup. I don't really care if it's in the Mini M1 Pro or a Studio M1 Pro.

Not having an M1 Pro model at all is a gap in the lineup, that Apple seems to recognize, since they are keeping the Intel Mini on sale for now. If they felt that base was covered they could discontinue the Intel Mini.

Or they might just be waiting for the M2 Mini as powerful enough to kill off the Intel Mini.

Nah, I think it makes sense to offer the Mini with regular M1 & M1 Pro, given that the Studio is offering with the M1 Max / Ultra. That completes a nice range for the SFF desktop line. Whether they wait until M2 to do a Pro version or just silently replace the x86 model with an M1 Pro model on some random day in the next few months, who knows.

 

[Eug]

With the different package shapes it seems almost mandatory to have different cooling solutions.

According to Apple:

They have the same 370W power supply. The additional weight is due to M1 Ultra having a larger copper thermal module, where as M1 Max has an aluminum heatsink.

Apple explains why the M1 Ultra-equipped Mac Studio is two pounds heavier

The power supplies are the same; the heatsinks are not.

www.theverge.com

 

[Roland00Address]

TSMC's reticle limit is 858 mm^2. Apple could have gone a LOT bigger if they wanted. It makes sense to reuse the M1 Max die in this way rather than create a special die just for the Mac Studio.

Another reason to shy away from larger dies is that it is a dead end for everyone soon. With the high NA EUV machines coming along in a couple years that reticle limit is halved. For everyone, not just TSMC.

Question since I do not know and this seems like a very basic concept (which I need to learn), how does something like Cerebras work with the reticile limit since the entire chip is the entire wafer except a rectangle?

 

[ashFTW]

I'm thinking along those lines but they might have 12/24/48 GPU cores, so it is a 50% increase on both CPU and GPU. The die size would grow a bit assuming it is based on N4/N4P. I'm assuming the M2 will have A16's cores, which may get a decent performance bump on their making a 50% increase in their number less necessary (particularly if they get bigger)

Sounds reasonable.

 

[Mopetar]

I think that a lot of people are going to miss the 27 inch iMac. The Studio even in it's base configuration with a non Apple monitor is going to be an expensive replacement.

I don't think it will be gone for long. Apple just doesn't have the replacement ready yet.

Given that the entry-level iMac has been bumped up to 24", we may see a high-end 30" model in the future.

Maybe they're waiting until they have a new chip (M2 and variants) before releasing it, but they've had a large and small iMac for some time now so it would be unusual for them to have a single model.

 

[Hitman928]

Question since I do not know and this seems like a very basic concept (which I need to learn), how does something like Cerebras work with the reticile limit since the entire chip is the entire wafer except a rectangle?

They are able to use a process called stitching to combine multiple reticle shots together.

Edit: I don't have experience in stitching, but my company has done it for some large sensor arrays (not my group). From what I understand, you basically put some special markers on the design and that lets the foundry align the reticle shots as it steps through each one, with a slight bit of overlap.

 

[Doug S]

They are able to use a process called stitching to combine multiple reticle shots together.

Edit: I don't have experience in stitching, but my company has done it for some large sensor arrays (not my group). From what I understand, you basically put some special markers on the design and that lets the foundry align the reticle shots as it steps through each one, with a slight bit of overlap.

Yep, I don't know the limitations to that, but I assume there's enough connectivity that chiplet type designs like the M1 Max could be connected on the wafer (if Apple didn't have to flip them around for that) but the cost savings for that (i.e. no carrier/interposer required) is probably lost by the fact you have to bin on multiple dies at once. Cerebras doesn't care about that since 'AI' designs have tons of redundancy already.

 

[Hitman928]

Yep, I don't know the limitations to that, but I assume there's enough connectivity that chiplet type designs like the M1 Max could be connected on the wafer (if Apple didn't have to flip them around for that) but the cost savings for that (i.e. no carrier/interposer required) is probably lost by the fact you have to bin on multiple dies at once. Cerebras doesn't care about that since 'AI' designs have tons of redundancy already.

Yeah, the defect rate obviously starts to bite you hard the more and more you stitch together if each stitch tile has to be fully functional (or near it due to limited redundancy). There's also an added factor of the possibility of misalignment between stitches. If you just are stitching two together and it's a tiny bit off, it's probably not a big deal, but if you're stitching dozens or more together, then a little misalignment over the whole thing can be disastrous. I'm sure there's extra checks involved as you go which will also increase cost.

There's not a lot of circuits out there that stitching would make sense, but something like Cerberus is a prime candidate for it, to the point that they just stitch the whole wafer together, which is pretty mind boggling.

 

[Doug S]

Yeah, the defect rate obviously starts to bite you hard the more and more you stitch together if each stitch tile has to be fully functional (or near it due to limited redundancy). There's also an added factor of the possibility of misalignment between stitches. If you just are stitching two together and it's a tiny bit off, it's probably not a big deal, but if you're stitching dozens or more together, then a little misalignment over the whole thing can be disastrous. I'm sure there's extra checks involved as you go which will also increase cost.

There's not a lot of circuits out there that stitching would make sense, but something like Cerberus is a prime candidate for it, to the point that they just stitch the whole wafer together, which is pretty mind boggling.

Well I would assume that since 'AI' calculations are embarrassingly parallel, there isn't a whole lot of communication between individual Cerebras "chips" required. That means they don't need too many stitches so they can be thicker which increases the margin for error. If you wanted to have 10,000 of them along one edge like the M1 Ultra, the margin for error is would be reduced.