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

 

[Eug]

https://twitter.com/i/web/status/1505708249135296515

Only Apple would do something wasteful like this.

I wonder if this is a technical limitation / OS compatibility issue.

Center Stage was introduced with A13, so they’d most likely need A13 minimum. However, no A13 device has ever shipped with less than 64 GB.

 

[Doug S]

I wonder if this is a technical limitation / OS compatibility issue.

Center Stage was introduced with A13, so they’d most likely need A13 minimum. However, no A13 device has ever shipped with less than 64 GB.

The limitation probably comes from the size of flash devices they are buying - 64GB is probably the smallest amount they can use with their controllers now. Given that SSDs are under $100 per TB retail, Apple is surely paying less than $5 for 64GB of flash. Is it worth special ordering something smaller? Who knows, maybe that extra space will be used in the future. 64GB leaves a lot of room for apps...

 

[Roland00Address]

https://twitter.com/i/web/status/1505708249135296515

Only Apple would do something wasteful like this.

Why is it wasteful? You are selling a 5k monitor in a time one can buy 50 inch 4k tv for $300. Of course it is not the same but when selling a $1.5k device one overbuilds it, making it nice so it lasts several years with the goal of not just the margins on this device but to keep the customer happy so they will buy future devices from the same brand.

Likewise in a world where businesses buy $9k Surface Hubs for a conference rooms throwing enough silicon into these devices that they can be a good ai powered webcam also makes sense. Of course I say this last sentence but it seems Apple is kind of messing this part up.

 

[Eug]

M1 Ultra vs M1 Max:



M1 Max:







M1 Ultra:

 

[Doug S]

Interesting that when the GPU test was run by itself on the M1 Max Studio the GPU could hit 100% utilization, but when the CPU and GPU were tested together the GPU only hit 75-80% utilization. That's despite the GPU cores barely exceeding 40C and CPU cores at 60C, and the fans basically idling.

Obviously not a power limitation with 370W at its disposal, or a heat limitation giving how cool everything is running and the fans at idle. So is there something type of internal limitation for how much power can be delivered to the SoC, or available bandwidth to the SLC or something of that nature? Of course the most likely explanation is it is just a bug, some power profile that's not properly tuned for the M1 Max in a device that's not battery powered.

Like I said, these are version 1.0 devices so there will be little wrinkles. Things will only get better as Apple cures the limitations and bugs so M2 and M3 will really fly.

 

[igor_kavinski]

Interesting that when the GPU test was run by itself on the M1 Max Studio the GPU could hit 100% utilization, but when the CPU and GPU were tested together the GPU only hit 75-80% utilization.

Have to know what the CPU utilization was when GPU was at 100%. GPU not maxing out with CPU being stressed points to the CPU being so busy that it can't properly saturate the GPU with workload.

 

[Eug]

Have to know what the CPU utilization was when GPU was at 100%. GPU not maxing out with CPU being stressed points to the CPU being so busy that it can't properly saturate the GPU with workload.

Yes, he was running a CPU benchmark at the same time. CPU was maxed out at 100% with Cinebench so it shouldn’t come as a big surprise that GPU couldn’t be maxed out.

Max’s conclusion that the OS was artificially limiting GPU performance due to power limitations doesn’t really make sense here.

 

[thigobr]

Nobody knows if the temperature sensors are reporting averages or hotspots within the chip... Maybe this is just a limitation of the silicon where they can't feed more power/do more work without having areas where temperature go over unsafe levels.

 

[Eug]

Nobody knows if the temperature sensors are reporting averages or hotspots within the chip... Maybe this is just a limitation of the silicon where they can't feed more power/do more work without having areas where temperature go over unsafe levels.

GPU temp was 42C during that combined test. CPU was 60C. So temp was very unlikely to be the issue.

BTW, IIRC during the individual tests the GPU maxed out was 44C and the CPU maxed out was 58C.

Actually, the most impressive part to me for all of this was that the fan remained quiet.

 

[repoman27]

That's really a stretch. Even if what you suggest was true, using the same PSU makes no difference in the ease of bringing a different system to market. That's literally the easiest thing they can do, switching to a lower wattage power supply if 370W is more than they need.

Its ridiculous though just based on the size of the Studio's internals. How in the heck were those giant heatsinks going to fit inside their most recent iMac form factor? If they were going to stick an M1 Ultra in there it wouldn't use the big heavy copper heatsink, it would probably have had something like phase change cooling heat pipes on the package to keep it very low profile and used the entire backside of the display to dissipate excess heat. I could see Apple having done something different and use thin copper fins instead of plastic for the rear of the iMac Pro display. Whatever they did, it would be nothing like the cooling solution they did with the Studio.

Seems pretty clear that they had already concluded integrating a midrange / entry "pro" level desktop with the display is a dumb idea. Not sure where you get the idea the Studio was a rush job. It seems exceptionally well engineered to me, not something slapped together at the last minute as you imply. Why would having a Mini LED display make an all-in-one Studio more attractive? The more expensive the display the worse the idea of integration is. A lot of professionals already own a quality display, and don't want to pay for something they don't need - or to have one that doesn't match if they are one of the many with more than one display.

You get that's why they sell the stand separately, right? Because they know a lot of professionals also have some type of VESA mounting solution (swing arm or whatever) for their monitor(s) which the iMac design also screws up since you can't remove a stand if part of the Mac is integrated within it and if you want a different display height it won't match your other ones unless you set the iMac on a book or something.

I guess you missed that the iMac Pro was already a thing, where Apple put a strikingly similar 370 W PSU, whopping big copper heatsink, and dual blowers into the existing 27-inch iMac chassis. And that was able to power and cool a 140 W TDP Xeon W-2191B CPU and 250 W Radeon Pro Vega 64X.

I never said the Mac Studio was a rush job. I'm sure Apple routinely evaluates different design concepts and develops them concurrently. The obvious route would have been to simply put the M1 Max and M1 Ultra in the existing 27-inch iMac chassis. However, they've been using the same design for nearly a decade at this point, and that 5K panel is pushing 5 years old. A different size / resolution mini-LED panel in an updated chassis is almost certainly something Apple was investigating, yet what they released instead was the Mac Studio.

A year after its introduction in October of 2014, the starting price of the iMac (Retina 5K, 27-inch) dropped to $1799 and remained there until that model was discontinued earlier this month. It was one of Apple's best selling desktop Macs. For the literally millions of customers who purchased that model, the Mac Studio is not a replacement or even an otherwise great product. Instead of being able to buy an Alder Lake version of that iMac for $1799, the transition to Apple Silicon leaves them paying $1599 just to get the same 5-year-old display. Add to that the cheapest Mac Studio at $1999, $178 for the keyboard and mouse that used to be included (or $298 for the new black versions, which cost more, because they're black), and $129 for a 1.8 m Thunderbolt cable so you can place the Studio somewhere other than directly under the display (which isn't possible given the absurdly short cable Apple includes in the box). That's $3905 for the new entry level Mac Studio and 27-inch Studio Display. Not to mention that the 27-inch Intel iMac it "replaces" could be upgraded to 128 GB of RAM by the end user for just $420, and the hypothetical Alder Lake version would have been able to beat the current Studio in peak single-threaded performance. So don't give me this garbage that Apple selling the 27-inch Studio Display and Mac Studio separately is somehow better for their customers than continuing to offer a 27-inch iMac, because that simply isn't true.

 

[StinkyPinky]

That Ultra is huge. Makes me wonder what else they can do with this setup? A Mac Pro with 2 x Ultra's would be expensive as hell you would think. Not to mention it would have a limitation of 256GB of memory compared to 1TB+ of current Mac Pros.

Makes me think the Mac Pro is still some time away, until the M2 Max/Ultra.

 

[Doug S]

That Ultra is huge. Makes me wonder what else they can do with this setup? A Mac Pro with 2 x Ultra's would be expensive as hell you would think. Not to mention it would have a limitation of 256GB of memory compared to 1TB+ of current Mac Pros.

Makes me think the Mac Pro is still some time away, until the M2 Max/Ultra.

Far denser LPDDR is available than what Apple has used with the M1. They could easily do 1 or 2 TB of LPDDR5X on a M2 based Mac Pro. It is possible they could have DDR5 DIMM slots but then you have to deal with a larger pool of far slower memory, and it complicates their vision for unified memory. So I think they will offer a Mac Pro with at least 1 TB as the top end. That's a bit less than the 1.5 TB max they have now, but how many customers have installed over 1 TB? I don't know, but I'll bet Apple does.

I'm not sure why you think the Ultra is expensive when it is replacing both a high end Intel CPU and a fairly high end AMD GPU. In the Mac Pro, a high end Xeon and a high end AMD GPU. The Ultra with 128 GB of LPDDR5 will cost Apple considerably less than a Core i9 or 3080 GPU and 128 GB of DDR4 DIMMs would cost them.

 

[MadRat]

Apple's vision is to cripple storage to force iCloud usage. 2TB on an Apple product will become normal perhaps in 2030. But its not coming soon unless they change strategies.

 

[jpiniero]

Apple's vision is to cripple storage to force iCloud usage. 2TB on an Apple product will become normal perhaps in 2030. But its not coming soon unless they change strategies.

You can get up to 8 TB on the Studio. Storage and memory upgrades are of course a big revenue stream for Apple.

 

[MadRat]

You can get up to 8 TB on the Studio. Storage and memory upgrades are of course a big revenue stream for Apple.

I saw $4k for Studio with 1TB. What is the 8TB price?

 

[biostud]

For desktop solutions, isn't storage less of a problem as you can always have a thunderbolt or network connected storage device?

 

[Eug]

I saw $4k for Studio with 1TB. What is the 8TB price?

Lowest config Mac Studio with 8 TB SSD and 32 GB RAM is US$4399 retail (or $3959 education). However...

For desktop solutions, isn't storage less of a problem as you can always have a thunderbolt or network connected storage device?

Indeed. Or even USB-C for bulk storage.

 

[jpiniero]

I saw $4k for Studio with 1TB. What is the 8TB price?

On the Ultra model, it's an extra $2200.

 

[repoman27]

Here are pricing tables for all of the M1 Mac options, as best I could make them out. Obviously not all options are available on every platform.

SoC	CPU/GPU Cores	Option Price Δ	SoC Retail Price
M1	8/7	$150 (baseline)	$150
M1	8/8	$50	$200
M1 Pro	8/14	$0	$200
M1 Pro	10/14	$200	$400
M1 Pro	10/16	$100	$500
M1 Max	10/24	$200	$700
M1 Max	10/32	$200	$900
M1 Ultra	20/48	$1200	$2100
M1 Ultra	20/64	$1000	$3100

DRAM Capacity	Option Price Δ	DRAM Retail Price	DRAM Price / GB
8 GB	$200 (baseline)	$200	$25.00
16 GB	$200	$400	$25.00
32 GB	$400	$800	$25.00
64 GB	$400	$1200	$18.75
128 GB	$800	$2000	$15.63

NAND Flash Memory Capacity	Option Price Δ	NAND Retail Price	NAND Price / TB
256 GB	$200 (baseline)	$200	$800
512 GB	$200	$400	$800
1 TB	$200	$600	$600
2 TB	$400	$1000	$500
4 TB	$600	$1600	$400
8 TB	$1200	$2800	$350

Ethernet	Option Price Δ	Ethernet Retail Price
no Ethernet	$0 (baseline)	$0
Gigabit Ethernet	$30	$30
10 GbE	$100	$130

Also, TechInsights put up a little blog post teaser about the M1 Ultra recently. Not much new info, but they did include a close-up showing the pitch of the UltraFusion interconnect:

 

[Eug]

Here are pricing tables for all of the M1 Mac options, as best I could make them out. Obviously not all options are available on every platform.

SoC	CPU/GPU Cores	Option Price Δ	SoC Retail Price
M1	8/7	$150 (baseline)	$150
M1	8/8	$50	$200
M1	8/14	$0	$200
M1 Pro	10/14	$200	$400
M1 Pro	10/16	$100	$500
M1 Max	10/24	$200	$700
M1 Max	10/32	$200	$900
M1 Ultra	20/48	$1200	$2100
M1 Ultra	20/64	$1000	$3100

DRAM Capacity	Option Price Δ	DRAM Retail Price	DRAM Price / GB
8 GB	$200 (baseline)	$200	$25.00
16 GB	$200	$400	$25.00
32 GB	$400	$800	$25.00
64 GB	$400	$1200	$18.75
128 GB	$800	$2000	$15.63

NAND Flash Memory Capacity	Option Price Δ	NAND Retail Price	NAND Price / TB
256 GB	$200 (baseline)	$200	$800
512 GB	$200	$400	$800
1 TB	$200	$600	$600
2 TB	$400	$1000	$500
4 TB	$600	$1600	$400
8 TB	$1200	$2800	$350

Ethernet	Option Price Δ	Ethernet Retail Price
no Ethernet	$0 (baseline)	$0
Gigabit Ethernet	$30	$30
10 GbE	$100	$130

Also, TechInsights put up a little blog post teaser about the M1 Ultra recently. Not much new info, but they did include a close-up showing the pitch of the UltraFusion interconnect:

Nice table.

Note that you mislabelled the entry-level M1 Pro as M1. Also, what is the difference between "baseline" and "$0"? You have the baseline M1 costing $150 which I assume is just your guesstimate, but then you have the baseline M1 Pro at $0.

 

[Heartbreaker]

Apple's vision is to cripple storage to force iCloud usage. 2TB on an Apple product will become normal perhaps in 2030. But its not coming soon unless they change strategies.

iCloud? You have TB, USB and Ethernet for attached storage, before you get to something as slow as the cloud.

My only storage concern is with soldered in desktop storage, like on the Mini being a significant detriment to long term repairs. Studio doesn't have that problem.

Hopefully an M2 Mini gets a storage socket as well.

 

[repoman27]

Nice table.

Note that you mislabelled the entry-level M1 Pro as M1. Also, what is the difference between "baseline" and "$0"? You have the baseline M1 costing $150 which I assume is just your guesstimate, but then you have the baseline M1 Pro at $0.

Thanks, I fixed the M1 Pro entry.

Baseline is indeed my guesstimate of the portion of the full retail price Apple attributes to that component. The M1 Pro is an odd one, because as best I can tell, the salvaged 8c/14c M1 Pro is actually the same price as the fully enabled 8c/8c M1. Note that we never see the M1 and M1 Pro in the same platform, even when you think it would make sense. Like why is there no M1 Pro mini or 24-inch iMac? Why keep the 13-inch MacBook Pro around?

Far denser LPDDR is available than what Apple has used with the M1. They could easily do 1 or 2 TB of LPDDR5X on a M2 based Mac Pro. It is possible they could have DDR5 DIMM slots but then you have to deal with a larger pool of far slower memory, and it complicates their vision for unified memory. So I think they will offer a Mac Pro with at least 1 TB as the top end. That's a bit less than the 1.5 TB max they have now, but how many customers have installed over 1 TB? I don't know, but I'll bet Apple does.

Apple does have a patent on combining different memory technologies in a single system, so they are clearly working on it. And Samsung has only said:

[T]he 16Gb LPDDR5X chip will enable up to 64 gigabytes (GB) per memory package...

So still 16 Gbit dies, and "will enable" doesn't mean currently available in sufficient quantity for Apple to use in a Mac. Samsung is also really good at die stacking, but they aren't Apple's preferred DRAM supplier. I'm not sure when Samsung or anybody else will be piling 32 SDRAM dies into a single package.

 

[Heartbreaker]

Apple does have a patent on combining different memory technologies in a single system, so they are clearly working on it. And Samsung has only said:

So still 16 Gbit dies, and "will enable" doesn't mean currently available in sufficient quantity for Apple to use in a Mac. Samsung is also really good at die stacking, but they aren't Apple's preferred DRAM supplier. I'm not sure when Samsung or anybody else will be piling 32 SDRAM dies into a single package.

On the likelihood of coming to market, I would rank a proven existing enabling technology, much further ahead than having a patent.

Layering has been increasing over time, I recall seeing a Samsung product using 12 layers, so 16 layers is probably not far behind, and that would give 32GB LPDDR packages...

 

[Eug]

Baseline is indeed my guesstimate of the portion of the full retail price Apple attributes to that component. The M1 Pro is an odd one, because as best I can tell, the salvaged 8c/14c M1 Pro is actually the same price as the fully enabled 8c/8c M1. Note that we never see the M1 and M1 Pro in the same platform, even when you think it would make sense. Like why is there no M1 Pro mini or 24-inch iMac? Why keep the 13-inch MacBook Pro around?

There is a big jump in price between the MacBook Air and the 14" MacBook Pro. The 13" MacBook Pro provides a nice middle ground price-wise.

One might ask, "But why then do the 13" Air and 13" Pro both have the same SoC?" Well, Kuo Ming-Chi is stating that the next MacBook Air will get a new form factor but will likely retain M1. If that turns out to be true, and if the MacBook Pro ends up with M2, then that provides another differentiating factor, justifying that pricing middle ground. I always did think having M1 in both the Air and Pro was an odd choice, but thing is that there was no other chip to use aside from A-series chips.*

I am still hopeful there will be an M1 Pro Mac mini, starting with 8c/14c. While in terms of die size, it will still be the same as 8c/16c, Apple would price it lower than 8c/16c of course, hopefully providing an exact price replacement for the remaining Intel Mac mini at $1299 (16/512 GB).


*Actually, prior to the release of the M1 Macs, I thought there was a decent possibility that the MacBook Air could use an A14, while MacBook Pro would use A14X (aka M1) or better. I am pleased that turned out to be very wrong regarding the MacBook Air.

 

[repoman27]

On the likelihood of coming to market, I would rank a proven existing enabling technology, much further ahead than having a patent.

Layering has been increasing over time, I recall seeing a Samsung product using 12 layers, so 16 layers is probably not far behind, and that would give 32GB LPDDR packages...

Oh, absolutely.

I've spent a good chunk of time thinking about how Samsung could even build an LPDDR device with 32 dies. There is more than just die stacking involved. The differences in the lengths of the wire bonds starts to cause signal integrity issues with DRAM once you go more than 4-high. Also, you can't really go beyond 2 ranks with LPDDR, and most dies are split into two planes each with their own 16-bit interface. So you'd need to be producing 16 Gbit dies with a single 16-bit interface in volume, configuring them in byte mode, and stacking them 8-high to form a 16 GB dual-ranked octal die stack with a 32-bit interface. Then you'd place two side by side on a substrate, and PoP stack a pair of them to get to a 64 GB x128 device? Or they're going the TSV route to stack LPDDR? Anyways, I don't really see who else would be buying those besides Apple.