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

 

[uzzi38]

What do you make of the claim that M4 uses 50% of the power of M2 at the same speed as M2?

Given it's 2 generations of improvements on the CPU core and a node shrink involved? Eh. Pretty standard, honestly.

The way modern CPU cores are in how they require much more power to increase clocks, if you can shift the frequency/voltage curve up via a node shrink/core improvements, then even assuming no IPC improvements you actually improve power consumption by a large margin. Bundle in IPC improvements (meaning you can clock lower for the same performance level) and it's very easy to reduce power by such a large margin to hit the same performance.

Zen 4 makes for a great example I can easily point to. The 7950X running at 65W TDP (88W PPT, which is the real power consumption from the socket) can put out MT numbers decently ahead (22%) of the 5950X at a bit over half the power at a 105W TDP (142W PPT). It's just a combination of the things I mentioned before that make it really easy to get there.

 

[Doug S]

That's my point. N3E had better yields, hence Apple has no reason to sell you an iPad with only 3 performance cores when the amount of area they take up is such a small part of the overall SoC.

The only other way that you get to three cores outside of Apple fusing one off just to be evil, is that they require each of those cores to be able to hit a certain clock speed within a given voltage.

Suppose that each core independently has a 70% chance of hitting this targets. That means that there's only a 25% chance that all of them can do it. There's the binning target if Apple thinks 25% of sales will be for that variant. Meanwhile the base product gets the next bin that the vast majority of silicon can hit.

There is unlikely to be that much interdie variation, certainly not at the level Apple is binning. Apple's speed binning is pretty lax since it is pure pass/fail, so the threshold must be set such that almost every functional die is a pass. They could get at least 10-20% more ST if they binned the way Intel/AMD do.

 

[SpudLobby]

Given it's 2 generations of improvements on the CPU core and a node shrink involved? Eh. Pretty standard, honestly.

The way modern CPU cores are in how they require much more power to increase clocks, if you can shift the frequency/voltage curve up via a node shrink/core improvements, then even assuming no IPC improvements you actually improve power consumption by a large margin. Bundle in IPC improvements (meaning you can clock lower for the same performance level) and it's very easy to reduce power by such a large margin to hit the same performance.

Zen 4 makes for a great example I can easily point to. The 7950X running at 65W TDP (88W PPT, which is the real power consumption from the socket) can put out MT numbers decently ahead (22%) of the 5950X at a bit over half the power at a 105W TDP (142W PPT). It's just a combination of the things I mentioned before that make it really easy to get there.

Precisely. Throw in inherent electric gains, and some extra cache and architecture gains, and two more cores where the relationship between frequency and power isn’t linear but frequency and performance is often close enough, and MT is a sum of those performances?

Very expected to pull off. What’s more telling is ST power/performance curve improvements. Not that the extra cores and MT stuff isn’t relevant to the product though.

 

[SpudLobby]

There is unlikely to be that much interdie variation, certainly not at the level Apple is binning. Apple's speed binning is pretty lax since it is pure pass/fail, so the threshold must be set such that almost every functional die is a pass. They could get at least 10-20% more ST if they binned the way Intel/AMD do.

Yes exactly Doug. I don’t even know what he’s talking about, it’s not logical in this context for Apple’s M4. One e core? N3E and Apple binning differences? They don’t even frequency bin by design!

 

[Eug]

Yeah that makes me think it’s basically artificial and more about segmentation here. The odds the one E core is that big of a deal yield wise is really low I think.

M4 binning is based on P core. Or am I misunderstanding you?

So are you guys saying that the new iPad Pro is a buy? I saw the thing today, and was quite tempted, because the iPad I currently have is from 2018, and I could use an upgrade.

It's a definite buy for me. I wanted OLED, landscape camera, 256 GB base storage, AV1 (just because), and a lighter Magic Keyboard. I got all of the above. The price went up a bit more than I was hoping, but that increase was way, way lower than the crazy $500 prediction from some of the doom-and-gloomers. I was hoping for a US$100-150 price increase, but it went up $200, from $799 base to $999 base at retail (but with double the base storage).

Canadian edu pricing:

M2 iPad Pro
128 GB CA$1099 (US$800)
256 GB CA$1229 (US$895)

M4 iPad Pro
256 GB CA$1249 (US$909)

Also, next month we should get a CA$100 (US$73) gift card back.

 

[Mopetar]

Based on the 'best' scores in GB6 ST Apple gained about 15% from M2 to M3. Apple E cores tend to be worth about 1/3 of a P core, so an 8 core M2 had 5.3 cores vs 10 core M4 having 6 cores, accounting for 12%. From N3E TSMC says you should get 6-7%. If you multiply all that together (can't add it, since they are cumulative) you get 38%. They need to gain another 9% from other sources to achieve a 50% overall gain.

Apple is saying "up to" so they could have picked something that benchmarks a bit better for whatever reason. I suspect that they've juiced the clocks a bit. I don't recall any mention of battery life improvements, so either I missed those or they don't exist. Some of that can be attributed to thinner device, but I think they may have let the SoC sip harder.

I don't think we'll see an M5 in the fall though. It's more likely that Apple just doesn't refresh some products until 2025 when they have an SoC successor. They can refresh the Air with a full die M4 and claim upgrade due to more efficiency cores even if the underlying cores are largely the same. Meanwhile an M4 Pro or Ultra can be whatever the hell Apple wants them to be.

 

[SpudLobby]

M4 binning is based on P core. Or am I misunderstanding you?

No they don’t bin based on P cores. The first guy is just totally misunderstanding N3E and yields in context of the M3 -> M4 etc.

It's a definite buy for me. I wanted OLED, landscape camera, 256 GB base storage, AV1 (just because), and a lighter Magic Keyboard. I got all of the above. The price went up a bit more than I was hoping, but that increase was way, way lower than the crazy $500 prediction from some of the doom-and-gloomers. I was hoping for a US$100-150 price increase, but it went up $200, from $799 base to $999 base (but with double the base storage).

Canadian edu pricing:

M2 iPad Pro
128 GB CA$1099 (US$800)
256 GB CA$1229 (US$895)

M4 iPad Pro
256 GB CA$1249 (US$909)

 

[SpudLobby]

Apple is saying "up to"

It’s very safe to assume that’s about the core counts. They’ve done this with the Mx Pro stuff and they’ve done it with iGPUs for a while now. They don’t like frequency binning. Doubt that they’d do it in an iPad if they did, either.

so they could have picked something that benchmarks a bit better for whatever reason. I suspect that they've juiced the clocks a bit. I don't recall any mention of battery life improvements, so either I missed those or they don't exist. Some of that can be attributed to thinner device, but I think they may have let the SoC sip harder.

I don't think we'll see an M5 in the fall though. It's more likely that Apple just doesn't refresh some products until 2025 when they have an SoC successor. They can refresh the Air with a full die M4 and claim upgrade due to more efficiency cores even if the underlying cores are largely the same. Meanwhile an M4 Pro or Ultra can be whatever the hell Apple wants them to be.

 

[Eug]

No they don’t bin based on P cores. The first guy is just totally misunderstanding N3E and yields in context of the M3 -> M4 etc.

The binned variant has one less P-core. All the chips get 6 E-cores, but only the high end variant gets 4 P-cores.

It’s very safe to assume that’s about the core counts. They’ve done this with the Mx Pro stuff and they’ve done it with iGPUs for a while now. They don’t like frequency binning. Doubt that they’d do it in an iPad if they did, either.

In the past, they sometimes had different frequencies in different devices.

 

[SpudLobby]

View attachment 98544

The binned variant has one less P-core. All the chips get 6 E-cores, but only the high end variants get 4 P-cores.

I stand corrected. It may well be more artificial but could be genuinely they decided it made sense for yields I guess.

In the past, they sometimes had different frequencies in different devices.

Yes but rarely. Like iPad Mini A15 or M2 Pro/Max ST freq.

 

[trivik12]

So far no benchmarks right. Normally we see some geekbench numbers if reporters already had access to it.

 

[Eug]

So far no benchmarks right. Normally we see some geekbench numbers if reporters already had access to it.

I'm thinking the benchmarks will start to show up by the end of the week or on the weekend. But I also think they will ALL be 10-core. No 9-core.

 

[mikegg]

Given it's 2 generations of improvements on the CPU core and a node shrink involved? Eh. Pretty standard, honestly.

The way modern CPU cores are in how they require much more power to increase clocks, if you can shift the frequency/voltage curve up via a node shrink/core improvements, then even assuming no IPC improvements you actually improve power consumption by a large margin. Bundle in IPC improvements (meaning you can clock lower for the same performance level) and it's very easy to reduce power by such a large margin to hit the same performance.

Zen 4 makes for a great example I can easily point to. The 7950X running at 65W TDP (88W PPT, which is the real power consumption from the socket) can put out MT numbers decently ahead (22%) of the 5950X at a bit over half the power at a 105W TDP (142W PPT). It's just a combination of the things I mentioned before that make it really easy to get there.

The person I quoted claims that M4 uses the same core design as M3. Therefore, there is no new core claimed. Furthermore, the node change likely doesn’t help much in performance. Some day N3B is faster while others claim N3E is faster.

So I was asking the person who made those claims for an opinion on how M4 is twice as efficient as M2 while Apple did not make the same claim for M3.

 

[Eug]

The person I quoted claims that M4 uses the same core design as M3. Therefore, there is no new core claimed. Furthermore, the node change likely doesn’t help much in performance. Some day N3B is faster while others claim N3E is faster.

So I was asking the person who made those claims for an opinion on how M4 is twice as efficient as M2 while Apple did not make the same claim for M3.

6 e-cores on M4 vs 4 on M2 and M3.

Apple didn't clearly state what they meant, but I suspect that they are talking about power utilization at some cherry-picked lower performance level.

 

[Mopetar]

It’s very safe to assume that’s about the core counts. They’ve done this with the Mx Pro stuff and they’ve done it with iGPUs for a while now. They don’t like frequency binning. Doubt that they’d do it in an iPad if they did, either.

They don't have to have different frequencies between the different categories. You can select for a frequency for which 20% (my hypothetical guess that may be more or less in actual reality) qualify for a certain amount on all 4 cores and the remaining percentage can hit if the weakest core is disabled.

Depending on what Apple wants to do with the M4, there's no reason to assume they only have these two bins. It's entirely possible that these iPads are getting unique bins that we don't see for any of their Macs.

 

[SpudLobby]

They don't have to have different frequencies between the different categories. You can select for a frequency for which 20% (my hypothetical guess that may be more or less in actual reality) qualify for a certain amount on all 4 cores and the remaining percentage can hit if the weakest core is disabled.

Yes dude, I fully realize that, you keep starting from first principles and then getting sidetracked. What they select at will of course determine how many they have left, and if they’re only deactivating *one* P core and *one* E core then probably the frequency boost won’t be as high as it would be otherwise if they did full and deliberate frequency selection across entire dice.

Depending on what Apple wants to do with the M4, there's no reason to assume they only have these two bins. It's entirely possible that these iPads are getting unique bins that we don't see for any of their Macs.

This part is true, sure

 

[Eug]

Yes dude, I fully realize that, you keep starting from first principles and then getting sidetracked. What they select at will of course determine how many they have left, and if they’re only deactivating *one* P core and *one* E core then probably the frequency boost won’t be as high as it would be otherwise if they did full and deliberate frequency selection across entire dice.

Just to be clear, they are not deactivating any e-cores. It's just the one p-core.

 

[uzzi38]

The person I quoted claims that M4 uses the same core design as M3. Therefore, there is no new core claimed. Furthermore, the node change likely doesn’t help much in performance. Some day N3B is faster while others claim N3E is faster.

So I was asking the person who made those claims for an opinion on how M4 is twice as efficient as M2 while Apple did not make the same claim for M3.

I would be very surprised if it's identical core, it seems much more likely to me there would have been at least some minor tweaks to improve power efficiency/improve the peak clock rate or catch some low hanging fruit in the new core design from last gen. Even if only very minor changes.

Also, it doesn't really matter what the person prior said, Apple's claim of 50% less power at the same performance that you were asking about is against M2, so yes there is a node shrink involved (N5P -> N3E).

As for why Apple didn't make the same claim for N3: the gains were likely even smaller for M3 vs M2, so Apple likely just didn't want to talk about it. In fact, I do remember them not saying very much about the CPU core despite the fact it was actually an enhanced core, and we later found out that yes, IPC improvements were minimal.

 

[SpudLobby]

RE: frequency

Mopestar if you mean that 3 Cores might indicate they really are pushing frequency just just barely enough to where one p core and one e core might not meet the cutoff and then axing them (even with N3E’s better yields) then I mean yeah that’s believable.

The point Doug and I make is that this is a small number of cores and a small part of the die and some of this is just artificial segmentation — and further while it’s true even this binning would be about voltages and drive current (frequency on some level - obviously — Apple is very conservative about how they push that, so the gains they get from this are probably going to he low if it’s just two cores they’re splitting off from and based on past history. I would bet most of this is segmentation more than a yield issue per se and Apple if they wanted to bin for frequency *as their final differentiator in the SKUs* across an entire die would have a vastly different lineup and also probably more performance variation.

 

[Eug]

I would be very surprised if it's identical core, it seems much more likely to me there would have been at least some minor tweaks to improve power efficiency/improve the peak clock rate or catch some low hanging fruit in the new core design from last gen. Even if only very minor changes.

Also, it doesn't really matter what the person prior said, Apple's claim of 50% less power at the same performance that you were asking about is against M2, so yes there is a node shrink involved (N5P -> N3E).

As for why Apple didn't make the same claim for N3: the gains were likely even smaller for M3 vs M2, so Apple likely just didn't want to talk about it. In fact, I do remember them not saying very much about the CPU core despite the fact it was actually an enhanced core, and we later found out that yes, IPC improvements were minimal.

Again, a huge difference here is that M4 has 50% more efficiency cores than both M2 and M3.

 

[uzzi38]

Again, a huge difference here is that M4 has 50% more efficiency cores than both M2 and M3.

Yeah in this power category where both P and E cores are stuck with <1W of power each the extra E cores will absolutely make a difference.

(~8W is the cap of what can be sustainably cooled in most small passive devices).

 

[Eug]

Yeah in this power category where both P and E cores are stuck with <1W of power each the extra E cores will absolutely make a difference.

(~8W is the cap of what can be sustainably cooled in most small passive devices).

Well, to be clear, my point was that Apple did not specify where on the power efficiency vs. performance curve they took that measurement of 50% power utilization for performance equal to M2.

They never even specified it was near peak M2 performance (or peak within the iPad's thermal limitations).

 

[Ajay]

Well, to be clear, my point was that Apple did not specify where on the power efficiency vs. performance curve they took that measurement of 50% power utilization for performance equal to M2.

Probably, which, ofc, is a bit disingenuous. But it's Apple after all - groomed by the king of marketing (Jobs).

 

[Eug]

No way I'm getting Nanotexture

Not only is it only available on the 1 & 2 TB models, I really don't like the look. Also, they left the bezels glossy so it's a sharp transition from the bezel to the screen. I guess they left the bezels glossy for the FaceTime camera.

Anyhow, still nothing on Geekbench, but like I said, given that machines aren't out until next week, it may be a couple of days before we see some Geekbench leaks.

 

[Eug]

Interesting. The Apple logo on the back of the new iPad Pro is actually a copper heatsink.