M3 roadmap speculation hints at next Apple Silicon generation chips

AppleInsider said:

The M3 Pro's base configuration is anticipated to have 12 CPU cores, again split evenly between performance and efficiency cores, and an 18-core GPU. The top configuration will use add two more performance cores, bringing the total to 14, as well as a 20-core GPU.

The M3 Max will start with a base configuration of 16 CPU cores, using 12 performance and four efficiency cores, and a 32-core GPU. On the high end, the M3 Max will have the same 16-core CPU but a 40-core GPU.

I'd expect the 12-core M3 Pro to be 8 performance-cores, it would be unusual to have 6 efficiency cores on Pro and 4 on Max. Plus M2 Pro 10-core already has 6 performance cores, 4 efficiency. Moving to 6p/6e only increases the efficiency cores. 8p/4e would increase CPU performance at least 50%.

The GPU core counts look like a small increase so I'd say they will increase transistor count per core:

https://wccftech.com/apple-a10-fusion-cores-bigger-than-competition/

"One reason why Apple is adamant is designing larger cores is because having more transistors per core helps when performance and efficiency per-watt metric is calculated. While this might not be a good approach when conserving space, clock efficiency greatly increases thanks to these decisions."

They might also have a strategy similar to Intel's tick-tock. 2nm won't be ready until late 2025/2026 so they have to make 3nm last for 2023/2024/2025. I doubt they would throw everything in with the first 3nm revision then have a small refresh in 2025. It's best to split it so that each refresh has a worthwhile improvement (~50% increase each time) so that M4 (2025) is 2x M2 performance.

Marvin said:

AppleInsider said:

The M3 Pro's base configuration is anticipated to have 12 CPU cores, again split evenly between performance and efficiency cores, and an 18-core GPU. The top configuration will use add two more performance cores, bringing the total to 14, as well as a 20-core GPU.

The M3 Max will start with a base configuration of 16 CPU cores, using 12 performance and four efficiency cores, and a 32-core GPU. On the high end, the M3 Max will have the same 16-core CPU but a 40-core GPU.

I'd expect the 12-core M3 Pro to be 8 performance-cores, it would be unusual to have 6 efficiency cores on Pro and 4 on Max. Plus M2 Pro 10-core already has 6 performance cores, 4 efficiency. Moving to 6p/6e only increases the efficiency cores. 8p/4e would increase CPU performance at least 50%.

The GPU core counts look like a small increase so I'd say they will increase transistor count per core:

https://wccftech.com/apple-a10-fusion-cores-bigger-than-competition/

"One reason why Apple is adamant is designing larger cores is because having more transistors per core helps when performance and efficiency per-watt metric is calculated. While this might not be a good approach when conserving space, clock efficiency greatly increases thanks to these decisions."

They might also have a strategy similar to Intel's tick-tock. 2nm won't be ready until late 2025/2026 so they have to make 3nm last for 2023/2024/2025. I doubt they would throw everything in with the first 3nm revision then have a small refresh in 2025. It's best to split it so that each refresh has a worthwhile improvement (~50% increase each time) so that M4 (2025) is 2x M2 performance.

Speaking of transistors, TSMC is staggering the full 3nm transition over N3 and N2. The first phase is the die shrink. N3/N3E and N3P (N3+) transistors are still FinFET-based, which TSMC has been using since 16nm (2013).

However, N3 and N3P are the last two generations of FinFET. To move beyond 3nm, the industry has adopted GAAFET (GAA = gate all-around, FET = field effect transistor), which TSMC calls "Nanosheet" transistors. I'm pretty sure I remember TSMC said in their initial press release they would do both things at the same time, the [1] die shrink and the [2] change in transistor architecture. That is what Samsung has done, with success, at least with regard to focused cryptocurrency-mining silicon. See TechInsights on this topic, here: https://www.techinsights.com/disruptive-event/samsung-3nm-gaa-process
But TSMC won't start using Nanosheet (GAAFET) transistors until N2 and N2P.

So N3 (and the more mainstream N3P refinement after it) isn't as radical a shift as was first thought. Right now, Apple can increase the number of transistors per core, as you suggest, without having to worry about the FinFET-to-Nanosheet transition as well. 

Here is a recent Anandtech article that provides a bit more context: https://www.anandtech.com/show/18960/samsung-foundry-s-3nm-and-4nm-yields-are-improving-report

Marvin said:

AppleInsider said:

The M3 Pro's base configuration is anticipated to have 12 CPU cores, again split evenly between performance and efficiency cores, and an 18-core GPU. The top configuration will use add two more performance cores, bringing the total to 14, as well as a 20-core GPU.

The M3 Max will start with a base configuration of 16 CPU cores, using 12 performance and four efficiency cores, and a 32-core GPU. On the high end, the M3 Max will have the same 16-core CPU but a 40-core GPU.

I'd expect the 12-core M3 Pro to be 8 performance-cores, it would be unusual to have 6 efficiency cores on Pro and 4 on Max. Plus M2 Pro 10-core already has 6 performance cores, 4 efficiency. Moving to 6p/6e only increases the efficiency cores. 8p/4e would increase CPU performance at least 50%.

The GPU core counts look like a small increase so I'd say they will increase transistor count per core

It is a good lineup of options in concept imo, but the details are still screwy and a bit nonsensical. As I read it, and in terms of CPU perf + CPU eff + GPU cores, the rumors looks like:

4+4+10
6+6+18
8+6+20
12+4+32
12+4+40

Why would the Pro have 6 eff cores while the Max has 4 eff cores? They are not following the existing plan of designing 1 chip, the Max, and "chopping" off part of it for the Pro, and using two for the Ultra? So, if the Pro has 6 eff cores, the Max would have 6 eff cores as well. It really should look like:

4+4+10
6+6+16 (M3 Pro binned)
8+6+20 (M3 Pro)
10+6+32 (M3 Max binned)
10+6+40 (M3 Max)

Or, something like this is more sensible, which assumed each CPU perf complex is 4 cores, each CPU eff complex is 4 cores. So they would be using the same CPU perf and eff complexes from M3 all the way to M3 Ultra. It just varies by number of complexes and binning.

4+4+10
6+4+16
8+4+20
10+4+32
12+4+40

The big hole in the current Pro vs Max SoC lineup is that the Max doesn't offer more CPU perf cores than the Pro. Adding 2 or 4 more perf cores for the Max would provide a better upsell. So, good to hear the rumors are intimating this. The 18 to 20 GPU cores in the Pro though? Doesn't make sense. 10% isn't much of difference. It needs to be about 15% difference to be able to upsell on order $200. The only way they could make +10% work is if they combine it with a CPU core increase.

tht said:

Marvin said:

AppleInsider said:

The M3 Pro's base configuration is anticipated to have 12 CPU cores, again split evenly between performance and efficiency cores, and an 18-core GPU. The top configuration will use add two more performance cores, bringing the total to 14, as well as a 20-core GPU.

The M3 Max will start with a base configuration of 16 CPU cores, using 12 performance and four efficiency cores, and a 32-core GPU. On the high end, the M3 Max will have the same 16-core CPU but a 40-core GPU.

I'd expect the 12-core M3 Pro to be 8 performance-cores, it would be unusual to have 6 efficiency cores on Pro and 4 on Max. Plus M2 Pro 10-core already has 6 performance cores, 4 efficiency. Moving to 6p/6e only increases the efficiency cores. 8p/4e would increase CPU performance at least 50%.

The GPU core counts look like a small increase so I'd say they will increase transistor count per core

Or, something like this is more sensible, which assumed each CPU perf complex is 4 cores, each CPU eff complex is 4 cores. So they would be using the same CPU perf and eff complexes from M3 all the way to M3 Ultra. It just varies by number of complexes and binning.

4+4+10
6+4+16
8+4+20
10+4+32
12+4+40

The big hole in the current Pro vs Max SoC lineup is that the Max doesn't offer more CPU perf cores than the Pro. Adding 2 or 4 more perf cores for the Max would provide a better upsell. So, good to hear the rumors are intimating this. The 18 to 20 GPU cores in the Pro though? Doesn't make sense. 10% isn't much of difference. It needs to be about 15% difference to be able to upsell on order $200. The only way they could make +10% work is if they combine it with a CPU core increase.

That number of CPU cores looks right, I can't think of any benefit to having 6 efficiency cores.

For the GPU core count, I'm guessing they must be significantly redesigned cores with more transistors each. In the mobile chips, A12 was 6-core:

https://en.wikipedia.org/wiki/Apple_A12

2p/4e CPU, 4 GPU cores, total 7b transistors

A16 is also 6-core CPU (2p/4e), 5-core GPU but 16b transistors:

https://en.wikipedia.org/wiki/Apple_A16

They must have roughly doubled the transistor count per core.

Since this is a new node, they will have redesigned the cores and in later revisions, they can increase the number of cores. 3nm allows 70% density increase so each core can increase that much, it may end up a little less. Lets say each GPU core has 40% more transistors and 10% more cores, then it's over 50% increase. Plus they are supposed to be faster at the same power on top of this.

I'm expecting 50% gain for CPU/GPU from M2->M3 at a lower power. Then with M4 they can add more cores at a higher power draw. 20-30TFLOPs GPU in a thin and light laptop where the fan barely comes on will be really nice.

Marvin said:

tht said:

Marvin said:

AppleInsider said:

The M3 Pro's base configuration is anticipated to have 12 CPU cores, again split evenly between performance and efficiency cores, and an 18-core GPU. The top configuration will use add two more performance cores, bringing the total to 14, as well as a 20-core GPU.

The M3 Max will start with a base configuration of 16 CPU cores, using 12 performance and four efficiency cores, and a 32-core GPU. On the high end, the M3 Max will have the same 16-core CPU but a 40-core GPU.

I'd expect the 12-core M3 Pro to be 8 performance-cores, it would be unusual to have 6 efficiency cores on Pro and 4 on Max. Plus M2 Pro 10-core already has 6 performance cores, 4 efficiency. Moving to 6p/6e only increases the efficiency cores. 8p/4e would increase CPU performance at least 50%.

The GPU core counts look like a small increase so I'd say they will increase transistor count per core

Or, something like this is more sensible, which assumed each CPU perf complex is 4 cores, each CPU eff complex is 4 cores. So they would be using the same CPU perf and eff complexes from M3 all the way to M3 Ultra. It just varies by number of complexes and binning.

4+4+10
6+4+16
8+4+20
10+4+32
12+4+40

The big hole in the current Pro vs Max SoC lineup is that the Max doesn't offer more CPU perf cores than the Pro. Adding 2 or 4 more perf cores for the Max would provide a better upsell. So, good to hear the rumors are intimating this. The 18 to 20 GPU cores in the Pro though? Doesn't make sense. 10% isn't much of difference. It needs to be about 15% difference to be able to upsell on order $200. The only way they could make +10% work is if they combine it with a CPU core increase.

That number of CPU cores looks right, I can't think of any benefit to having 6 efficiency cores.

For the GPU core count, I'm guessing they must be significantly redesigned cores with more transistors each. In the mobile chips, A12 was 6-core:

https://en.wikipedia.org/wiki/Apple_A12

2p/4e CPU, 4 GPU cores, total 7b transistors

A16 is also 6-core CPU (2p/4e), 5-core GPU but 16b transistors:

https://en.wikipedia.org/wiki/Apple_A16

They must have roughly doubled the transistor count per core.

Since this is a new node, they will have redesigned the cores and in later revisions, they can increase the number of cores. 3nm allows 70% density increase so each core can increase that much, it may end up a little less. Lets say each GPU core has 40% more transistors and 10% more cores, then it's over 50% increase. Plus they are supposed to be faster at the same power on top of this.

I'm expecting 50% gain for CPU/GPU from M2->M3 at a lower power. Then with M4 they can add more cores at a higher power draw. 20-30TFLOPs GPU in a thin and light laptop where the fan barely comes on will be really nice.

The core count configurations are too hard to predict, and something looking right is often wrong. The 2 eff cores in the M1 Pro/Max generation was really from left field. The eff cores are really small and Apple deemed in necessary to only have 2? Just weird. They couldn't have been that desperate for die area. So, there must have been some other metric driving the decision that I haven't heard yet.

So, 6 efficiency cores are possible imo. It all depends on the prioritization of performance or runtime in the MBP models, or whatever metric they are using. The quibble I had was that the Max chips should have the same or more efficiency cores than the Pro chips. That part of the rumor doesn't make sense if Apple is doing the same thing they did with prior Pro and Max chips. The specific number of eff cores could mean the Max only has 2 more perf cores as the total number cores could be gleaned from GB testing. 10 versus 12 performance cores will impact upsell.

You probably only should expect 20% level performance gains per CPU core at the same power consumption, with a good portion of it from higher clocks. The days of 50% gain is over. For GPU cores, I can see 100% for some benchmarks if hardware raytracing is there, but aggregate benchmarks are going to be 30% maybe.