Apple: M1 vs. M2

[Cmaier]

A nice summary table from 9to5mac (other than the typos ) :

M1 versus M2 chip: Here's everything we know so far

AnandTech has taken a deep dive into the new M2 chip announced yesterday, focusing in particular on the M1 versus M2 chip performance. These chips are available in the all-new MacBook Air, and in an updated version of the entry-level 13-inch MacBook Pro. The site says that while Apple has been...

9to5mac.com

 

[Andropov]

Given the 1500€ price tag of the M2 MacBook Air in Europe, it makes sense that it includes the ProRes and ProRes RAW codecs in the Media Engine, since its price is well into the market segment of prosumer laptops.

 

[Cmaier]

Given the 1500€ price tag of the M2 MacBook Air in Europe, it makes sense that it includes the ProRes and ProRes RAW codecs in the Media Engine, since its price is well into the market segment of prosumer laptops.

It’s edging into being a really good content creation machine for most people.

 

[Colstan]

This looks like a decent upgrade, better than some had expected. I wonder if they revved the clock higher than 3.2Ghz? Apple seems much more concerned with IPC, unlike the PC guys with their upcoming 5.4Ghz space heaters and 600W graphics cows.

I still use an Intel Mac mini, like a savage. I'm waiting for M3, but this appears impressive, and I'm curious about the Mac Pro, even though I'm definitely not the target market.

 

[Andropov]

The fact that Intel's 12 core CPUs (which can be found in PC laptops in that price range) beat the M2 in peak performance is going to be the source of endless debates.

 

[Cmaier]

The fact that Intel's 12 core CPUs (which can be found in PC laptops in that price range) beat the M2 in peak performance is going to be the source of endless debates.

Well, endless *dumb* debates, sure.

 

[DT]

Not that it was a major performance concern (especially with the typical use case for a base config), but I see the M1 MBA is now only available in the 7-core GPU flavor.

 

[Cmaier]

Not that it was a major performance concern (especially with the typical use case for a base config), but I see the M1 MBA is now only available in the 7-core GPU flavor.

Makes sense. People who really needed that little extra oomph now have better choices.

 

[Renzatic]

It’s edging into being a really good content creation machine for most people.

I'm disappointed that the GPU still doesn't sport RT cores, but the M2 Air looks to be roughly as stout as a 14" MBP with an M1 Max. The only downside is that you can only max the machine out with 24GB RAM.

 

[Cmaier]

I'm disappointed that the GPU still doesn't sport RT cores, but the M2 Air looks to be roughly as stout as a 14" MBP with an M1 Max. The only downside is that you can only max the machine out with 24GB RAM.

They’re going to need a process shrink to be able to get RT into the die area and power envelope.

 

[Renzatic]

They’re going to need a process shrink to be able to get RT into the die area and power envelope.

They don't necessarily need to do a die shrink, though the addition would require more power, and better heat dispersion. It's something that would be a better fit for the Studio, rather than any of the laptops.

 

[Cmaier]

They don't necessarily need to do a die shrink, though the addition would require more power, and better heat dispersion. It's something that would be a better fit for the Studio, rather than any of the laptops.

RT tracing hardware, by my understanding, takes quite a bit of die area. My thought was that unless they were willing to make the die much bigger, which causes lots of other problems, they need a shrink. Also, without a shrink, it may be impossible to even fit the M2 Max with RT into the reticle.

 

[Cmaier]

Also should go without saying: no Arm v9 on M2. So tell your MR friends so that they have something to complain about

 

[Renzatic]

RT tracing hardware, by my understanding, takes quite a bit of die area. My thought was that unless they were willing to make the die much bigger, which causes lots of other problems, they need a shrink. Also, without a shrink, it may be impossible to even fit the M2 Max with RT into the reticle.

I'm far from an expert on this, but I've always assumed that the RT cores would probably take up roughly the same amount of space as the Neural Engine, which would lead to a ~25% increase in the GPU's size.

This is all semi-uneducated guessing on my part, since I equate the Neural Engine to being somewhat similar in form and function to the Tensor Cores on a Geforce chip, which take up the same amount of space as the RT cores.

 

[Cmaier]

I'm far from an expert on this, but I've always assumed that the RT cores would probably take up roughly the same amount of space as the Neural Engine, which would lead to a ~25% increase in the GPU's size.

This is all semi-uneducated guessing on my part, since I equate the Neural Engine to being somewhat similar in form and function to the Tensor Cores on a Geforce chip, which take up the same amount of space as the RT cores.

Could be. I’ve never designed either kind of hardware so I claim no expertise in the matter. Someone just told me once that RT would take space.

 

[theorist9]

They’re going to need a process shrink to be able to get RT into the die area and power envelope.

Why would they need to fit the current die area? Couldn't they just make the die bigger, as they did in going from the M1 to the M2? Or is the issue that they could fit RT on the M2 by expanding the die, but wouldn't be able to do it on the M2 Max, because the size of the M1 Max die is already close to the upper limit for their process (plus power concerns and cost)?

When they go to 3 nm, they will have higher density and efficiency. Might that be when they introduce RT?

 

[Cmaier]

Why would they need to fit the current die area? Couldn't they just make the die bigger, as they did in going from the M1 to the M2? Or is the issue that they could fit RT on the M2 by expanding the die, but wouldn't be able to do it on the M2 Max, because the size of the M1 Max die is already close to the upper limit for their process (plus power concerns and cost)?

When they go to 3 nm, they will have higher density and efficiency. Might that be when they introduce RT?

Well, first, M2 would be even bigger than it is, and that costs them money, of course. And, yeah, I was more thinking about M1 Max being close to the reticle size (I think I mentioned that above). It might be fine, but it depends on how much space is actually required to add the hardware.

I’m pretty sure 3nm will bring RT, though, along with Arm v9. So middle of next year, I think.

 

[Yoused]

Maybe they will implement PIG: Progressive Image Generation, working on a principle similar progressive jpegs, and feed the output through the neural engine to assess which areas are more homogenous and which need finer rendering. If that could be implemented with effective hardware acceleration, it might significantly improve speed and efficiency versus straight raster RT/PT.

 

[leman]

The fact that Intel's 12 core CPUs (which can be found in PC laptops in that price range) beat the M2 in peak performance is going to be the source of endless debates.

Le't wait for the benchmarks. I have a suspicion that M2 will score around 2000 in GB5 which will put it ahead of any mobile Alder Lake. And frankly, none of the current Alder Lake-P SKUs even beats M1 in single-threaded performance. Folks like to point out that high-end desktop Golden Cove is faster, but that comes with extreme per-core power draws not achievable on a laptop.

 

[Yoused]

I made a little chart to look at some Apple Silicon performance metrics based on GeekBench 5, using iPad SoCs

	chip	cores	GB5 score	GB5 multicore	GHz	score per Ghz	multicore per core
2013	A7	2	278	526	1.4	198.6	94.6%
2014	A8X	3	378	1049	1.5	252	92.5%
2015	A9X	2	648	1195	2.2	294.5	92.2%
2016	A9X	2	643	1176	2.1	306.2	91.4%
2017	A10X	6	831	2264	2.3	361.3	45.4%
2018	A12X	8	1113	4607	2.5	445.2	51.7%
2019	A12Z	8	1116	4617	2.5	446.4	51.7%
2020	A14	8	1584	4124	3.0	528	32.5%
2021	M1	8	1708	7145	3.2	533.8	52.3%

The last column is kind of silly: if the multicore score reflected single-core times core count, it would be 100% – at 2017, it falls off a lot because the SoC is Big.little and the single core score is for the big core.

The second column from the right is the interesting one: the single core score divided by the clock speed. It clearly shows the progress in big core performance efficiency. Clock speed rises steadily as the die process shrinks, but core performance has been rising even faster – core efficiency has increased by more than two-and-two-thirds over the past nine years (M2 will probably be more efficient by a factor of 3).



As a side note, when Alder Lake is mapped into the last column, if you count 16 cores, the performance is a respectable 54.2%, but if you count the full capacity of 24 threads, it drops off to a sad 36.2%.