Does the M2 chip have the 32MB TLB cache limitation that the M1 line of chips has?

[Marswarrior462]

I’ve been watching Max Tech, which is in my opinion, one of the best Apple tech channels besides EverythingApplePro, and they discovered that because the M1 chips were developed up to 7 years in advance, Apple underestimated a serious problem with cache limitations that has caused the performance, especially on the graphics side, to scale horribly, especially in the case of the M1 Max and even more so with the M1 Ultra. As a result of the 32MB TLB cache limitations, the unbinned Mac Studio is barely any faster than the binned M1 Ultra. Now that Apple has introduced the M2, I’m wondering if that limitation is fixed, but I’ve also heard that due to complications with the chip shortage, Apple was forced to settle with a more minor upgrade for the M2 chip and the real upgrade will be the M3 line of chips, which has me concerned that the cache limitation is still there. I’ve been wanting to upgrade my obsolete mid-2012 Retina 15-inch MacBook Pro to a new 16.2 inch model for quite some time now, but skipped the M1 Max because I couldn’t afford it. Should I save up for the M2 Max MacBook Pro or wait for the M3 Max? If the cache limitation is not there, I’m waiting for the M3 Max. Also, how bad is the problem on lower end chips like the base M1 and M2 chips? Will the cache limitation throttle the performance of low-end Macs and high-end iPads as well?

 

[ian87w]

I’ve been watching Max Tech, which is in my opinion, one of the best Apple tech channels besides EverythingApplePro, and they discovered that because the M1 chips were developed up to 7 years in advance, Apple underestimated a serious problem with cache limitations that has caused the performance, especially on the graphics side, to scale horribly, especially in the case of the M1 Max and even more so with the M1 Ultra. As a result of the 32MB TLB cache limitations, the unbinned Mac Studio is barely any faster than the binned M1 Ultra. Now that Apple has introduced the M2, I’m wondering if that limitation is fixed, but I’ve also heard that due to complications with the chip shortage, Apple was forced to settle with a more minor upgrade for the M2 chip and the real upgrade will be the M3 line of chips, which has me concerned that the cache limitation is still there. I’ve been wanting to upgrade my obsolete mid-2012 Retina 15-inch MacBook Pro to a new 16.2 inch model for quite some time now, but skipped the M1 Max because I couldn’t afford it. Should I save up for the M2 Max MacBook Pro or wait for the M3 Max? If the cache limitation is not there, I’m waiting for the M3 Max. Also, how bad is the problem on lower end chips like the base M1 and M2 chips? Will the cache limitation throttle the performance of low-end Macs and high-end iPads as well?

Typical questions (to see what you need, not to justify M1 vs M2): what do you use your Macbook Pro for? If your 2012 actually can still do the job, then you probably don't even need the Max Apple Silicon. Also, what's your budget?

If large screen is your priority, then right now, your only option is to sve up for the 16" Macbook Pro (unless Apple released the fabled 15" Macbook Air). If you already have the money, the current M1 Pro/Max Macbook Pro should be leaps and bounds over your 2012 Macbook Pro. If you are saving some money up, then wait until the M2 Pro/Max version is released.

Whatever technical limitation of the M1 and M2 you mentioned, probably won't be a significant thing in real life, especially when we compare it to 2012 intel Macbook Pro.

 

[Marswarrior462]

Typical questions (to see what you need, not to justify M1 vs M2): what do you use your Macbook Pro for? If your 2012 actually can still do the job, then you probably don't even need the Max Apple Silicon. Also, what's your budget?

If large screen is your priority, then right now, your only option is to sve up for the 16" Macbook Pro (unless Apple released the fabled 15" Macbook Air). If you already have the money, the current M1 Pro/Max Macbook Pro should be leaps and bounds over your 2012 Macbook Pro. If you are saving some money up, then wait until the M2 Pro/Max version is released.

Whatever technical limitation of the M1 and M2 you mentioned, probably won't be a significant thing in real life, especially when we compare it to 2012 intel Macbook Pro.

I plan to do video and photo editing and learning how to do machine learning and 3D modelling with courses. I also want to have a nice MacBook Pro that will last me a very long time so I don’t have to spend that kind of money again anytime soon after I buy it. Until then, I’ll settle with an M2 iPad Pro until I can get a MacBook Pro. I’m also holding off for cellular connectivity and upgraded ports (hopefully with HDMI 2.1 and UHS-III)

 

[russell_314]

I like Max Tech but much of the content is clickbait and fluff. He does have some good info but when I watch a video of him trying to compare a Mac to another one he doesn't have...

 

[Gnattu]

Except MaxTech is NOT a reliable source of technology information.

He does not even correctly understand how things work but still making nonsense claims. This is typical for a Youtuber but is also really toxic.

What is a TLB? Well, Mr. Yuryev called it "Transaction lookaside buffer", this is so wrong, the correct name is Translation lookaside buffer. If he understands what is virtual memory, what is physical memory, and how CPU/OS manages the memory, he would not call it wrong.

Plus, he even cited an article from Alyssa Rosenzweig, an experienced GPU driver developer who is writing driver for M1 GPUs, to "prove" his point. However, that article's content has nothing to do with the TLB. It is talking about TVB/PV and that is dynamically resized to avoid the performance hit after a coupe of frames.

Mr. Yuryev from Max Tech does not know what he is talking about, and he does not know what the professionals are talking about. Are you still worrying about the clickbait issue he mentioned in his video now?

 

[russell_314]

Except MaxTech is NOT a reliable source of technology information.

He does not even correctly understand how things work but still making nonsense claims. This is typical for a Youtuber but is also really toxic.

What is a TLB? Well, Mr. Yuryev called it "Transaction lookaside buffer", this is so wrong, the correct name is Translation lookaside buffer. If he understands what is virtual memory, what is physical memory, and how CPU/OS manages the memory, he would not call it wrong.

Plus, he even cited an article from Alyssa Rosenzweig, an experienced GPU driver developer who is writing driver for M1 GPUs, to "prove" his point. However, that article's content has nothing to do with the TLB. It is talking about TVB/PV and that is dynamically resized to avoid the performance hit after a coupe of frames.

Mr. Yuryev from Max Tech does not know what he is talking about, and he does not know what the professionals are talking about. Are you still worrying about the clickbait issue he mentioned in his video now?

The problem is most people even the more nerdy ones like me have no clue what that is so he can just use fancy words to sound good. It gets views and likes so that means $$$$$$ on YouTube

 

[leman]

I’ve been watching Max Tech, which is in my opinion, one of the best Apple tech channels

It has already been pointed out that Maxim is not a very good source of technical information and the particular video you are referencing is full of confusion and improper use of technical terms

Here is a discussion of this video on Hacker News, a reputable technology communication platform, with real software and hardware experts. You will notice that they are as puzzled about Maxim's claims as we are: https://news.ycombinator.com/item?id=31020484

As a result of the 32MB TLB cache limitations, the unbinned Mac Studio is barely any faster than the binned M1 Ultra.

What does this even mean? TLB is not measured in megabytes, it is measured in number of entries. Ok, let us just assume that Maxim misspoke and what he really meant was that the TBL is only large enough to cover a range of 32MB. Given 16KB pages on Apple hardware, this gives you a TLB with 2'048 entries. I am not quite sure how they arrived at this number, since Apple does not publish this information and they did not offer any information how exactly they measured it. At any rate, a TLB of that size is absolutely in line with the hardware from other vendors, and I have no idea why someone would claim that it is "too small". Besides, the point of the TLB is to cover the available cache, and 32MB is more than plenty for M1 designs.

If you want to look for reasons why M1 Ultra doesn't seem to perform well in practice — here is a much simpler one: the software is likely not optimised for this kind of hardware. M1 Ultra has 8'192 ALUs, you need to give it a steady supply of a massive amount of work to saturate the GPU properly. I suspect that most applications don't do that, simply because they were not tested with that kind of hardware. Here is a tech talk directly from Apple — not a sensationalist YouTuber — where they discuss the problem and possible solutions in more details: https://developer.apple.com/videos/play/wwdc2022/10159/ In particular, see how interleaving multiple work packages can dramatically improve scaling.

 

[darngooddesign]

...Should I save up for the M2 Max MacBook Pro or wait for the M3 Max?

The M2 MBPs won't be out until second half of 2023 and the M3 MBPs wont be out until second half od 2025. There will not be an Ultra laptop.

At some point you have to buy a computer.

 

[jav6454]

Except MaxTech is NOT a reliable source of technology information.

He does not even correctly understand how things work but still making nonsense claims. This is typical for a Youtuber but is also really toxic.

What is a TLB? Well, Mr. Yuryev called it "Transaction lookaside buffer", this is so wrong, the correct name is Translation lookaside buffer. If he understands what is virtual memory, what is physical memory, and how CPU/OS manages the memory, he would not call it wrong.

Plus, he even cited an article from Alyssa Rosenzweig, an experienced GPU driver developer who is writing driver for M1 GPUs, to "prove" his point. However, that article's content has nothing to do with the TLB. It is talking about TVB/PV and that is dynamically resized to avoid the performance hit after a coupe of frames.

Mr. Yuryev from Max Tech does not know what he is talking about, and he does not know what the professionals are talking about. Are you still worrying about the clickbait issue he mentioned in his video now?

Goes to show that not everything that shines id Gold. I for one, think thought that OP was trying to reference CPU level caches

 

[CraigJDuffy]

I plan to do video and photo editing and learning how to do machine learning and 3D modelling with courses. I also want to have a nice MacBook Pro that will last me a very long time so I don’t have to spend that kind of money again anytime soon after I buy it. Until then, I’ll settle with an M2 iPad Pro until I can get a MacBook Pro. I’m also holding off for cellular connectivity and upgraded ports (hopefully with HDMI 2.1 and UHS-III)

An m2 iPad Pro will cost more than a m2 mba and do less and last less time…

 

[fameios-phil]

The M2 MBPs won't be out until second half of 2023 and the M3 MBPs wont be out until second half od 2025. There will not be an Ultra laptop.

You don't know that. There have been MacBook releases as short as 6 months apart in the past (not that it is going to be THAT short). The M3 will probably be that far out but...

They could have the M2 Pro/Max already queued up and ready to launch Fall 2022 or Spring 2023. The main difference between the M1 Pro and M2 is the additional CPU and GPU cores. Now that both have the ProRes engine I would expect the M2 Pro to be 8 performance cores and either 2 or 4 efficiency cores and increase the GPU up to 20.

It is actually in Apple's best interest to swap and release the Pro versions before the consumer versions.

Intel for example always releases the new high end chips first and then rolls out the more mass market consumer versions after the manufacturing changes impacts have been minimized.

The fact that they didn't launch an M2 Pro at the same time or before the M2 makes little sense from a business standpoint. Why not upsell to the Pro that is available "now" instead of the cheaper version that is coming in a few months?

I think a lot of Apple's plans have been impacted by the manufacturing shortages since Covid hit and there is a lot of hidden info that might make their launch decisions make more sense.

 

[ian87w]

I plan to do video and photo editing and learning how to do machine learning and 3D modelling with courses. I also want to have a nice MacBook Pro that will last me a very long time so I don’t have to spend that kind of money again anytime soon after I buy it. Until then, I’ll settle with an M2 iPad Pro until I can get a MacBook Pro. I’m also holding off for cellular connectivity and upgraded ports (hopefully with HDMI 2.1 and UHS-III)

Why waste money on the iPad Pro? Keep using your current MacBook and put the money into getting the MacBook Pro that you want. I’d say don’t bother with the expensive iPad, especially if you’re concerned with longevity.

If you have the money now, just get the 16” MacBook Pro. Done, and you can continue doing things that you want instead of wasting time and money on the iPad. IMO in your stage, even the M2 MacBook Air would be more useful than an iPad. If you don’t mind a smaller screen, just get the M2 MacBook Air and an external monitor. Better investment imo than buying an iPad.

 

[dgdosen]

I like Max Tech but much of the content is clickbait and fluff. He does have some good info but when I watch a video of him trying to compare a Mac to another one he doesn't have...

ha! ... clickbait and fluff... I can probably count on one hand the number of Mac videos I've watched on Youtube that I didn't regret moments after starting...

And of course, that makes me think of:

 

[ian87w]

Just to add, I used to watch both MaxTech and EverythingApplePro. Their general positive enthusiasm were enjoyable to watch compared to other toxic gaslighting youtubers. However, after a while, I realized that they are mostly talking about nothing. Other than actual benchmarks and test that they did, their analysis and opinions are generally empty and even made up. So I stopped watching them routinely. I maybe catch MaxTech once in a while for their benchmark testing and camera comparisons, but their other content are generally well produced of nothing, at least in terms of actual tech analysis.

 

[russell_314]

ha! ... clickbait and fluff... I can probably count on one hand the number of Mac videos I've watched on Youtube that I didn't regret moments after starting...

And of course, that makes me think of:

Well I don’t mean to say all of their videos are terrible but you have to take some of them with a grain of salt. They give some good information sometimes but they try it seems like they try to hard to the point of making stuff up when they don’t know. It’s like most things on the Internet you have to separate the information from the crap

 

[thenewperson]

Just to add, I used to watch both MaxTech and EverythingApplePro. Their general positive enthusiasm were enjoyable to watch compared to other toxic gaslighting youtubers. However, after a while, I realized that they are mostly talking about nothing. Other than actual benchmarks and test that they did, their analysis and opinions are generally empty and even made up. So I stopped watching them routinely. I maybe catch MaxTech once in a while for their benchmark testing and camera comparisons, but their other content are generally well produced of nothing, at least in terms of actual tech analysis.

Yeah if you watch MaxTech enough you simply notice a lot of their theories are hot air (Still waiting for M1 Macs to rule gaming lol).

The newest one I've seen floated by Vadim is that the M2 we have now is actually an M1.5 that was quickly whipped up due to 3nm being delayed till next year so the real M2 (which will now be the M3) will simply be delayed until 3nm is in sufficient quantity next year.

Wild.

 

[joema2]

...Max Tech...they discovered...cache limitations that has caused the performance, especially on the graphics side, to scale horribly, especially in the case of the M1 Max and even more so with the M1 Ultra. As a result of the 32MB TLB cache limitations, the unbinned Mac Studio is barely any faster than the binned M1 Ultra...

With GPU scaling it's possible TLB cache can impact performance. That is why Apple mentioned it this year in the WWDC22 talk mentioned above by Leman. In that talk they revealed they are adding a perf. counter for the Memory Management Unit's TLB miss rate.

But even if that was an issue I don't see why it would be worse for the M1 Ultra than the M1 Max. The Ultra is composed of two M1 Max dies, and each one has a separate MMU and TLB cache. IOW the Ultra has (up to) twice the physical RAM but also twice the TLB cache.

The other scalability problem observed by Max Tech and others relates to M1 Max/Ultra limited improvement on video encode/decode workflows. However most of their tests are single stream, and it's difficult (though not impossible) for multi-threaded software to segment an input or output video file and dispatch those segments to multiple encode/decode units. To a degree Apple's Compressor already does this for ProRes files, if on Ultra you configure Compressor's advanced preferences for multiple instances. Those ProRes segments are split across multiple processes and concatenated for final output to a single file. However in my tests it has limited performance benefit, so something is serialized and limiting scalability. But it's very unlikely TLB cache.

If you test 4-stream decode in FCP for a difficult codec, that is significantly faster and smoother on M1 Ultra than M1 Max. On Ultra the viewer update rate and timeline smoothness is considerably better than M1 Max. That is for a workload without any effects and which is not limited by GPU, CPU or I/O. So even now the M1 Ultra has some encode/decode benefit over M1 Max, but software improvements at the application, framework and system layers could probably further improve things.

 

[EugW]

It has already been pointed out that Maxim is not a very good source of technical information and the particular video you are referencing is full of confusion and improper use of technical terms

Here is a discussion of this video on Hacker News, a reputable technology communication platform, with real software and hardware experts. You will notice that they are as puzzled about Maxim's claims as we are: https://news.ycombinator.com/item?id=31020484





What does this even mean? TLB is not measured in megabytes, it is measured in number of entries. Ok, let us just assume that Maxim misspoke and what he really meant was that the TBL is only large enough to cover a range of 32MB. Given 16KB pages on Apple hardware, this gives you a TLB with 2'048 entries. I am not quite sure how they arrived at this number, since Apple does not publish this information and they did not offer any information how exactly they measured it. At any rate, a TLB of that size is absolutely in line with the hardware from other vendors, and I have no idea why someone would claim that it is "too small". Besides, the point of the TLB is to cover the available cache, and 32MB is more than plenty for M1 designs.

If you want to look for reasons why M1 Ultra doesn't seem to perform well in practice — here is a much simpler one: the software is likely not optimised for this kind of hardware. M1 Ultra has 8'192 ALUs, you need to give it a steady supply of a massive amount of work to saturate the GPU properly. I suspect that most applications don't do that, simply because they were not tested with that kind of hardware. Here is a tech talk directly from Apple — not a sensationalist YouTuber — where they discuss the problem and possible solutions in more details: https://developer.apple.com/videos/play/wwdc2022/10159/ In particular, see how interleaving multiple work packages can dramatically improve scaling.

The OP has brought this topic up more than once. I'm not sure why s/he is so concerned about this, but I guess we can blame Max Tech at least in part.

 

[russell_314]

The OP has brought this topic up more than once. I'm not sure why s/he is so concerned about this, but I guess we can blame Max Tech at least in part.

That’s like everything people focus on stuff. A lot of people focus on specifications but if you blindfolded them and asked them to tell you the difference between two computers with different specifications they probably couldn’t tell you. It’s something that looks good on paper but can use it doesn’t really make a difference. At least not one that’s observable by human perception.

Max tech is a fairly decent channel but you have to take things with a grain of salt. When it comes down to it the ultimate goal is views so if that means sensationalizing features then of course they’re going to do it. Making YouTube videos is a paying job so you make what people want to buy. The Apple review community is very crowded on YouTube so you have to set yourself apart. They do that by sensationalizing things