Aspects where Intel Macs still outperform the M1 models?

[pshufd]

I'm sorry, but what relevance does that have? He posed a question, I answered. You've just spouted fairly meaningless conjecture.

He said most use cases. Most users do not require a 5800X. The people on this forum are outliers.

 

[NotTooLate]

Most people wouldn't notice the difference.

My task manager stats:

In use 31.7 GB
Committed 39.2/73.4 GB
Cached 31.0 GB

This is without my two main daily work programs running. It's nice to know that I could run five or six virtual machines at the same time if I wanted to, and I have done this in the past for testing.

The point I was going to make is such , if I put the 64GB memory on a single channel and put 8GBx4 channels for the 32GB I will be getting 4x the BW (assuming the memory controller and uncore are capable enough) , having bigger sized DRAM that needs to open and close banks all the time is a MEGA penalty on the system and a single channel system will be BW starved and will lose almost every benchmark you ever do between the systems , more DRAM is not always better , and the I3 that were mentioned before have either a single or dual channel configurations traced on the motherboard which are bad.

Second point i wanted to make in regards to "soldered vs non soldered" , Apple uses high BW low power memories for the CPU`s , the LP4 memories are NOT available in a DIMM configuration , and you cannot buy them as such , so its not that Apple "soldered them down" to prevent upgrades or any of that , its just not a DIMM like we used normally in the desktop environment , you can argue that you want a desktop memory inside your laptop for the upgradeability , but this will cost BW , size , active power (more pins , more voltage for the cells , more voltage for the logic) and refresh power cost are also higher , so you trade capacity for a lot of advantages you get from the LPDDR.

Lastly , the LP4 memory module is much smaller , so on a given channel you can put less DRAM capacity per module (I think its something like 1GB/2GB per module , at least according to the internet/wiki !!) if you use LP4 vs DDR4 , this is again a design decision by Apple , but should be taken into account when analyzing the memory subsystem capabilities.

TLDR - when ppl say Apple 8GB is more then Intel 8GB , a big reason for it is that Apple uses more channels then Intel in those machines , thus having much better BW and less bank/page hotspots that can cause performance issues.

It will be interesting to see the BW numbers Apple will post in the more performant chips , as this will determine the performance of the machine , much more then the max capacity , which I assume will also increase.

Nvidia new Grace CPU`s will use LP5 memories and not the DDR5 desktop class memories , you can bet they will be "non upgradeable" , which is the way those memories are built.

 

[pshufd]

The point I was going to make is such , if I put the 64GB memory on a single channel and put 8GBx4 channels for the 32GB I will be getting 4x the BW (assuming the memory controller and uncore are capable enough) , having bigger sized DRAM that needs to open and close banks all the time is a MEGA penalty on the system and a single channel system will be BW starved and will lose almost every benchmark you ever do between the systems , more DRAM is not always better , and the I3 that were mentioned before have either a single or dual channel configurations traced on the motherboard which are bad.

Second point i wanted to make in regards to "soldered vs non soldered" , Apple uses high BW low power memories for the CPU`s , the LP4 memories are NOT available in a DIMM configuration , and you cannot buy them as such , so its not that Apple "soldered them down" to prevent upgrades or any of that , its just not a DIMM like we used normally in the desktop environment , you can argue that you want a desktop memory inside your laptop for the upgradeability , but this will cost BW , size , active power (more pins , more voltage for the cells , more voltage for the logic) and refresh power cost are also higher , so you trade capacity for a lot of advantages you get from the LPDDR.

Lastly , the LP4 memory module is much smaller , so on a given channel you can put less DRAM capacity per module (I think its something like 1GB/2GB per module , at least according to the internet/wiki !!) if you use LP4 vs DDR4 , this is again a design decision by Apple , but should be taken into account when analyzing the memory subsystem capabilities.

TLDR - when ppl say Apple 8GB is more then Intel 8GB , a big reason for it is that Apple uses more channels then Intel in those machines , thus having much better BW and less bank/page hotspots that can cause performance issues.

It will be interesting to see the BW numbers Apple will post in the more performant chips , as this will determine the performance of the machine , much more then the max capacity , which I assume will also increase.

Nvidia new Grace CPU`s will use LP5 memories and not the DDR5 desktop class memories , you can bet they will be "non upgradeable" , which is the way those memories are built.

My wife has a 2018 i3 Mini with 8 GB of RAM. She does email, watches videos and browses the web. And her Mini is fine for those purposes. We all complain about RAM but there's a reason why Apple sells these configurations - they sell a ton of them. Even on the iMac 21.5 and the 27 - the base is 8 and I see lots of them for resale with the original 8 GB of RAM.

 

[NotTooLate]

My wife has a 2018 i3 Mini with 8 GB of RAM. She does email, watches videos and browses the web. And her Mini is fine for those purposes. We all complain about RAM but there's a reason why Apple sells these configurations - they sell a ton of them. Even on the iMac 21.5 and the 27 - the base is 8 and I see lots of them for resale with the original 8 GB of RAM.

Sure , but my case in hand , is that 8GB of Intel != Apple`s 8GB , your wife Mini is using a single channel 8GB stick or 2x4GB (best case) , while Apple are using more channels for their 8GB`s and getting much better memory BW out of it, you can read anandtech review on the M1 memory BW (68GB/s!!!!) , its impressive for such a small machine and one of the reason Apple`s 8GB are better then Intel 8GB.

For reference in wikichip , Ryzen 5900X has 2 channels max for DDR4 memory with 48GB/s max total BW....

 

[pshufd]

Sure , but my case in hand , is that 8GB of Intel != Apple`s 8GB , your wife Mini is using a single channel 8GB stick or 2x4GB (best case) , while Apple are using more channels for their 8GB`s and getting much better memory BW out of it, you can read anandtech review on the M1 memory BW (68GB/s!!!!) , its impressive for such a small machine and one of the reason Apple`s 8GB are better then Intel 8GB.

For reference in wikichip , Ryzen 5900X has 2 channels max for DDR4 memory with 48GB/s max total BW....

Yes. But for the majority of PC users, none of this matters.

 

[GiantKiwi]

Yes. But for the majority of PC users, none of this matters.

And yet you keep bleating on about how much it doesn't matter.

 

[Bruninho]

My short and direct answer to the question in this thread: None.

 

[pshufd]

And yet you keep bleating on about how much it doesn't matter.

It doesn't matter for most people. But the people here generally have more needs than the typical user.

 

[Yebubbleman]

As I know you are a tech oriented person , i think we can have a technical discussion that wont go over your head like so many folks in this and other forums.

I present to you 2 systems , please tell me which has the better performance for most use cases :
1) 64GB memory + AMD Ryzen 5800x.
2) 32GB memory + AMD Ryzen 5800x.

Everything in the 2 computers are equal aside from the memory , i would also give you "same DRAM timings" on both memories.

I am not trolling , I promise!

I am unsure as to the point of your question?

 

[Idec50]

I can't tell if it's Wifi or what, but the load times for Websites is all over the place. It's either nearly instant or takes a long time (I even wonder if it's pre-loading some Websites based on history). My wired Windows desktop is very consistent. Actually, when this is connected to the ethernet it still varies. It could be related to having too many tabs open and only 8gb RAM. Safari needs a tab sleep setting like Microsoft Edge. It's funny how overall now Microsoft is improving on software while Apple improves on hardware (Microsoft has Edge, Windows Subsystem for Linux, VSCode, Github, Xbox Game Pass).

My 2017 MBP loaded Websites a little slower on average but more consistently than this machine. Perhaps also because it had 16gb, it could run more in the background, like various VMs and resource hog applications. This machine does slow down when I really push it with say 12 applications and I'm working as fast as possible. 80% of the time it performs better. I can't even use my 2017 MBP now because it runs way too hot and loud. Overall, an amazing improvement but I can't wait to grab the fully loaded M2 with 16 cores, 32gb ram, 2tb+. Also can't wait for R 4.1 to come out with native AS support, but it runs fine under emulation.

 

[NotTooLate]

I am unsure as to the point of your question?

So many ppl are hung up on capacity , where they forget that a lot of the memory performance comes from bandwidth and latency , Apple , as shows in Anandtech review , has stellar BW while a lot of ppl read 16gb is less then 64gb , thus 64gb will be a better machine from memory POV , my example was provided in this thread , I won’t copy paste it , but Apple has a wide 8 channel memory subsystem vs 2 channel systems from x86 folks for anything that is in a laptop and most desktop , they can put more capacity due to the nature of DDR vs LPDDR , but they can’t match the bw.

 

[Yebubbleman]

So many ppl are hung up on capacity , where they forget that a lot of the memory performance comes from bandwidth and latency , Apple , as shows in Anandtech review , has stellar BW while a lot of ppl read 16gb is less then 64gb , thus 64gb will be a better machine from memory POV , my example was provided in this thread , I won’t copy paste it , but Apple has a wide 8 channel memory subsystem vs 2 channel systems from x86 folks for anything that is in a laptop and most desktop , they can put more capacity due to the nature of DDR vs LPDDR , but they can’t match the bw.

Machine A and Machine A with more RAM is always a match that favors the latter. You might not see a day-to-day benefit for many years, but you do eventually and inevitably see a benefit. Apple has deprecated support for both iOS and iPadOS based on RAM at many points (including the last time the minimum system requirements for iOS and iPadOS changed). They will inevitably require greater RAM for new OS releases on the Mac as well.

 

[NotTooLate]

Machine A and Machine A with more RAM is always a match that favors the latter. You might not see a day-to-day benefit for many years, but you do eventually and inevitably see a benefit. Apple has deprecated support for both iOS and iPadOS based on RAM at many points (including the last time the minimum system requirements for iOS and iPadOS changed). They will inevitably require greater RAM for new OS releases on the Mac as well.

Your statement is of course FALSE which was the point i was trying to make , even in a Machine A vs Machine A , more RAM is not always better , this point is more adamant when we compare different systems as ppl usually do with the M1 vs Intel.

In the PC world , you put a 64GB DIMM stick vs putting (8GB) x (4 channels) ,and you will find out FAST how BW trumps capacity for most usecases , on the edge case where you run a single program that will use the single DIMM BW to the max and nothing else , a 64GB might (depending on the app , if its BW starved or not) better then my 8GBx4 configuration , but in a normal user experience the single channel memory will struggle to keep up as it has 1/4 of the memory BW.

Apple uses 8 channel LP4 memory memory subsystem (even on the 8GB configuration i assume , looking at Anandtech review , the BW is crazy high , this is the main point that makes the "magic" happen for most users and use case , in additional to the unified memory arch that reduce overhead of copying data needlessly.

The question at point was not M1 8GB vs M1 16GB , as of course the 16GB will outperform the 8GB , is the fact that ppl mix up capacity of an Intel machine vs M1 , which uses different memory and memory topology.

So again , to finalize , an Intel mac mini who uses a single or dual channel DDR , will be FAR behind M1 in BW and unless you use a single non BW starved app that can leverage the extra capacity efficiently you will be better off with 16GB M1 vs 32GB Intel for most of the usage of the machine in any given time (aside from the corner case i just mentioned).

 

[Yebubbleman]

Your statement is of course FALSE which was the point i was trying to make , even in a Machine A vs Machine A , more RAM is not always better , this point is more adamant when we compare different systems as ppl usually do with the M1 vs Intel.

In the PC world , you put a 64GB DIMM stick vs putting (8GB) x (4 channels) ,and you will find out FAST how BW trumps capacity for most usecases , on the edge case where you run a single program that will use the single DIMM BW to the max and nothing else , a 64GB might (depending on the app , if its BW starved or not) better then my 8GBx4 configuration , but in a normal user experience the single channel memory will struggle to keep up as it has 1/4 of the memory BW.

Apple uses 8 channel LP4 memory memory subsystem (even on the 8GB configuration i assume , looking at Anandtech review , the BW is crazy high , this is the main point that makes the "magic" happen for most users and use case , in additional to the unified memory arch that reduce overhead of copying data needlessly.

The question at point was not M1 8GB vs M1 16GB , as of course the 16GB will outperform the 8GB , is the fact that ppl mix up capacity of an Intel machine vs M1 , which uses different memory and memory topology.

So again , to finalize , an Intel mac mini who uses a single or dual channel DDR , will be FAR behind M1 in BW and unless you use a single non BW starved app that can leverage the extra capacity efficiently you will be better off with 16GB M1 vs 32GB Intel for most of the usage of the machine in any given time (aside from the corner case i just mentioned).

Bandwidth is not everything. You open up every chrome tab and spin up all the VMs, the bandwidth of the RAM won't mean anything because you'll still need the capacity of your RAM to be high. People act like the bandwidth of the RAM is enough to make up for the fact that every M1 configuration will cap out at 16GB. It won't. And if you have the ability to spec out an M1 with 16GB instead of 8GB, you should always do so because, at the rate they usually obsolete iOS and iPadOS devices, the maximum RAM capacity WILL matter.

 

[quarkysg]

Bandwidth is not everything. You open up every chrome tab and spin up all the VMs, the bandwidth of the RAM won't mean anything because you'll still need the capacity of your RAM to be high. People act like the bandwidth of the RAM is enough to make up for the fact that every M1 configuration will cap out at 16GB. It won't. And if you have the ability to spec out an M1 with 16GB instead of 8GB, you should always do so because, at the rate they usually obsolete iOS and iPadOS devices, the maximum RAM capacity WILL matter.

You're right that memory bandwidth is not everything, tho. it is important to improve overall system performance and responsiveness.

If a particular workload needs a lot of RAM, using fast but small amount of it will hurt the work through put, as memory swapping comes into play.

If a particular workload does not need gobs of RAM, using slow but a large amount of RAM will result is a less responsiveness.

The reason why memory standards are getting faster and faster is to solve the issue of alleviating data starvation that modern CPUs are facing, and also why CPUs are getting bigger caches.

Well, engineering have always been about trade offs. In the case of the released M1 Macs, Apple have decided that 16GB would be the limit and designed it according to their target market. I don't think there's any technical limit for the M1 SoCs to address more than 16GB of memory.

Apple could design the M1 memory controller to support DDR4 and went with socketed SO-DIMMs modules, but that would likely reduced the memory bandwidth by half. In this situation, I think the performance level obtained will probably not be very good compared to what they have now, especially for the GPU performance.

Similarly the M1 can be made to use more than 2 LPDDR4X RAM chips, thus allowing greater memory capacity. But this will likely cause additional latency (thus lowering performance) and increase cost, and probably be priced higher than what the target market might be willing to bear.

 

[leman]

Bandwidth is not everything. You open up every chrome tab and spin up all the VMs, the bandwidth of the RAM won't mean anything because you'll still need the capacity of your RAM to be high.

It's a balancing act. If you are in a situation where you need a lot of active RAM at one given moment (browser tabs don't really count because they an be seamlessly swapped in and out), capacity is not sufficient either — you also want bandwidth, large caches and more CPU cores. In a traditional desktop system with 128-bit memory bus, RAM capacity is subject to diminishing returns as the system simply cannot utilize all the RAM at the same time efficiently.

Apple could design the M1 memory controller to support DDR4 and went with socketed SO-DIMMs modules, but that would likely reduced the memory bandwidth by half.

DDR4 is not that much slower in absolute terms, but LPDDR4 offers better access granularity due to using smaller data batches and having more independent memory channels. So it ends up better for multithreaded workloads, especially when the GPU is considered.

Still, M1 is sill only using 128-bit LPDDR4 bus, so it doesn't offer any better RAM connection than a run of the mill Ice Lake/Tiger Lake laptop. The main reason for integrating memory on package is superior energy efficiency. M1's RAM uses laughably little power.

Similarly the M1 can be made to use more than 2 LPDDR4X RAM chips, thus allowing greater memory capacity. But this will likely cause additional latency (thus lowering performance) and increase cost, and probably be priced higher than what the target market might be willing to bear.

It really just boils down to cost and balance, I doubt latency would have been affected in any meaningful way. The main limitation of M1's memory capacity is the RAM density itself. The design tradeoff is fairly obvious: M1's choices of memory capacity are totally adequate for consumer-level computer usage (and yes, 8GB RAM is still more than sufficient, especially paired with super-efficient swapping of Apple Silicon) and single SoC package form factor allows for simplified manufacturing and logistics. Having a separate M1 variant with more RAM would either raise the prices by $100 across the board or the 32GB option would be priced at close to $800-1000, which wouldn't make much sense to anyone.

 

[drison]

Absolutely. Thanks to the efforts of the open source community and Homebrew, everything in my dev toolchain is completely ARM native. Only a few non-development-related stragglers running in Rosetta right now, like Dropbox.

Speaking of Dropbox... that was my biggest pain point in converting. The client continues to be a massive RAM and CPU hog. So, I had to significantly scale back the number of files managed there, as well as shift completely to Smart Sync, in order to keep it under control.

Other than that, the 16GB RAM ceiling has not been a problem at all. Memory Pressure rarely goes above 35% on my Mini, and the system feels extremely snappy, despite running all the same apps and tools that I used to run on my 64GB iMac Pro.

Thanks for your feedback. I currently run a Linux pc with 64GB as well as Macbook Pro with 32GB for work. As an enterprise Java developer, I was concerned that 16GB would not be enough. I know I won't be able to run several VMs but am considering using the new M1 for my development workstation.

Do you have any experience using Docker, MySQL, Elasticsearch, NPM, etc on the M1 machine? Any issues if so?

Thank you!
Dave

 

[jdb8167]

Thanks for your feedback. I currently run a Linux pc with 64GB as well as Macbook Pro with 32GB for work. As an enterprise Java developer, I was concerned that 16GB would not be enough. I know I won't be able to run several VMs but am considering using the new M1 for my development workstation.

Do you have any experience using Docker, MySQL, Elasticsearch, NPM, etc on the M1 machine? Any issues if so?

Thank you!
Dave

When you several VMs are you referring to different versions or running the VMs simultaneously? Most versions of the JDK are available natively after version 8. As for simultaneous JVMs that should actually work reasonably well considering how fast swap is on the M1 Macs. Is there something you want me to test?

 

[drison]

When you several VMs are you referring to different versions or running the VMs simultaneously? Most versions of the JDK are available natively after version 8. As for simultaneous JVMs that should actually work reasonably well considering how fast swap is on the M1 Macs. Is there something you want me to test?

Sorry, I was referring to OS VM's using VirtualBox or VMWare, etc. Sorry for the confusion. I tend to have to run a Windows VM and sometimes isolated Linux VMs for things like Elasticsearch, Oracle, etc. in addition to Java VMs for Tomcat, IntelliJ, etc.

It sounds like it should be a good machine for a Java developer except that I might need to run my OS VMs on another machine.

Thanks for the reply!

Dave

 

[Gudi]

Aspects where Intel Macs still outperform the M1 models?

• heat production
• fan noise
• energy consumption

 

[pshufd]

Aspects where Intel Macs still outperform the M1 models?

• heat production
• fan noise
• energy consumption

It was chilly last night so I was in my recliner, dressed warmly with a comforter on top of me and a 2014 MacBook Pro watching YouTube videos. The laptop was nicely toasty.

 

[jent]

Looping back just to say that the M1 Pro and M1 Max announcement today basically rendered moot my original points.

There is the exception of support for booting into x86_64 Windows and Linux, and although Parallels will support ARM Windows 11 on M1 Macs, I don't know if that support is considered official for the final OS version.

Apple is doing great with the new M1 versions.

 

[pshufd]

Looping back just to say that the M1 Pro and M1 Max announcement today basically rendered moot my original points.

There is the exception of support for booting into x86_64 Windows and Linux, and although Parallels will support ARM Windows 11 on M1 Macs, I don't know if that support is considered official for the final OS version.

Apple is doing great with the new M1 versions.

At this point, if you need to run Windows or Linux, just buy a separate machine. That's my situation right now.

 

[ADGrant]

Looping back just to say that the M1 Pro and M1 Max announcement today basically rendered moot my original points.

There is the exception of support for booting into x86_64 Windows and Linux, and although Parallels will support ARM Windows 11 on M1 Macs, I don't know if that support is considered official for the final OS version.

Apple is doing great with the new M1 versions.

As far as MacBooks are concerned most of your original points are moot. There are still desktop Macs that can be configured with more GPU power and while 64gb of RAM is enough for most uses, the 2020 iMac supports 128GB and the Mac Pro supports over a terabyte.

While Parallels may support Windows 11 on ARM running in a VM on a Mac, Microsoft does not and has no plans to.

 

[Madhatter32]

It was chilly last night so I was in my recliner, dressed warmly with a comforter on top of me and a 2014 MacBook Pro watching YouTube videos. The laptop was nicely toasty.

I have the same device. I enjoy it immensely -- especially in winter.