Concerning the "M1 is better with RAM" myth

[leman]

I wonder if the RAM is standard - that is if you could just buy them from an industrial supplier. I've never seen that form-factor available at Amazon.

It's standard LPDDR4X RAM from what we know, yes. I don't think you can buy them in regular shops, too you'll need to have a microchip supplier with access to this stuff, but it certainly should possible to procure them.

This was always my theory on this. With the faster SSDs, swapping isn't going to slow everything down as much as the previous generation of SSDs. Assuming this is the case I would expect to see the same performance if Apple were to put these faster SSDs in an Intel system. This would prove that, no, an M1 doesn't do magic.

It's not that much about the SSD speed, all of these SSDs are fast enough (and there are PC systems with faster SSDs on the market). It's about the algorithmic overhead of accessing memory pages that have been flushed to the SSD. Th paper I quote makes the observation that both Linux and Windows are badly optimized when it comes to this, probably since most of this code was designed long before fast SSDs were available and the cost of paging was dominated by very slow HDD access.

 

[Reason077]

I don't see how that's even remotely possible as the RAM is on the CPU chip itself...

No, it's not. Here's an image from an M1 Mac mini teardown of the M1 with the heat spreader removed. Note the two distinct Hynix RAM chips to the right hand side of the silver M1 SoC:

 

[Reason077]

That is fascinating in and of itself. Are we on the cusp of RAM and storage becoming indistinct?

The downside is that NVM/NAND doesn't last forever: the number of writes it can support are limited, and eventually the SSD will die.

My 8GB 2017 MacBook Pro managed to accumulate 119 TB of writes on it's 256GB SSD in 3.5 years of use, which according to smartmontools, has used up at least 15% of it's lifespan/spare cells:

SMART/Health Information (NVMe Log 0x02)

Available Spare: 79%
Available Spare Threshold: 2%
Percentage Used: 15%
Data Units Read: 256,215,308 [131 TB]
Data Units Written: 232,512,023 [119 TB]

My 8GB M1 has a 512GB SSD, but seems to be accumulating writes at a much faster rate, accumulating over 8 TB in about 5 weeks. At this rate it will hit around 300 TB in 3.5 years!

 

[leman]

My 8GB M1 has a 512GB SSD, but seems to be accumulating writes at a much faster rate, accumulating over 8 TB in about 5 weeks. At this rate it will hit around 300 TB in 3.5 years!

Which would still put your expected SSD lifespan based on the endurance alone close to 15 years... I don't think it's anything to worry about in practical terms.

 

[Reason077]

Which would still put your expected SSD lifespan based on the endurance alone close to 15 years... I don't think it's anything to worry about in practical terms.

Yes, if it lasts 15 years then I agree that's fine. But it's also been suggested that NAND decline is non-linear. ie: The "available spare" will decline faster and faster over time?

 

[bobcomer]

No, it's not. Here's an image from an M1 Mac mini teardown of the M1 with the heat spreader removed. Note the two distinct Hynix RAM chips to the right hand side of the silver M1 SoC:

Still on the same package and no handy socket to plug/unplug a RAM chip into.

 

[leman]

Still on the same package and no handy socket to plug/unplug a RAM chip into.

And that's why you need some serious soldering skills and possibly professional equipment. Here is a video of RAM replacement on a MacBook Air, I suppose the procedure is comparable...

 

[pshufd]

I suspect the folks who upgraded the RAM ran into the same problem as the folks who tried to get 32GB RAM into a 2015 MacBook Pro. The firmware doesn’t support that kind of RAM configuration, so while it’s physically possible, it’s pointless because the machine will never see it, and may not even boot.

This is why I prefer upgradeable systems.

It's why I am running Windows right now and that the 2020 iMac is a consideration as it provides user-installable RAM.

 

[leman]

This is why I prefer upgradeable systems.

We are quickly approaching the limits of upgradeable RAM though... if you want high bandwidth, low latency, energy efficient RAM, tightly integrated solutions become the only viable option.

 

[pshufd]

We are quickly approaching the limits of upgradeable RAM though... if you want high bandwidth, low latency, energy efficient RAM, tightly integrated solutions become the only viable option.

I'm running on a desktop and have more flexibility.

I had a look at the 2021 LG Gram and it's quite nice with decent battery life. RAM is soldered in which is a change from 2020 where you could change it yourself. They give you an extra NVMe slot, though, which is nice.

 

[Coheebuzz]

True, but how would you get it apart and soldered back together? The Bus for the RAM is part of the M1 package. Oh well, never mind, it's not something easy in any case, so there's no upgrade path for us users.

Expect these sort of services to become mainstream in response to demand, a spark plug can take more time to replace than desoldering and soldering RAM and yet we think nothing of it.

 

[leman]

Nice analysis. I was thinking the unified memory would save a lot of holding copies, such as buffers of vertex or texture memory that serve as staging areas for copying to graphics memory. Certainly for games it simplifies a lot of things.

I think this is less clear-cut. True, UMA systems don't need to keep a copy of GPU resources in system memory, but the default API is still very copy-oriented (unless you use zero-copy APIs like newBufferWithBytesNoCopy), so most software will have to pay some overhead here anyway. Also, there are some GPU-only resources that can take substantial amount of memory like framebuffer attachments or various scratch buffers. With a traditional dGPU, these don't have to take any system memory at all. Then again, Apple TBDR GPU can often get away with using transient on-chip storage for many of these purposes and won't need to allocate any system memory here at all.

In the end, I think that a game with Apple-specific optimizations might end up using less RAM overall due to combination of transient resources and texture compression (Apple Silicon seems to be the only desktop platform with robust ASTC support...)

 

[Coheebuzz]

Also not mentioned are the 4 efficiency cores which are reserved for background system processes so system response is always snappy irregardless of what it's munching, there's nothing stopping a hypothetical 1 million core system with a billion petabytes of RAM that isn't optimised around the user-experience to experience slowdowns because we can always find something to fully-occupy them with and leave nothing for kernel and user stuff.

 

[Reason077]

Also not mentioned are the 4 efficiency cores which are reserved for background system processes

They're not necessarily reserved for background or system processes. Any[*] thread/process can be scheduled to run on the efficiency cores, although of course macOS will tend to run the more demanding ones on the performance cores.

[*] except for Intel (Rosetta) processes, which can only run on performance cores

 

[Coheebuzz]

They're not necessarily reserved for background or system processes. Any[*] thread/process can be scheduled to run on the efficiency cores, although of course macOS will tend to run the more demanding ones on the performance cores.

[*] except for Intel (Rosetta) processes, which can only run on performance cores

Reserved was the wrong word then, i didn't mean to suggest that they were exclusive to system processes but they can act as a buffer for such.

 

[rui no onna]

They only replaced the 8GB with 16GB, and I wouldn’t be surprised if the SSD was taken out from another M1 Mac. It’s a nice way to show off your skills with soldering, but that’s about it.

Don't you just need $$$$ equipment for that (BGA/SMD rework station or something)?

The controller or "brain" of the SSD is part of the M1 chip. As far as I'm aware, for the actual storage portion, the M1 Macs just use standard NAND flash packages.

Regardless, swapping these is not something one can easily do at home.

 

[leman]

Don't you just need $$$$ equipment for that (BGA rework station or something)?

If you watch the video I posted, they do it using a special heat gun (no idea how expensive those are), a pair of tweezers and a simple soldering iron. That said, I thought that video was terrifying. So many things can go wrong… and you need extreme precision when aligning these chips…

 

[rui no onna]

If you watch the video I posted, they do it using a special heat gun (no idea how expensive those are), a pair of tweezers and a simple soldering iron. That said, I thought that video was terrifying. So many things can go wrong… and you need extreme precision when aligning these chips…

I believe the SMD Hot Air Rework Stations are relatively affordable. I know several users on MobileRead who have those and used them to replace failed NAND flash on Kindles (apparently a not uncommon problem).

As you say though, it's nowhere near as easy as replacing slotted parts.

 

[neinjohn]

This one is a pretty good example of a great SMD Hot Air Station ( https://www.unionrepair.com/hot-air-rework-station-quick-861dw.html ) that is not Chinese crap.

A good soldering station is between 200-400$ and a handy microcoscope anywhere from 250-500$. Of course you can spend way more than that for any of those items if going with the more renowned brands, which may or not, work better, and less on some workable stuff if you're Russia trained.

 

[Andropov]

The expensive thing about this replacements is not the hot air station itself, but the fact that you have to account for a lot of bricked boards from time to time, even if you know what you're doing. Replacing surface mounted components reliably is hard and requires skill. And every time you screw up you ruin a $1000 logic board

 

[Octopuss]

M1 is better with (16GB then 8 GB) RAM

 

[leons]

The downside is that NVM/NAND doesn't last forever: the number of writes it can support are limited, and eventually the SSD will die.

My 8GB 2017 MacBook Pro managed to accumulate 119 TB of writes on it's 256GB SSD in 3.5 years of use, which according to smartmontools, has used up at least 15% of it's lifespan/spare cells:

SMART/Health Information (NVMe Log 0x02)

Available Spare: 79%
Available Spare Threshold: 2%
Percentage Used: 15%
Data Units Read: 256,215,308 [131 TB]
Data Units Written: 232,512,023 [119 TB]

My 8GB M1 has a 512GB SSD, but seems to be accumulating writes at a much faster rate, accumulating over 8 TB in about 5 weeks. At this rate it will hit around 300 TB in 3.5 years!

But it is good for about 10X that.

 

[quarkysg]

Replacing surface mounted components reliably is hard and requires skill.

Yup. I once tried to replace an old router's 8 pin SOIC flash chip and completely decimated the board's traces. Granted, I was using a cheap soldering iron and some solder wicks. Using specialised equipments helps, but experience and skills counts for a lot in such operations.

 

[bluecoast]

Kudos to the OP for starting this thread.

I’m going to give a very unscientific contribution, I’m afraid!

My contention is that many Mac users on forums (such as this) are expecting the low end M1 machines to sustain ‘mid-level’ work loads which (presumably) the next tranche of Macs (32+ GB of RAM etc. will be intended for.

The current M1 systems are (presumably) the weakest Mx Macs that Apple will ever release.

They’re aimed at consumers and those with low end pro and prosumer work flows & as such, we shouldn’t expect more from them.

So to finish up, for me, the more interesting comparisons are going to be when the ‘mid-level’ pro workflow Mx systems are released & we can see how this architecture - and its performance - scales.

 

[pshufd]

Kudos to the OP for starting this thread.

I’m going to give a very unscientific contribution, I’m afraid!

My contention is that many Mac users on forums (such as this) are expecting the low end M1 machines to sustain ‘mid-level’ work loads which (presumably) the next tranche of Macs (32+ GB of RAM etc. will be intended for.

The current M1 systems are (presumably) the weakest Mx Macs that Apple will ever release.

They’re aimed at consumers and those with low end pro and prosumer work flows & as such, we shouldn’t expect more from them.

So to finish up, for me, the more interesting comparisons are going to be when the ‘mid-level’ pro workflow Mx systems are released & we can see how this architecture - and its performance - scales.

There's a 2019 8-core MacBook Pro 16 with 32 GB of RAM and 1 TB SSD for $1,800 locally. I was thinking that's $1,400 off retail for a machine that's 18 months old. It would meet my desire for RAM, CPU, screen size, and battery life. Or I could wait for the M1X that will likely have 32 GB of RAM and the ports I want and be lighter with more battery life and far more CPU compute (jury is still out on GPU) that would likely cost $2,400 to $3,000.

I'm keeping my eye open on 2019s for deals. I won't jump on one but they would be acceptable at the right price.