Apple Likely Planning to Skip M3 Mac Mini for M4 Refresh in Late 2024

[Motorola68000]

What an earth are you on about?

Apple’s M series chips use unified memory, it works differently to normal RAM you’d buy from a shop to install in a PC and due to Apple controlling all the hardware and software at this point, it’s massively optimised.

Now, if you’re buying a machine to run an application that needs 16gb of RAM, you should get a machine with 16gb of RAM but the everyday user doesn’t need that and won’t need that for years to come to do what they currently do on their machines.

You certainly don't need to explain unified memory to me, although a basic explanation is that technically the RAM is not different to normal RAM its how it is situated and its proximity and its interaction with GPU/CPU and where unified RAM is soldered onto the SoC, but where ironically I never raised that, so quite why you are inserting an answer to a question I never raised or using it as an argument seems rather strange as if anything it makes getting 8Gb even more tenuous when it can't be upgraded after purchase.

"If you are buying a machine to run an application that needs 16GB" well Apple now have backtracked and confirmed 8Gb is suitable for basic tasks, and where I doubt many people even check what software RAM requirements are let alone that these things change quite rapidly with evolving software and multitasking, which makes your argument very implausible.

If you don't consider any future proofing, then in my opinion its not the way to ensure longevity of your devices. How many upgrades to Mac OS do we get? Frequent?

How many upgrades of existing software used even in basic tasks, where they WILL evolve and require more RAM?

 

[DavidSchaub]

You seem quite angry about it. Apple is just a business, they've always been expensive and I don't expect that to ever change.

I feel that this is bothering people because “the expensiveness” seems to be non-linear.

To maintain Apple’s target average margins across the Max business, I think the margins on BTO options are way higher than the margins on the base SKUs.

Would it make people happier if the base SKUs cost a bit more and RAM and Storage cost a lot less? I don’t know

 

[chucker23n1]

"Unified Memory" just means the CPU and GPU cores share the same address space. Which means, for example, you can build a large amount of data in the CPU, then pass it on to the GPU in zero time, since it can just reuse your existing addresses. That's useful, and when you add Apple's relatively high-bandwidth memory controllers, you end up with quite good memory latency/throughput numbers.

"Unified Memory" does not, however, mean that the RAM chips are anything special. They're LPDDR5X; quite a few laptops use that now.

 

[chucker23n1]

I feel that this is bothering people because “the expensiveness” seems to be non-linear.

To maintain Apple’s target average margins across the Max business, I think the margins on BTO options are way higher than the margins on the base SKUs.

Would it make people happier if the base SKUs cost a bit more and RAM and Storage cost a lot less? I don’t know

It's OK to criticize Apple's base configs while also buying Apple products. The $200 BTO upgrade for 8 to 16, or $400 for 16 to 32, has a big margin, but isn't that out of line with the rest of the industry. But we're fast approaching the point where defaulting to 8 on the $1,099, $1,299, $1,599 portable Macs is too low.

 

[truthsteve]

Unified memory also means that RAM is also being consumed as video memory

and?

unified memory also means you have more available video memory than the average non-gaming laptop
same goes for AI use cases
same goes for audio acceleration
etc...

and you don't have to allocate textures on main memory and then copy to video memory, the GPU simply grabs from the unified memory.

I don't think you understand this at all if you think that was a counter-argument.

 

[DavidSchaub]

and?

unified memory also means you have more available video memory than the average non-gaming laptop
same goes for AI use cases
same goes for audio acceleration
etc...

Yes, and all those cases could be eating up RAM that the CPUs might have used itself otherwise... so less RAM in general.

 

[truthsteve]

Yes, and all those cases could be eating up RAM that the CPUs might have used itself otherwise... so less RAM in general.

except you have more memory for the GPU for example.

 

[truthsteve]

The RAM chips Apple sources are somewhere between $40 and $60, both for 8 GiB and 16 GiB. So just to put "much more complex and costly" in perspective, upgrading would cost them something like $20. Probably less, since they get better bulk discounts than I do.

all you're looking at is BOM which doesn't even cover other costs.

 

[Motorola68000]

"Unified Memory" just means the CPU and GPU cores share the same address space. Which means, for example, you can build a large amount of data in the CPU, then pass it on to the GPU in zero time, since it can just reuse your existing addresses. That's useful, and when you add Apple's relatively high-bandwidth memory controllers, you end up with quite good memory latency/throughput numbers.

"Unified Memory" does not, however, mean that the RAM chips are anything special. They're LPDDR5X; quite a few laptops use that now.

A better description than mine. It does show that some do not realise its not the RAM, which as you rightly suggest is nothing special. Thank you for a better description.

 

[chucker23n1]

all you're looking at is BOM which doesn't even cover other costs.

There are no other costs when it comes to BTO options. You're literally soldering thing A or thing B.

Do they use high BTO margins to recoup R&D? Sure. Are the parts "much more complex and costly to manufacture"? No.

 

[truthsteve]

There are no other costs when it comes to BTO options. You're literally soldering thing A or thing B.

that's a completely different argument from your "if chips cost XYZ to pc manufacturers, it should cost relatively the same for Apple to build it into UMA architecture" argument. and this new argument is not even getting to my point at all. 🤦‍♂️

Are the parts "much more complex and costly to manufacture"? No.

demonstrably false. for one, soldering the memory onto the SoC makes the heat problem exponentially more difficult, especially when coupling it with the GPU/neural engine.

plus the more memory put into the SoC, the lower the yields on the entire SoC chip. simple facts.

pricing UMA tiers isn't just about "how much can I buy samsung chips for".

 

[DavidSchaub]

plus the more memory put into the SoC, the lower the yields on the entire SoC chip. simple facts.

pricing UMA tiers isn't just about "how much can I buy samsung chips for".

I'm a little confused about this.

The memory is not in the SoC.

The M* die SoC is just a CPU, GPU, and other stuff that computers used to have on their motherboards; except, it still doesn't include the RAM or mass storage.

Instead of RAM chips being soldered onto a DIMM board, these dense (so a bit more expensive) RAM chips are being soldered onto a package board (like a little motherboard) that only the SoC and the RAM on.

Apple literally buys some of their RAM chips from Samsung, yes.

 

[chucker23n1]

that's a completely different argument from your "if chips cost XYZ to pc manufacturers, it should cost relatively the same for Apple to build it into UMA architecture" argument.

Which is true. Apple just sources LPDDR5X chips from SG Hynix. Others can do so as well, and several laptop manufacturers do just that.

and this new argument is not even getting to my point at all. 🤦‍♂️

Feel free to elaborate.

demonstrably false. for one, soldering the memory onto the SoC makes the heat problem exponentially more difficult,

They're on-package, not on-SoC. And besides, the company Apple buys the RAM chips from doesn't do the soldering, so why would that factor into the RAM cost?

especially when coupling it with the GPU/neural engine.

How does the amount of chips inside the SoC that use the RAM affect "the heat problem"?

plus the more memory put into the SoC, the lower the yields on the entire SoC chip. simple facts.

The RAM isn't inside the SoC.

pricing UMA tiers isn't just about "how much can I buy samsung chips for".

That's true, I suppose, since they buy from SG Hynix.

(Yes it is. That's all the RAM pricing is about. And, again, "unified memory" has zero relevance to this.)

 

[Serqetry]

Are you sure he said that ?

I can’t seem to recall.

Another casualty of the Gurman mind-wipe ray. It's his secret weapon that makes it possible for him to keep that perfect track record while continuing to be wrong all the time.

 

[truthsteve]

The memory is not in the SoC.

Gruber says the memory is literally in the SoC

The M1 really is an entire system on a chip. Everything is on the M1. The various processors, of course: the CPU cores, the GPU cores, the Neural Engine cores. But everything else is on the M1 too: the storage controller, the Secure Enclave, the memory controller, and, yes, the memory itself.

One More Thing: The M1 Macs

The bottom line is that for over a decade, iPhones and iPads have had Apple-designed chips the competition could not and still cannot match. Now the Mac does too.

daringfireball.net

and then from Apple:
"...chips for prcessor/IO/security/and memory...now with M1, these technologies are combined into a single SoC"

(7:20)

even if it wasn't, the SiP is still produced with the memory with all of the constraints and downsides (heat/complexity in manufacturing/lower yields).

 

[truthsteve]

Which is true.

actually no, it's not.

Feel free to elaborate.

already did later in the post.

They're on-package, not on-SoC.

One More Thing: The M1 Macs

The bottom line is that for over a decade, iPhones and iPads have had Apple-designed chips the competition could not and still cannot match. Now the Mac does too.

daringfireball.net

The M1 really is an entire system on a chip. Everything is on the M1. The various processors, of course: the CPU cores, the GPU cores, the Neural Engine cores. But everything else is on the M1 too: the storage controller, the Secure Enclave, the memory controller, and, yes, the memory itself.

and then from Apple:
"...chips for prcessor/IO/security/and memory...now with M1, these technologies are combined into a single SoC"

(7:20)

Even if they weren't, it still increases heat and lowers yields should the final production of the M1 contain a defective memory chip.

And besides, the company Apple buys the RAM chips from doesn't do the soldering, so why would that factor into the RAM cost?

again, you're doing BOM only.

adding ram into the SoC increases heat, increases complexity, lowers yields

How does the amount of chips inside the SoC that use the RAM affect "the heat problem"?

more chips = more energy consumption = more heat. this should be obvious to anyone.

The RAM isn't inside the SoC.

see above.

That's true, I suppose, since they buy from SG Hynix.

(Yes it is. That's all the RAM pricing is about. And, again, "unified memory" has zero relevance to this.)

Unified Memory includes additional custom memory controllers, lowers yields, increases complexity in manufacturing the SoC, therefore increases prices of the ram upgrades.

 

[DavidSchaub]

The memory is literally in the SoC

One More Thing: The M1 Macs

The bottom line is that for over a decade, iPhones and iPads have had Apple-designed chips the competition could not and still cannot match. Now the Mac does too.

daringfireball.net

This is now an "internet definition debate", which is always pointless to enter... but oh well...

This page has an image:

Edge AI Evangelist’s Thoughts Vol4: Why is the Apple M1 processor so fast? - HACARUS INC.

Hello everyone, I’m Haruyuki Tago, Edge Evangelist at HACARUS Tokyo R&D Center. In this series of articles, I will share some insights from my decades long experience in the semiconductor industry and comment on different AI industry related topics from…

hacarus.com

From there, I labeled this image:

 

[Boil]

The memory is literally in the SoC

One More Thing: The M1 Macs

The bottom line is that for over a decade, iPhones and iPads have had Apple-designed chips the competition could not and still cannot match. Now the Mac does too.

daringfireball.net

From provided link...

The DRAM for M1-based Macs is on the package (“on the substrate”, I believe, is the technical lingo).

The above is the M2 Ultra PACKAGE; note the actual SoC (consisting of two M2 Max chips) in the middle, with the eight LPDDR5 DRAM chips attached to the PACKAGE that carries both the SoC and the RAM...

THE RAM IS NOT, AND HAS NEVER BEEN, PRINTED WITH THE SOC...!

Take a breath and admit, at least to yourself, that you are wrong about this...

 

[truthsteve]

This is now an "internet definition debate", which is always pointless to enter... but oh well...

This page has an image:

Edge AI Evangelist’s Thoughts Vol4: Why is the Apple M1 processor so fast? - HACARUS INC.

Hello everyone, I’m Haruyuki Tago, Edge Evangelist at HACARUS Tokyo R&D Center. In this series of articles, I will share some insights from my decades long experience in the semiconductor industry and comment on different AI industry related topics from…

hacarus.com

From there, I labeled this image:
View attachment 2371176

apple literally said it combines the memory onto a single chip.

you were the one that decided to enter that debate, not me, despite the fact that the outcome would still be the same: producing the M1 would be more costly with UMA vs standard traditional PC memory architectures.

are you going to continue the definition debate or provide a valid argument as to how Apple can build UMA with zero overhead compared to traditional PC memory architectures?

 

[DavidSchaub]

"Unified Memory" is just a good implementation of "shared video memory with an integrated GPU".

Apple's implementation was just to get as close as they could to HBM (High Bandwidth Memory) using off the shelf high density LPDDR memory chips. Make everything close and tight and soldered together, and things run really nicely.

Apple's implementation is also even "better" because by forcing it to be iGPU only, they can build everything to copy video data around less, which is a huge savings.

 

[truthsteve]

From provided link...

The above is the M2 Ultra PACKAGE; note the actual SoC (consisting of two M2 Max chips) in the middle, with the eight LPDDR5 DRAM chips attached to the PACKAGE that carries both the SoC and the RAM...

THE RAM IS NOT, AND HAS NEVER BEEN, PRINTED WITH THE SOC...!

Take a breath and admit, at least to yourself, that you are wrong about this...

"...chips for prcessor/IO/security/and memory...now with M1, these technologies are combined into a single SoC"

APPLE THEMSELVES SAID THE MEMORY IS COMBINED INTO A SINGLE SOC

I'm just going by what the person who largely designed the M1 chip is saying. You're going to tell me he's wrong? 🤣

There's no way you're getting out of this. Admit you're wrong and be done with it.

 

[Boil]

No, I'm going to tell you that your interpretation of what is presented is wrong...

Also, if you are trying to make a point using a 49 minute YouTube video, at least have the courtesy/decency to use the "Copy video url from current time" option, as I have done above...

 

[truthsteve]

Make everything close and tight and soldered together,

which increases heat, increases complexity in production of the M1, and lowers yields. this makes the product more expensive, therefore you cannot say "because PC manufacturer can provide XYZ ram, Apple should be able to do the exact same thing for the exact same cost"

 

[truthsteve]

No, I'm going to tell you that your interpretation of what is presented is wrong...

Also, if you are trying to make a point using a 49 minute YouTube video, at least have the courtesy/decency to use the "Copy video url from current time" option, as I have done above...

lol you're completely ignoring the quote.

watch at 7:20. there's really nothing left to debate. I'm quoting word for word what Johnny said. they literally drew an outline around the DDR4 chips and animated into a box to what Johnny describes as "into a single SoC"

I can't get any clearer than this.

if you're accusing me of "misinterpreting" it, then you're misinterpreting what I said because I used SoC in the same EXACT manner as Johnny.

 

[Boil]

Dude, you are wrong...

Johny misspoke, he follows up with...

"M1 unifies it's high-bandwidth low-latency memory into a single pool within a custom package."

Just a simple Google search for a de-lidded Mn-series SoC package will show that the RAM, which is an item purchased from a vendor not named TSMC, is a separate discrete item ON the package...

How, pray tell, does one print RAM onto a wafer, when said RAM is purchased from a third-party...?

Two examples have been provided for you so far, but I guess we are Photoshop masters who take the time to create false images to gaslight you...?

And if you watch the animation closely, you will see the two DRAM outlines are kept separate from the SoC outline, with the marketing variant heat spreader image imposed over it all at the end of the animation...