Unified memory increases 16” base RAM

[aednichols]

I believe the first 16” MBP with Apple Silicon may increase the base RAM to 32 GB.

The reason is that ASi no longer distinguishes between main and video memory and uses a single unified address space.

With the 4 or 8 GB of Radeon memory going away, I think this is our memory doubling moment. A user who needs 16 GB of main memory and 8 GB video memory cannot use a 16 GB ASi Mac, and Apple will not (should not) make a product that is strictly worse than the one it replaces.

 

[Erehy Dobon]

Based on Apple's normal M.O., I think you are wrong.

Apple's base RAM configuration will be 16GB and they will charge a premium for the 32GB RAM upgrade.

One thing I am absolutely sure about is that Apple will not substantially erode their gross margins. AAPL shareholders would not find that satisfying.

I am an indirect AAPL shareholder and I would not be happy if AAPL indulged in your fantasy scenario.

 

[aednichols]

Don’t forget that they will now start pocketing the fat chip margins they used to pay to Intel.

They will also no longer be paying AMD for that 4-8 GB of memory, nor for any third-party GPU at all for that matter.

 

[Erehy Dobon]

My guess is that Apple's initial yields won't allow them to rake in the big margins on the new ASi Mac chips. They are using cutting edge process nodes.

Apple will gain better margins once they have more volume and can use binning to spread ASi Mac IC production to various product lines.

Right now, they have exactly zero shipping product lines.

 

[bill-p]

Yeah, "fat chip margin" is doubtful. Intel ate up R&D and manufacturing on those chips. Now Apple has to shoulder R&D and manufacturing, plus they can't sell these chips to another company so they can only "eat it up" if yield sucks. People will argue "but iPhone and iPad..." but you're all saying it as if scaling a mobile chip to desktop level performance is completely free, with no drawback whatsoever, and yields, etc... remain perfectly linear. I don't think so, but anyways, everyone is entitled to their own opinions.

As for the 16" MacBook Pro, I don't think Apple will increase base RAM to 32GB. It'll be 16GB just like before. When the GPU needs memory, it won't take a full 4GB or 8GB all at once. Memory allocation happens on an "as needed" basis. So it's not like your 16GB MacBook can only access 12GB or 8GB. That's not how it works.

 

[aednichols]

As for the 16" MacBook Pro, I don't think Apple will increase base RAM to 32GB. It'll be 16GB just like before. When the GPU needs memory, it won't take a full 4GB or 8GB all at once. Memory allocation happens on an "as needed" basis. So it's not like your 16GB MacBook can only access 12GB or 8GB. That's not how it works.

Indeed, but if I need 16 GB main and 8 GB video today, that’s 24 GB and there’s no way to partition 16 GB to achieve that.

 

[bill-p]

Indeed, but if I need 16 GB main and 8 GB video today, that’s 24 GB and there’s no way to partition 16 GB to achieve that.

And in what scenario will you need 16GB main and 8GB video simultaneously? (I'm honestly genuinely curious)

 

[Erehy Dobon]

At least in consumer computing VRAM is more expensive than regular RAM because of the former's typically higher performance characteristics.

It's cheaper to work off of one chunk of SRAM which is how computers with integrated GPUs work.

Like all consumer PC product design decisions, there's a consideration between performance and price. Some people on this website don't appear to get that.

 

[aednichols]

And in what scenario will you need 16GB main and 8GB video simultaneously? (I'm honestly genuinely curious)

2D and 3D graphics, video editing, machine learning, scientific computing. There is a reason the Mac Pro can be configured with up to 2 x 32 GB GPUs.

 

[aednichols]

At least in consumer computing VRAM is more expensive than regular RAM because of the former's typically higher performance characteristics.

It's cheaper to work off of one chunk of SRAM which is how computers with integrated GPUs work.

Like all consumer PC product design decisions, there's a consideration between performance and price. Some people on this website don't appear to get that.

The Apple Silicon architectural sessions from WWDC mentioned a high speed cache that may play an equivalent role. It's been a while but my interpretation at the time was that it would be an additional slower/larger tier than typical CPU cache.

 

[Erehy Dobon]

2D and 3D graphics, video editing, machine learning, scientific computing. There is a reason the Mac Pro can be configured with up to 2 x 32 GB GPUs.

Not every transistor performs the same function.

Some eggheads can't move beyond this.

There are various circuits that are optimized for various tasks, machine learning, ray tracing, 2D rasterization, etc.

One of the massive shortcomings of synthetic benchmarks is a predetermined mix of limited assessment of various capabilities of today's 3D chips typically in an environment that doesn't reflect a real world usage case.

That's why the Windows PC guys rely heavily on multiple real world gaming benchmarks like AAA titles Shadow of the Tomb Raider, Metro Exodus or Rainbow Six Siege, because THAT'S WHAT PEOPLE PLAY not a synthetic.

Hell, I use Unigine Heaven to generate good fan curves on my custom build PC, I don't use it for benchmarking.

 

[Puonti]

Indeed, but if I need 16 GB main and 8 GB video today, that’s 24 GB and there’s no way to partition 16 GB to achieve that.

I don't have any insights into what kind of RAM steps Apple will offer with the Apple Silicon Macs, but them not offering a 24 GB option does "feel" like an Apple move. I could see them offering 8, 16, 32, 64 and 128 GB options (not all of them for all of their Mac product lines, of course), because in their view those options would still cater to the needs of their customers - the 32 GB one in your example.

Apple is no stranger to shifting things around in ways that require customers to pay a bit more for the configuration that exceeds their needs, vs. offering a configuration that meets them exactly.

Edit: just in case it's not clear, this is an observation, not critique, of how Apple operates. Their choices have always worked out for me in the end, even if they've not always perfectly matched with what I've wanted at the time.

 

[aednichols]

them not offering a 24 GB option does "feel" like an Apple move

Indeed, implicit in my first post is the assumption they would never offer an "unnatural" amount of RAM like this.

 

[leman]

RAM is RAM, you get what you need. Unified memory simplifies things. Your considerations are purely hypothetical and I am not sure whether they apply to reality. A user that “needs” 8GB VRAM for productivity purposes will most likely need 32GB of RAM anyway since they are most likely working with large videos or things like that.

Besides, relationship between RAM and VRAM on Intel Macs is not necessarily linear. The system will often mirror GPU data in system RAM to speed up data transfers, so it’s not like you get X+Y usable RAM in the first place.

 

[Andropov]

Besides, relationship between RAM and VRAM on Intel Macs is not necessarily linear. The system will often mirror GPU data in system RAM to speed up data transfers, so it’s not like you get X+Y usable RAM in the first place.

This. It's not uncommon to see data copied from the RAM to the GPU VRAM, processed and then written back to RAM. See Metal "Managed Mode" buffers for example.

 

[leman]

This. It's not uncommon to see data copied from the RAM to the GPU VRAM, processed and then written back to RAM. See Metal "Managed Mode" buffers for example.

It's even worse for OpenGL, where driver is often forced to keep the copy of the data in the system RAM because it has no idea how you are going to use it.

 

[Kostask]

My guess is that Apple's initial yields won't allow them to rake in the big margins on the new ASi Mac chips. They are using cutting edge process nodes.

Apple will gain better margins once they have more volume and can use binning to spread ASi Mac IC production to various product lines.

Right now, they have exactly zero shipping product lines.

Apple's yields are exactly zero, because Apple doesn't fabricate silicon. TSMC, the company that makes silicon for Apple, has been running the 5nm process since at least the beginning of the year (for other Apple SoCs), and any process issues have been worked out. In the same way, TSMC controls the process, not Apple. From the latest reports, the TSMC yields at this point in the timeline are better than they were for the 7nm process (exact terms were "the defect rate was lower than the 7nm process was") at the same point in its timeline. TSMC has probably been producing AS Mac SoCs for at least a quarter, maybe even two; the basis of that is that Apple has 2.5M AS Macs (of all flavours that will be announced, and reveaked tomorrow) on order. To get that many AS Macs made by the end of the year, they need 2.5M SoCs. You don't build that many SoCs in a week, even with perfect yields and with production running full tilt, as the AS Mac SoCs are not the only parts that TSMC is making for Apple.

Apple is shipping a number of product lines using the same 5nm process as will be used in the AS Macs; iPad air 4s, iPhones (all 4 varieties), and the Apple Watch 6.

 

[thingstoponder]

I believe the first 16” MBP with Apple Silicon may increase the base RAM to 32 GB.

The reason is that ASi no longer distinguishes between main and video memory and uses a single unified address space.

With the 4 or 8 GB of Radeon memory going away, I think this is our memory doubling moment. A user who needs 16 GB of main memory and 8 GB video memory cannot use a 16 GB ASi Mac, and Apple will not (should not) make a product that is strictly worse than the one it replaces.

The base model isn’t 8GB and the system ram is split for the Intel GPU anyways which is used most of the time. It’s a moot point.

Maybe they won’t double it but they’ll do 20GB or so. They could depending on how the memory channels are configured.

Indeed, but if I need 16 GB main and 8 GB video today, that’s 24 GB and there’s no way to partition 16 GB to achieve that.

Then upgrade the RAM. 8GB is an upgrade as is anyways.