Will 16GB of Memory Finally Come Standard Alongside the M3 Chip?

[thebart]

Since people seem to keep insisting that the price of something should essentially be the sum of the costs to create it, let's say Apple priced their memory upgrades at their cost. So adding going from 8GB to 32GB RAM would wind up meaning, say $50. Going from 256GB to 1TB would cost another $50 or so. Give or take, no exact numbers, but the point everyone wants to make here is that it would be much, much less than Apple currently charges.

Do you think Apple would release a series of products with only $100 between them? Would that be cost effective for the company or of any use to the consumer? Or if you see this as evidence that the low end products aren't worth producing at all, would they decide to only sell one model? What price would they likely sell it at?

when I say commensurate, I don't mean the same as cost. Of course I expect Apple to charge a premium, so no I don't expect to pay $50 for +24GB. Don't ridicule other people's points with absurdities. But $200 for +8GB is maybe, just maybe, a bit price gougy?

Other monopolistic / monopol-ish companies like Amazon and walmart leverage their position to squeeze suppliers and vendors. Apple is somewhat unique in that they use their position to squeeze everyone -- from Facebook to EA right down to end users

 

[Analog Kid]

Don't ridicule other people's points with absurdities.

Nobody's ridiculing anyone. I simply asked how you would propose pricing the systems. How would you set up their product line to meet your upgrade pricing goals and maintain the profit margins they need to sustain their Mac business?

 

[Analog Kid]

In my opinion, pricing is based on the perceived value to the customer, not the manufacturing price. It doesn't matter whether Apple pays $1, $50, or $100 for the RAM, as long as the customer is willing to pay their upgrade prices, it's just basic business. And you know, if it works, it works. People were grumbling a bit about Apple's high prices, but they were still happy to pay them.

Agreed with the exception that I'd say if the customer is willing to pay the system price.

Apple prices their systems where they do because they have different customers with different perceived value of the system. Parents buying their high schooler a machine for writing papers and companies equipping staff for web apps value the hardware less, and people working on bigger projects, writing code, content creation, etc value it more.

Turns out that those customers also have diverging needs for memory and storage and that's how Apple addresses those different customers. They sell different priced systems to those different customers. It just so happens that the specification that varies is memory so people think that's what's driving the price difference because people can't get past thinking of a Mac as a bag of parts.

However, the recent discussions suggest that the attitude is shifting a bit and customers find it more and more difficult to justify the prices. This likely means that Apple will have to adapt. Historically, they like their $200 upgrade points, which makes sense, but maybe the hardware can be distributed a bit differently across them. Maybe we'll see 12GB RAM jumps for 200$ instead of the current 8GB. Maybe there will be some other pricing structure. The only think I'm fairly sure about is that Apple will have to change things a bit, because their laptops having half the RAM and SSD of the closest competitors just doesn't look good.

And herein lies the problem. People keep insisting on looking at it as upgrade prices. It's not. They are different systems aimed at different markets.

What I expect will happen is that people will keep complaining, at some point the available memory parts will change in way that allows them to raise the memory at each price point, and the people complaining won't realize this is one of the two times their broken clock is right and believe that Apple changed because of customer pressure.

Another alternative is that Apple raises the price of their most discounted model and outfit it with more RAM. People will still complain about price but they'll somehow be satisfied that the "base model" specs increased and tie the price increase to a new process node, supply chain something, or inflation.

If Apple needs to find another way to differentiate machines, I feel it's worse for everyone. They could, for example, charge cost for RAM but modify the system in some other way, such as differentiating more granularly by processor performance.

At the moment all machines in a particular class have the same level of performance which is a great bargain for the low end market. YouTube seems to have people hypnotized into believing that memory restrictions are impacting their Chrome performance. People will find that scaling down the processor performance actually will.


What Apple has done with their pricing is rather remarkable. They've saved a ton of engineering effort to keep their products as low cost as possible because they really only need to design a fixed number of variants and customers of the lowest cost Mac get the same killer single core benchmarks that the highest end pros do.

People just need to stop thinking of the Mac as only the cost of its parts.

 

[leman]

A random thought:

If running AI on end-user devices becomes a thing, we are going to see a rapid increase in memory requirements, because bigger models are more useful. In that hypothetical future, we would soon be arguing whether 128 GB RAM should be the standard in the next-generation models.

This is also something where unified memory has a clear advantage. Apple has a clear path towards shipping consumer devices with very large amounts of RAM accessible to the GPU, if that becomes desirable. Achieving the same with traditional discrete GPUs is going to be more difficult.

The concept I've seen mentioned is in-memory processing. When bandwidth becomes a bottleneck, why not have the matrix engines inside the RAM itself?

P.S. And here is a relevant Apple patent — https://patentscope.wipo.int/search/en/detail.jsf?docId=US392326017&_cid=P22-LFQFPR-61308-2

 

[leman]

And herein lies the problem. People keep insisting on looking at it as upgrade prices. It's not. They are different systems aimed at different markets.

What I expect will happen is that people will keep complaining, at some point the available memory parts will change in way that allows them to raise the memory at each price point, and the people complaining won't realize this is one of the two times their broken clock is right and believe that Apple changed because of customer pressure.

Another alternative is that Apple raises the price of their most discounted model and outfit it with more RAM. People will still complain about price but they'll somehow be satisfied that the "base model" specs increased and tie the price increase to a new process node, supply chain something, or inflation.

If Apple needs to find another way to differentiate machines, I feel it's worse for everyone. They could, for example, charge cost for RAM but modify the system in some other way, such as differentiating more granularly by processor performance.

At the moment all machines in a particular class have the same level of performance which is a great bargain for the low end market. YouTube seems to have people hypnotized into believing that memory restrictions are impacting their Chrome performance. People will find that scaling down the processor performance actually will.


What Apple has done with their pricing is rather remarkable. They've saved a ton of engineering effort to keep their products as low cost as possible because they really only need to design a fixed number of variants and customers of the lowest cost Mac get the same killer single core benchmarks that the highest end pros do.

People just need to stop thinking of the Mac as only the cost of its parts.

I fully agree with you that looking at the pricing from the perspective of parts is naive. It's about device capability, with more capable devices getting progressively more expensive to cater to more demanding (and niche) users.

But value proposition and the available alternatives on the market also play an important role. Right now Macs are in a decent enough spot pricing wise IMO. Sure, they are expensive, and people might complain, but in the end of the day they are not any more expensive than competing, often inferior, products. But should, say, Microsoft release a Surface Laptop 6 with a 16GB/512GB config for $999, many users will ask themselves whether the more expensive M3 Air is a good value (even if the Air has better performance and battery). So it becomes the game of which decision can better optimise the revenue: keep the margins (and risk losing sales), lower the margins (and maintain or only slightly decrease the revenue), compensate by increasing the value proposition somewhere else, etc. After all, businesses have to adapt to circumstances. Failure to do so will quickly lead to ruin.

 

[theorist9]

I fully agree with you that looking at the pricing from the perspective of parts is naive. It's about device capability, with more capable devices getting progressively more expensive to cater to more demanding (and niche) users.

But value proposition and the available alternatives on the market also play an important role. Right now Macs are in a decent enough spot pricing wise IMO. Sure, they are expensive, and people might complain, but in the end of the day they are not any more expensive than competing, often inferior, products. But should, say, Microsoft release a Surface Laptop 6 with a 16GB/512GB config for $999, many users will ask themselves whether the more expensive M3 Air is a good value (even if the Air has better performance and battery). So it becomes the game of which decision can better optimise the revenue: keep the margins (and risk losing sales), lower the margins (and maintain or only slightly decrease the revenue), compensate by increasing the value proposition somewhere else, etc. After all, businesses have to adapt to circumstances. Failure to do so will quickly lead to ruin.

When I've priced out current Macs vs. PCs, comparing models with equal computing power, storage, and RAM (the Macs always win for efficiency, and thus quietness and battery life), I've found the Macs cost less when everything is spec'd with low storage and RAM. But the cost quickly switches in the PCs' favor when you go to higher storage and RAM.

That's a simple consequence of how much Apple charges to upgrade storage and RAM. And of course that's always been the Apple model--have the higher-spec'd devices subsidize the lower-spec'd ones. That helps them maintain high average margins, while at the same time enabling them to advertise lower starting prices. And it also helps them provide lower-cost entry-level devices for the all-important student market, since what people buy when they're young influences what ecosystem they'll adopt when they're older.

 

[Analog Kid]

Failure to do so will quickly lead to ruin.

Eventually lead to ruin. I think both Apple and Microsoft stand as examples of how hard it is for a really big and diversified company to die. Both have had eras of bad decision making and both managed to turn it around before flatlining. MS probably carries more inertia than Apple does because they're so ingrained in business purchasing, but Apple still has a really long way to fall before hitting earth.

 

[theorist9]

[Doing my best to sound like a Boston mechanic]
"There's your problem. You've got a real bad case of what we call 'Wolfram deposits' all over your memory.":

View attachment 2154695

You are a poster child for the difference between "I have too many tabs" and "I'm swapping because I do real work". Swap a tab and you're hitting that disk every few minutes, but you're probably hammering disk every few microseconds to analyze whatever dataset you're working with.

I'm curious though, while you're doing whatever it is you're doing, if you run 'vm_stat 10' in the terminal it will give you statistics about how much memory is getting paged every 10 seconds.

Sorry, forgot to reply to this. Here's what I found by doing an eyeball average over a few dozen vm_stat 10 ouputs:
640,000 pages/(10 s) x 4096 bytes/page x 1 MB/(10^6 bytes) = 260 MB/s.

 

[theorist9]

wow, what exactly is Wolfram Kernel doing to require such amounts of RAM?

In addition to the numeric calc I mentioned earlier, solving certain symbolic problems can also require a lot of RAM, particularly if there are a lot of variables.

 

[bombardier10]

Today 16 GB memory cost is about 50$...For buyers of course. But with the Apple logo already ten times more. Paying for a brand is nothing new in business .

 

[enb141]

Today 16 GB memory cost is about 50$...For buyers of course. But with the Apple logo already ten times more. Paying for a brand is nothing new in business .

Right now the price for upgrading from 8 GB to 16 GB is just $180, what a bargain 😏.

 

[dmccloud]

Today 16 GB memory cost is about 50$...For buyers of course. But with the Apple logo already ten times more. Paying for a brand is nothing new in business .

For DDR4, yes. But where are you going to find DDR5 (let alone LPDDR5X like Apple is using) for anything even close to $50 for 16GB?

 

[mi7chy]

Right now the price for upgrading from 8 GB to 16 GB is just $180, what a bargain 😏.

You can get 64GB (2x32GB) DDR5 SODIMM for $189.98.

https://www.amazon.com/Crucial-4800MHz-Laptop-Memory-CT16G48C40S5/dp/B09RVNMGFH/

For desktop, 32GB was free with CPU purchase.

 

[JouniS]

For DDR4, yes. But where are you going to find DDR5 (let alone LPDDR5X like Apple is using) for anything even close to $50 for 16GB?

A year ago, DDR5 was still expensive, because it was a fancy state of the art product with limited availability in the market. Today it's the standard baseline for memory, and its prices have fallen to where DDR4 prices used to be.

Faster memory is fundamentally no more expensive than slower memory, just like the next generation of CPUs is fundamentally no more expensive than the previous generation. Once the manufacturing capacity is high enough that there is no acute shortage, prices will fall to the old level.

 

[enb141]

You can get 64GB (2x32GB) DDR5 SODIMM for $189.98.

https://www.amazon.com/Crucial-4800MHz-Laptop-Memory-CT16G48C40S5/dp/B09RVNMGFH/

For desktop, 32GB was free with CPU purchase.

I was talking about the bargain prices that apple charges people for upgrading from 8 GB to 16 GB, which is just $180.00

 

[dmccloud]

A year ago, DDR5 was still expensive, because it was a fancy state of the art product with limited availability in the market. Today it's the standard baseline for memory, and its prices have fallen to where DDR4 prices used to be.

Faster memory is fundamentally no more expensive than slower memory, just like the next generation of CPUs is fundamentally no more expensive than the previous generation. Once the manufacturing capacity is high enough that there is no acute shortage, prices will fall to the old level.

Just a cursory glance on Newegg shows a 32GB DDR4 kit costs between $30-$80 less than a 32GB DDR5 kit from the same manufacturer. Factor in that on the AMD side, upgrading from a 5000 series Ryzen to the 7000 series requires both new RAM and motherboard in addition to the processor itself, which means that the associated costs are higher across the board. There is a reason Intel has made the 13th gen i series CPUs work with either DDR4 or DDR5, and it's cost considerations.

 

[JouniS]

Just a cursory glance on Newegg shows a 32GB DDR4 kit costs between $30-$80 less than a 32GB DDR5 kit from the same manufacturer. Factor in that on the AMD side, upgrading from a 5000 series Ryzen to the 7000 series requires both new RAM and motherboard in addition to the processor itself, which means that the associated costs are higher across the board. There is a reason Intel has made the 13th gen i series CPUs work with either DDR4 or DDR5, and it's cost considerations.

DDR5 is cheaper today than DDR4 was a couple of years ago. Of course DDR4 is even cheaper today, because the demand is lower and manufacturers and resellers want to get some money for the stocks they have.

The real point was that memory is memory. Once you have the production pipeline up and running, there is no substantial difference between making current-generation memory and previous-generation memory. And because there is at least some competition in the market, it's not possible to maintain artificial price premiums for faster memory very long. You can maintain some price differentiation by binning the chips and charging more for the ones that pass stricter quality control, but that's it. DDR5 is the same product as DDR4 was, and it will be priced in a similar way as DDR4 was priced in the same point of the product cycle. And while the prices fluctuate a lot over a few years, the price per gigabyte for current-generation memory goes down in the long term.

 

[senttoschool]

A random thought:

If running AI on end-user devices becomes a thing, we are going to see a rapid increase in memory requirements, because bigger models are more useful. In that hypothetical future, we would soon be arguing whether 128 GB RAM should be the standard in the next-generation models.

This is also something where unified memory has a clear advantage. Apple has a clear path towards shipping consumer devices with very large amounts of RAM accessible to the GPU, if that becomes desirable. Achieving the same with traditional discrete GPUs is going to be more difficult.

Absolutely. There is finally a demand for more memory. Most normal users and even some pro users have been fine with 8-16GB for the last 10 years. LLMs are the first mass consumer-level technology I know of that will demand much more than that.

My next Mac will likely have 64GB.

 

[enb141]

Absolutely. There is finally a demand for more memory. Most normal users and even some pro users have been fine with 8-16GB. LLMs are the first mass consumer-level technology I know of that will demand much more than that.

My next Mac will likely have 64GB.

A Mac with 64 GB in RAM, with that money you could get a PC with 256 GB or two PCs with 128 each.

 

[senttoschool]

A Mac with 64 GB in RAM, with that money you could get a PC with 256 GB or two PCs with 128 each.

Who cares? You need VRAM for LLMs.

Look up how much a GPU with 64GB of VRAM costs.

Apple's UMA now has a distinct advantage when it comes to LLMs.

 

[mi7chy]

64GB or even 128GB isn't going to cut it. GPT-3 175B requires 600GB RAM and 250GB VRAM. Only x64 workstations can accommodate those requirements or cloud compute rental.

 

[enb141]

Who cares? You need VRAM for LLMs.

Look up how much a GPU with 64GB of VRAM costs.

Apple's UMA now has a distinct advantage when it comes to LLMs.

Your beloved mac uses shared RAM for both GPU and CPU, so a GeForce 4090 with 24 GB in RAM + a PC with 32 GB RAM is cheaper than getting a mac mini / studio with 64 GB in RAM.

 

[senttoschool]

64GB or even 128GB isn't going to cut it. GPT-3 175B requires 600GB RAM and 250GB VRAM. Only x64 workstations can accommodate those requirements or cloud compute rental.

First, do you have a reliable source for those numbers?

Second, no, not only a x64. There are plenty of ARM server CPUs that can easily go beyond that CPU ram number.

Third, People have gotten lesser models to work on 16GB of unified memory such as Meta's LLama 30B model. The larger the VRAM, the bigger the model you can run. You can run a model that is fairly close to GPT3 on a 32GB or 64GB AS Mac such as LLama 64B model.

Fourth, no one is going to do LLM inference on a CPU.

 

[senttoschool]

Your beloved mac uses shared RAM for both GPU and CPU, so a GeForce 4090 with 24 GB in RAM + a PC with 32 GB RAM is cheaper than getting a mac mini / studio with 64 GB in RAM.

It doesn’t work like that. You need VRAM. System RAM matters very little. You can’t combine the two. You need to load the model on high bandwidth RAM, which is what VRAM is, and then use a GPU to do inference. CPUs are way too slow to do any useful inference.

Take for example, in order to run LLama's 64B model on Nvidia GPUs, you need 3x RTX 3090s with 24GB each for a total of 72GB of VRAM.

 

[sgtaylor5]

With that money I can build a PC with 64/1TB SSD + 8TB HD, Core i9 and probably still enough money left for a 4090

And you still wouldn’t have macOS on that computer. That’s the major differentiator between a Mac and a Windows computer. Oh, and the better quality of Apple parts.