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

[enb141]

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.

Define better quality of Apple parts, specially SSD, which for example Samsung Pro 980/990 pretty much smokes all SSD macs have.

RAM, macs are not using DDR5 or GDDR6, so again define better quality.

Keyboards, if you get a mechanical keyboard, any of those will smoke any apple keyboard, same thing for mice.

Internal sound card of macs are trash, some motherboards have way way way way better sound card that any apple.

So saying mac have better quality parts is BS.

 

[sgtaylor5]

Define better quality of Apple parts, specially SSD, which for example Samsung Pro 980/990 pretty much smokes all SSD macs have.

RAM, macs are not using DDR5 or GDDR6, so again define better quality.

Keyboards, if you get a mechanical keyboard, any of those will smoke any apple keyboard, same thing for mice.

Internal sound card of macs are trash, some motherboards have way way way way better sound card that any apple.

So saying mac have better quality parts is BS.

When I was writing about better quality, I was not thinking better performance. I meant they’d last longer, that they wouldn’t fail earlier than normal electronics. Macs usually last longer than PC’s.

 

[leman]

Define better quality of Apple parts, specially SSD, which for example Samsung Pro 980/990 pretty much smokes all SSD macs have.

It would be curious to see how these parts compare in terms of write endurance and reliability. For example, Apple hardware has built-in protection against data loss on power failure and it offers string data synchronization guarantees. These are features usually reserved for very costly enterprise SSDs, as consumers (esp. gamers) are more tolerant to data corruption.

RAM, macs are not using DDR5 or GDDR6, so again define better quality.

They use LPDDR5, which is the most expensive and premium RAM on the market, not to mention that they have the RAM modules made to custom specifications to enable ultra-low energy consumption. You won’t be able to find those kinds of parts anywhere except on high end smartphones.

 

[leman]

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.

A 24GB GPU is not sufficient for really large GPU workloads. Nvidia has a GPU for these applications - it comes with 80GB RAM and is nominally slower than the RTX 4090 on paper (but doubles tensor performance). You won’t be able to buy this GPU in a normal computer store, and prices I’ve seen mentioned vary between $30000 and $50000.

 

[enb141]

A 24GB GPU is not sufficient for really large GPU workloads. Nvidia has a GPU for these applications - it comes with 80GB RAM and is nominally slower than the RTX 4090 on paper (but doubles tensor performance). You won’t be able to buy this GPU in a normal computer store, and prices I’ve seen mentioned vary between $30000 and $50000.

So according to you, a mac with 32 GB ram smokes a Geforce 4090?

 

[enb141]

It would be curious to see how these parts compare in terms of write endurance and reliability. For example, Apple hardware has built-in protection against data loss on power failure and it offers string data synchronization guarantees. These are features usually reserved for very costly enterprise SSDs, as consumers (esp. gamers) are more tolerant to data corruption.

For Desktops you can use a RAID of SSD, 4 disks Raid 10 will solve pretty much solve that issue.

They use LPDDR5, which is the most expensive and premium RAM on the market, not to mention that they have the RAM modules made to custom specifications to enable ultra-low energy consumption. You won’t be able to find those kinds of parts anywhere except on high end smartphones.

Some laptops have LPDDR5, is more expensive yes, but not $200 for upgrading from 8GB to 16 GB, and also is not better than DDR5, it just uses less power and dissipates less heat that's it.

 

[enb141]

When I was writing about better quality, I was not thinking better performance. I meant they’d last longer, that they wouldn’t fail earlier than normal electronics. Macs usually last longer than PC’s.

Macs are just better if you compare them with budget to normal priced Dells or Lenovos PCs, not to any high end PC, specially custom made ones.

Mac pricing is only decent if you only need to read your emails online, use small office documents (8GB/256GB), if you need a little bit more, then 16/1TB is the normal now days, the price is high, almost twice from the base model, after that, the price is just nuts, is better to build your own PC or look for a prebuilt desktop.

For laptops, their battery can last twice about a normal PC, so if you need a laptop that lasts lots of hours without charging it, then a mac is a good option.

If you are looking for the bests price / performance then if you need lots of RAM/storage, macs can get soooo expensive, so a PC is the way to go in 99% of the cases.

 

[senttoschool]

So according to you, a mac with 32 GB ram smokes a Geforce 4090?

Do you have a ChatGPT account?

Here's how much VRAM is required for each variant of GPT-3 according to GPT4:

1. GPT-3 125M (125 million parameters): Around 2.5 GB
2. GPT-3 355M (355 million parameters): Around 7 GB
3. GPT-3 760M (760 million parameters): Around 15 GB
4. GPT-3 1.3B (1.3 billion parameters): Around 26 GB
5. GPT-3 2.7B (2.7 billion parameters): Around 54 GB
6. GPT-3 6.7B (6.7 billion parameters): Around 134 GB
7. GPT-3 13B (13 billion parameters): Around 260 GB
8. GPT-3 175B (175 billion parameters): Around 350 GB

Keep in mind that this is VRAM - not your typical DDR4/DDR5 RAM connected to the CPU. So if you have a 24GB 4090, you can only run the 760 million parameter model. The production version of ChatGPT uses the 175b variant. This means OpenAI needs to stitch together 5 Nvidia A100 80GB together in order to inference. Each Nvidia A100 is about $15,000. You need 5 of them so the total is $75,000.

No one is saying that Apple Silicon can do inference as fast as Nvidia's professional line of GPUs. But, when AI software is more optimized for Apple Silicon, you could in theory, run a decently sophisticated LLM on consumer-grade Macs because of Apple's unified memory architecture. It doesn't cost an arm and a leg to get 120GB of VRAM on Macs. It only costs an arm.

Why VRAM? GPT-4 answers:

High bandwidth RAM is important for Large Language Model (LLM) inference due to the following reasons:

1. Parallelism: LLMs, such as GPT-3, rely heavily on parallelism for performing numerous mathematical operations like matrix multiplications, which are a core component of neural networks. High bandwidth RAM enables faster data transfer between the memory and processing units (such as GPUs or TPUs), improving the overall speed of these parallel computations.
2. Model size: LLMs have a large number of parameters that need to be stored in memory. High bandwidth RAM allows for quicker access to these parameters during inference, reducing the time taken to process each input.
3. Intermediate computations: During inference, LLMs generate intermediate values that must be stored in memory. Fast access to these intermediate values is crucial for maintaining the flow of computations and avoiding bottlenecks.
4. Batching: To optimize the performance of LLMs, multiple inputs are often processed simultaneously in a single batch. High bandwidth RAM is essential for managing the increased memory requirements and ensuring fast data transfer between the memory and processing units for each input in the batch.
5. Caching: LLMs often cache intermediate results to speed up computation. High bandwidth RAM allows for faster access to these cached values, improving the model's overall performance.

In summary, high bandwidth RAM is crucial for LLM inference because it improves the speed and efficiency of data transfer between memory and processing units. This allows for faster parallel computations, quicker access to model parameters, and more efficient handling of intermediate values, batching, and caching, all of which contribute to better overall performance.

 

[senttoschool]

For Desktops you can use a RAID of SSD, 4 disks Raid 10 will solve pretty much solve that issue.


Some laptops have LPDDR5, is more expensive yes, but not $200 for upgrading from 8GB to 16 GB, and also is not better than DDR5, it just uses less power and dissipates less heat that's it.

Macs are just better if you compare them with budget to normal priced Dells or Lenovos PCs, not to any high end PC, specially custom made ones.

Mac pricing is only decent if you only need to read your emails online, use small office documents (8GB/256GB), if you need a little bit more, then 16/1TB is the normal now days, the price is high, almost twice from the base model, after that, the price is just nuts, is better to build your own PC or look for a prebuilt desktop.

For laptops, their battery can last twice about a normal PC, so if you need a laptop that lasts lots of hours without charging it, then a mac is a good option.

If you are looking for the bests price / performance then if you need lots of RAM/storage, macs can get soooo expensive, so a PC is the way to go in 99% of the cases.

I used to build PCs when I was younger. My first one was an AMD XP 2500+ that I overclocked to XP 3200+. My second was a Core 2 Duo Conroe. My last one was an Intel 750 Lynnfield.

You're coming from the DIY x86 world. I get how you think that you can get better value by buying different parts and putting them together. I really do.

But for myself, Apple computers provide more value than building my own. I no longer have time or desire to tinker with DIY computers. Apple's quality and performance is generally top notch. I need macOS for work because of *nix. I need a laptop, not a desktop. I no longer play video games.

So you can say talk about how much more memory you can get if you build a DIY desktop as much as you want. It still doesn't matter to me.

I generally look at overall industry trends over forum anecdotes. In industry trends, Macs are outselling and outgrowing PCs. This tells me that other people are also beginning to find more value in Macs than PCs.

 

[enb141]

I used to build PCs when I was younger. My first one was an AMD XP 2500+ that I overclocked to XP 3200+. My second was a Core 2 Duo Conroe. My last one was an Intel 750 Lynnfield.

You're coming from the DIY x86 world. I get how you think that you can get better value by buying different parts and putting them together. I really do.

But for myself, Apple computers provide more value than building my own. I no longer have time or desire to tinker with DIY computers. Apple's quality and performance is generally top notch. I need macOS for work because of *nix. I need a laptop, not a desktop. I no longer play video games.

So you can say talk about how much more memory you can get if you build a DIY desktop as much as you want. It still doesn't matter to me.

I generally look at overall industry trends over forum anecdotes. In industry trends, Macs are outselling and outgrowing PCs. This tells me that other people are also beginning to find more value in Macs than PCs.

More value will depend on your needs, in my case 1 boot drive PCIe 5 gen (intel optane smokes any other SSD on the market, the PRO ones), 64 GB in ram, that amount of ram with a mac would cost me a fortune, secondary SSD, HDD for non fast files (music), a decent video card, 4070 which smokes pretty much what apple offers.

Video games, even casual ones are a no on Macs, the transition to apple chips, killed the last remnants of gaming on mac.

With my needs, if I go mac, I had to spend a fortune, so doing a little research, will save me a lots of money, besides, some parts are reusable, so I don't have to buy everything every time, for example monitor, power supply, case, those parts I can reuse for a long time, 8 years or more, my old power supply served me for 16 years, now is time to go, it didn't die but I know that can't work for many more years, so I replaced it.

 

[enb141]

Do you have a ChatGPT account?

Here's how much VRAM is required for each variant of GPT-3 according to GPT4:

Keep in mind that this is VRAM - not your typical DDR4/DDR5 RAM connected to the CPU. So if you have a 24GB 4090, you can only run the 760 million parameter model. The production version of ChatGPT uses the 175b variant. This means OpenAI needs to stitch together 5 Nvidia A100 80GB together in order to inference. Each Nvidia A100 is about $15,000. You need 5 of them so the total is $75,000.

No one is saying that Apple Silicon can do inference as fast as Nvidia's professional line of GPUs. But, when AI software is more optimized for Apple Silicon, you could in theory, run a decently sophisticated LLM on consumer-grade Macs because of Apple's unified memory architecture. It doesn't cost an arm and a leg to get 120GB of VRAM on Macs. It only costs an arm.

Why VRAM? GPT-4 answers:

Interesting post about GPT requirements, but still I don't think only VRAM would the only factor to take into account, I don't think VRAM will be only one part, but a good GPU is also needed, right now macs GPU is a joke in comparison to high end Nvidia cards, even gaming ones.

 

[senttoschool]

Interesting post about GPT requirements, but still I don't think only VRAM would the only factor to take into account, I don't think VRAM will be only one part, but a good GPU is also needed, right now macs GPU is a joke in comparison to high end Nvidia cards, even gaming ones.

It isn't a joke. It's very powerful but the software needs to be optimized for it. Nvidia is obviously the leader in AI chips. No one here is going argue against this.

What people here are saying, if you are willing to listen just for a tiny bit, is that Apple's unified memory architecture has a unique advantage in an LLM-dominated world because it's far cheaper to get more high bandwidth RAM on a Mac than from a GPU. VRAM is VRAM. You can't just download more.

 

[senttoschool]

More value will depend on your needs, in my case 1 boot drive PCIe 5 gen (intel optane smokes any other SSD on the market, the PRO ones), 64 GB in ram, that amount of ram with a mac would cost me a fortune, secondary SSD, HDD for non fast files (music), a decent video card, 4070 which smokes pretty much what apple offers.

Video games, even casual ones are a no on Macs, the transition to apple chips, killed the last remnants of gaming on mac.

With my needs, if I go mac, I had to spend a fortune, so doing a little research, will save me a lots of money, besides, some parts are reusable, so I don't have to buy everything every time, for example monitor, power supply, case, those parts I can reuse for a long time, 8 years or more, my old power supply served me for 16 years, now is time to go, it didn't die but I know that can't work for many more years, so I replaced it.

That's fine. I get it. I outlined my thoughts in the post you quoted.

Just an FYI, there have been a lot more DIY x86 diehards like yourself coming to this forum and telling us how PCs provide more value. Perhaps it's the insecurities. Perhaps it's the envy. Everything you’ve said has been repeated here many times before. Keep in mind that there are a lot of pretty high technical posters here. Many of them have built DIY PCs before like myself. We are not your average Mac user.

 

[leman]

For Desktops you can use a RAID of SSD, 4 disks Raid 10 will solve pretty much solve that issue.

You are joking, right?

So according to you, a mac with 32 GB ram smokes a Geforce 4090?

If the 4090 spends most of time trashing the VRAM and waiting until the data comes via the PCIe, yes, it’s likely that it’s gonna be trashed. I can probably outrun Usain Bolt two if he needs to make a toilet break every second.

Some laptops have LPDDR5, is more expensive yes, but not $200 for upgrading from 8GB to 16 GB, and also is not better than DDR5, it just uses less power and dissipates less heat that's it.

What does “better” mean? LPDDR is a more complex device compared to DDR, that’s all.

 

[enb141]

It isn't a joke. It's very powerful but the software needs to be optimized for it. Nvidia is obviously the leader in AI chips. No one here is going argue against this.

What people here are saying, if you are willing to listen just for a tiny bit, is that Apple's unified memory architecture has a unique advantage in an LLM-dominated world because it's far cheaper to get more high bandwidth RAM on a Mac than from a GPU. VRAM is VRAM. You can't just download more.

Unified RAM has its advantages, consoles have it, and is one of the reasons why they are cheap and still can provide good performance without a lot of horsepower.

But then, saying that mac unified RAM can smoke Nvidia's cards, that's another thing.

 

[enb141]

You are joking, right?

My point is, if you need something that doesn't corrupts your data, you have options.

If the 4090 spends most of time trashing the VRAM and waiting until the data comes via the PCIe, yes, it’s likely that it’s gonna be trashed. I can probably outrun Usain Bolt two if he needs to make a toilet break every second.

32 GB in RAM won't be enough for CPU and GPU.

What does “better” mean? LPDDR is a more complex device compared to DDR, that’s all.

Smaller, consumes less power, that's all, better performance, no, scalability, no.

 

[Rafterman]

So according to you, a mac with 32 GB ram smokes a Geforce 4090?

A 38 core MacBook Pro runs about 13 teraflops. A 4090 runs at abut 80 tflops. The Mac might do some things better, but "smokes?" Not exactly.

 

[enb141]

That's fine. I get it. I outlined my thoughts in the post you quoted.

Just an FYI, there have been a lot more DIY x86 diehards like yourself coming to this forum and telling us how PCs provide more value. Perhaps it's the insecurities. Perhaps it's the envy. Everything you’ve said has been repeated here many times before. Keep in mind that there are a lot of pretty high technical posters here. Many of them have built DIY PCs before like myself. We are not your average Mac user.

That's why I'm here reading from people that know about computers. Some forums have very knowledgeable people, for example hackintosh community exists because people know that they can build a computer to run OSX, also if you want to keep in the mac ecosystem, not so long ago people bought base computers and upgraded the HDD and memory by themselves because that apple charges you is just insane.

I've been DIY since 1998, and I have no plans to buy a prebuilt PC.

Hackintosh is pretty annoying to keep it working in the long run, I don't have the time anymore to do that and also, in a few years, there won't be Mac OS for x86.

With DIY you can get for example a core i3 with 16 TB Hard Drive, or you can upgrade your video card every 2 years instead of upgrading the whole system because you want / need a newer GPU.

My point is, if you are a DIY, you won't like to be robbed by insanely overpriced upgrades.

 

[Rafterman]

32 GB in RAM won't be enough for CPU and GPU.

For 90 percent, yes it would be. Maybe some day, Macs will catch up to PCs in gaming titles available, including super heavy ones. For productivity, most - like Photoshop and video editing - will work fine for most, with 32GB. You'd have to get into 8K video editing to need more.

 

[enb141]

A 38 core MacBook Pro runs about 13 teraflops. A 4090 runs at abut 80 tflops. The Mac might do some things better, but "smokes?" Not exactly.

"Power" is on one part of the equation, what to do with that power and how, is another question.

For example, anecdotal if you wish but, a year ago I upgraded my GeForce 1050ti with an AMD 6400 for my HTPC, the 6400 is way faster than my old 1050 ti, but sucks for my purpose, the 6400 doesn't has 10 bit color if I use an smart TV, which my old 1050 ti was able to do, also the 6400 doesn't has VRR, so the powerful 6400 ended up being a waste of money because doesn't do what the old one did better.

TLDR; Nvidia has way better GPU drivers/utilities than anyone in the GPU world.

 

[Rafterman]

That's why I'm here reading from people that know about computers. Some forums have very knowledgeable people, for example hackintosh community exists because people know that they can build a computer to run OSX, also if you want to keep in the mac ecosystem, not so long ago people bought base computers and upgraded the HDD and memory by themselves because that apple charges you is just insane.

I've been DIY since 1998, and I have no plans to buy a prebuilt PC.

Hackintosh is pretty annoying to keep it working in the long run, I don't have the time anymore to do that and also, in a few years, there won't be Mac OS for x86.

With DIY you can get for example a core i3 with 16 TB Hard Drive, or you can upgrade your video card every 2 years instead of upgrading the whole system because you want / need a newer GPU.

My point is, if you are a DIY, you won't like to be robbed by insanely overpriced upgrades.

I used to hackintosh long ago. I actually got a Dell laptop working with MacOS, including the built in broadband cellular modem. Which, ironically. meant I had a cellular equipped "MacBook" before Apple.

Used to DIY too. There was a certain joy with building your own rig. But I got lazy and now I just get properly spec'ed prebuilds.

 

[enb141]

For 90 percent, yes it would be. Maybe some day, Macs will catch up to PCs in gaming titles available, including super heavy ones. For productivity, most - like Photoshop and video editing - will work fine for most, with 32GB. You'd have to get into 8K video editing to need more.

There won't be mac gaming, if apple wanted mac gaming, the transition to Apple Silicon, killed the small gaming macs had, their only way to get games is the mobile ones (iPad games and iPhone games), you won't see Cyberpunk, Crysis, Far Cry on macs.

 

[enb141]

I used to hackintosh long ago. I actually got a Dell laptop working with MacOS, including the built in broadband cellular modem. Which, ironically. meant I had a cellular equipped "MacBook" before Apple.

Used to DIY too. There was a certain joy with building your own rig. But I got lazy and now I just get properly spec'ed prebuilds.

I still DIY because I reuse parts until they die or get too old, my old antec case has served me well for 16+ years and so my Power supply with another 16+ years, in those years I got at least 4 computers, so what a waste of money would be to get rid of them every 4 years when they could last way way more than those 4 years.

 

[senttoschool]

My point is, if you are a DIY, you won't like to be robbed by insanely overpriced upgrades.

How much money do you make per hour?

Now how much time do you spend researching parts, keeping up with new parts/releases?

How much time do you spend to put the PC together?

How much time do you spend to RMA bad parts?

 

[enb141]

How much money do you make per hour?

Now how much time do you spend researching parts, keeping up with new parts/releases?

How much time do you spend to put the PC together?

How much time do you spend to RMA bad parts?

- I don't have a fixed job but about ~ $4-$5
- I love to read about hardware, so to me research is like reading newspapers (probably hundreds or hours).
- One day to build the machine, a few more days to move my data from my old to my new build.
- I almost haven't need to RMA, I try to buy parts that doesn't fail, if they do, they probably I'll buy another part in the time that I don't have the item due to RMA.