Apple is falling behind, M4 Max nowhere close to RTX 5090.

[Unregistered 4U]

Besides all of that, the only real problem I have with Apple M strategy is to limit power draw on their desktop computers. That doesn't make any sense. Unless scalability is not that great or there is poor stability or perhaps it requires a lot of cooling. We will never know.

It’s just not required. It’s not like anyone is going to ship a faster macOS system. Chasing the teeny percent of users that would spend the money on such a desktop system is for AMD and Intel, now. Apple’s more than happy to make the profitable chunk of change they’re making with their laptops.

Apple’s rapidly approaching a point where very few will ever need anything like a Mac Studio or Mac Pro. They’re not putting a lot of emphasis on desktops as they don’t see a future there. I mean, just the fact that a Mac laptop is being compared against a power guzzling desktop in order to show a performance difference is a testament to the fact that Apple’s on the right track.

 

[theorist9]

I use Nvidia workstation and M4 max. For anything that fits with in the paltry Ram of Nvidia it smokes m4 max. But more than half of my use cases need at least 100 GB of VRAM/ unified memory. If I am training or running full analysis of my datasets, I use cloud. But my M1 Max is go to for all dev/testing, and work out all kinks before I use expensive cloud infrastructure.

It’s quite simple, anything less than 24-32 GB RAM, Nvidia is still the best choice. But anything more than 32GB, Nvidia 50XX is useless.

It will be interesting to see performance comparisions of these two machines:

$3,000 128GB VRAM/1 TB SSD NVIDIA Grace/Blackwell DGX Spark (formerly Project DIGITS), 273 GB/s mem. bandwidth

vs.

$3,700 40-GPU-Core 128 GB URAM/1TB SSD M4 Max Studio, 546 GB/s mem. bandwidth.

Interestingly, the DGX Spark can be 2x bridged to give more effective VRAM. But if you could do 4x DGX Spark (512 GB VRAM, $12k + connectors), you could compare that to a $10k 512 GB URAM M3 Ultra.

And I've read the Ultra itself can bridged, which would give 1 TB VRAM.

 

[Unregistered 4U]

I don't think they're "falling" behind, Apple Silicon has always been behind in terms of GPU.

Yeah, and it literally doesn’t matter as they’ll still get their 20 million plus in sales this year. And, it’s certain that if they were ahead in GPU… they’d still only sell 20 million plus. There’s nothing about a more performant GPU than they’re already shipping that would significantly drive sales of more Macs.

 

[Unregistered 4U]

Shouldn’t AI on Apple Silicon be using the Neural Engine rather then the GPU in the SoC.

Apple were first to utilize dedicated Neural Engines for consumer devices, I wouldn’t put it past them to create yet another new efficient dedicated block and it won’t be positioned ONLY for the high end. The entire processor line will get a boost with the desktop ones performing some decent multiple over the performance of the MacBook Air.

 

[diamond.g]

It’s to Steam’s benefit to attract more customers, same as with Apple. Apple’s currently selling 20+ million Apple Silicon Macs a year without providing incentives to Steam and, they’ll likely sell 20+ million Macs THIS year without providing incentives to Steam. If Steam’s looking at 20+ million users and going “Nah, we’d rather not try to win those customers” then that’s on them. Games will never drive Mac sales in a huge way as anyone not in the market and looking to play a given game has a wide range of options, almost all less expensive than a Mac, to play that game.

Wouldn’t it be on us macOS users to show up as more than 5% (ideally 10%) in the hwsurvey first?

 

[Unregistered 4U]

Do Mac users even have the *option* of GPUs that powerful? Efficiency is great but if we can't even reach that power, that's kind of an issue. External GPUs used to seemingly be a way around that but I don't think they're supported anymore

Nope, if any user doesn’t need macOS and needs peak GPU performance, they shouldn’t even be casting a shadow in Apple’s direction. Apple’s not in that space, hasn’t wanted to be in that space for years and likely doesn’t plan to change the calculus that says that, at any given time, I will ALWAYS be able to custom build a PC that, according to a LOONG list of PC benchmarks, will beat the most performant Mac being sold at that time.

 

[Unregistered 4U]

Wouldn’t it be on us macOS users to show up as more than 5% (ideally 10%) in the hwsurvey first?

In that equation, the macOS users are the ones with the dollars that Steam may or may not want. If Steam wants those dollars, they’ll do what’s necessary to increase those numbers to however high they want them to be. If Steam does NOT want those dollars, they’re free to turn them away.

Steam is well aware that people are downloading games for the PC then sweating through the details to get them running on their Macs. In THAT equation, Steam is doing nothing, Apple is incentivizing nothing, YET Steam gets Mac users. It’s likely an equation they’re happy with, keep doing the same thing and let the user figure it out.

 

[diamond.g]

In that equation, the macOS users are the ones with the dollars that Steam may or may not want. If Steam wants those dollars, they’ll do what’s necessary to increase those numbers to however high they want them to be. If Steam does NOT want those dollars, they’re free to turn them away.

Steam is well aware that people are downloading games for the PC then sweating through the details to get them running on their Macs. In THAT equation, Steam is doing nothing, Apple is incentivizing nothing, YET Steam gets Mac users. It’s likely an equation they’re happy with, keep doing the same thing and let the user figure it out.

I wonder what the ROI could be for Valve in buying Codeweavers and integrating their GPTK work into Steam natively.

 

[Unregistered 4U]

I wonder what the ROI could be for Valve in buying Codeweavers and integrating their GPTK work into Steam natively.

I’m guessing negligible in the short term and close to zero in the long term, especially considering that they’d have to keep it updated over time. As long as the claim to fame that Apple has over any other system being served games by Steam is “most expensive method”, the number of users they’re likely to draw in is not going to be worth their time. Possibly.

It’s true everywhere else that Apple users, even though smaller in number, drive larger amounts spending than their numbers would show. The fact that it’s likely NOT true for Steam would seem to indicate they are leaving an awful lot of dollars on the table. With more Apple users, they might not have to have as many sales!

 

[TechnoMonk]

It will be interesting to see performance comparisions of these two machines:

$3,000 128GB VRAM/1 TB SSD NVIDIA Grace/Blackwell DGX Spark (formerly Project DIGITS), 273 GB/s mem. bandwidth

vs.

$3,700 40-GPU-Core 128 GB URAM/1TB SSD M4 Max Studio, 410 GB/s mem. bandwidth.

Interestingly, the DGX Spark can be 2x bridged to give more effective VRAM. But if you could do 4x DGX Spark (512 GB VRAM, $12k + connectors), you could compare that to a $10k 512 GB URAM M3 Ultra.

And I've read the Ultra itself can bridged, which would give 1 TB VRAM.

M3 ultra has much higher memory B/W. I am still on 4090, probably would need to upgrade my workstation in the near future. It would be interesting to see how well Digits performs. And yes you could stack multiple ultras to get to 2 TB of vram but mem bandwidth on thunderbolt will be limiting factor.

 

[splifingate]

It will be interesting to see performance comparisions of these two machines:

$3,000 128GB VRAM/1 TB SSD NVIDIA Grace/Blackwell DGX Spark (formerly Project DIGITS), 273 GB/s mem. bandwidth

vs.

$3,700 40-GPU-Core 128 GB URAM/1TB SSD M4 Max Studio, 410 GB/s mem. bandwidth.

Interestingly, the DGX Spark can be 2x bridged to give more effective VRAM. But if you could do 4x DGX Spark (512 GB VRAM, $12k + connectors), you could compare that to a $10k 512 GB URAM M3 Ultra.

And I've read the Ultra itself can bridged, which would give 1 TB VRAM.

I'm interested in the comparisons, also.

I have the Spark Duo reserved (whatever that may eventually come to mean), and the kit was touted as ~USD8K . . . that would make a quad ~USD16K (before whatever tariff mayhem will ensue).

Things are going to get really interesting in the next few months....

 

[ZipZilla]

Gaming has sucked on Macs for the almost 30 years I have used one. I don't buy them for that. I buy it for content creation and general daily use.

 

[throAU]

The only drawback I can see (and I might be wrong) is increased power consumption.

Manufacturing complexity I guess? At some point there will be a yield/cost crossover vs. making bigger single dies though presumably.

 

[throAU]

Maybe it's indicative of several things, but using an MLX LLM on M4 and M4 Pro/Max returns wildly different performance, even though both have the same NPE. They have much different memory bandwidth, CPU, and GPU prowess. MLX should offload a portion go the NPE.

This is because the LLM is running on the GPU for the most part and the Pro/Max have much bigger/more powerful GPUs. Check the GPU utilisation on a machine running an LLM query.

 

[throAU]

As it is now, Macs are still awesome basic laptop-workstations, but that's that. Let's hope for a paradigm shift announced at WWDC for the fabled "hidra" chip in a real workstation/server. Mostly since it would be "fun", since I do not think we will get something with good price/perf ratio. After all, a h100 @30k is at least 10 times faster than a m3 ultra that costs about 10k

You have to ask though - who's going to use Macs for training? It's about as far away from an end user shiny device task as you can get. That wants bulk, cheap, headless hardware to run in massive clusters of thousands of machines rather than shiny friendly to use end user machines.

Unless you have your own - If you need to train fast, rent somebody's cluster, basically.

Apple make great hardware, but cost effective ML training is just NOT their wheelhouse - they don't have the hardware, they don't have the software and they don't have the customer base to justify building it.

And thats ok, there are others who build bespoke hardware ideal for that!

 

[throAU]

I don't think they're "falling" behind, Apple Silicon has always been behind in terms of GPU.

Depends how you measure.

If you measure in terms of power efficiency (Processing per watt) they're WAY in front.

 

[star-affinity]

Apple make great hardware, but cost effective ML training is just NOT their wheelhouse - they don't have the hardware, they don't have the software and they don't have the customer base to justify building it.

And thats ok, there are others who build bespoke hardware ideal for that!

Isn't ML training on the M3 Ultra good since it can have 512 GB Unified Memory which means the GPU has access to more RAM than any other GPU out there? I don't know.

 

[leman]

Manufacturing complexity I guess? At some point there will be a yield/cost crossover vs. making bigger single dies though presumably.

Oh yes, that was implied. It would be very interesting to learn more about the economics of these things — especially with newer processes costing as much as they do. What is cheaper: making a 600mm2 die on N3 or paying more for 3D packaging but making a 300mm2 N3 die + 300mm2 N5 die?

You have to ask though - who's going to use Macs for training? It's about as far away from an end user shiny device task as you can get. That wants bulk, cheap, headless hardware to run in massive clusters of thousands of machines rather than shiny friendly to use end user machines.

ML is a big field. I don't think anyone is seriously considering training a commercial-level LLM on a Mac, but for training smaller specialized models or fine-tuning, a local machine with a large RAM pool and good ML performance could work very well.

 

[picpicmac]

Not if the disaggregated dies cost less overall. E.g. if the new packaging allows them to use 30% smaller N3 dies

Which is the question.

When I look at images of the M1 and M2 variants - M4 I can't find online - it's pretty clear that the CPU and GPU cores take up only about half the surface (higher % on the Max, lower % on the base M).

Yet, TSMC and others have stated that their SoIC will be more expensive than the simpler processes.

The number of marketable product Apple gets per wafer has to be a larger increase than the cost per item or there's no reason for Apple to want to go down this path. Otherwise they'd just go with the regular 2nm process that will be available later this year (and if chosen would mean new Apple products would appear with such in early 2026.)

My conclusion is that indeed Apple has figured out that the per SoIC cost is about the same as their current procurements.

Yet, "about the same" is not exactly the same (or less.) And that could be why the base M5 will be made on the simpler process, as the mass market devices (e.g., MacBook Air) can't bear additional parts costs.

 

[TechnoMonk]

You have to ask though - who's going to use Macs for training? It's about as far away from an end user shiny device task as you can get. That wants bulk, cheap, headless hardware to run in massive clusters of thousands of machines rather than shiny friendly to use end user machines.

Unless you have your own - If you need to train fast, rent somebody's cluster, basically.

Apple make great hardware, but cost effective ML training is just NOT their wheelhouse - they don't have the hardware, they don't have the software and they don't have the customer base to justify building it.

And thats ok, there are others who build bespoke hardware ideal for that!

who is training ML on a 4090 or 5090? unless it’s for fine tuning or for dev/test before moving it to cloud. As some one who uses both Nvidia workstation and a Mac, they compliment each other. Large unified memory is helpful to test and tweak before blowing up money on High end Nvidia GPUs in cloud. Nvidia could potentially increase memory on future 6090, but that would eat in to their data center revenue.

 

[shadowboi]

Nowadays when I buy a Macbook I know what to expect, i.e. pristine build quality, well-functioning OS, compatibility with most of my apps, good security and nice display quality. When shopping for a Windows laptop you play a roulette – you either get this or that, not both. Also Windows… it is a big, big mess. And I am not some nolifer to waste my time polishing Linux and installing homebrew drivers for it.

Also as for the battery life – these are synthetic tests, before M1 MacBook every Windows laptop could not last more than 7 hours. In real use scenarios this is not enough and I would not trust synthetic tests since Windows can suck up battery out of the blue moon, for example every time it wants to install an update you don’t want or when it uses some process

 

[novagamer]

It will be interesting to see performance comparisions of these two machines:

$3,000 128GB VRAM/1 TB SSD NVIDIA Grace/Blackwell DGX Spark (formerly Project DIGITS), 273 GB/s mem. bandwidth

vs.

$3,700 40-GPU-Core 128 GB URAM/1TB SSD M4 Max Studio, 546 GB/s mem. bandwidth.

Interestingly, the DGX Spark can be 2x bridged to give more effective VRAM. But if you could do 4x DGX Spark (512 GB VRAM, $12k + connectors), you could compare that to a $10k 512 GB URAM M3 Ultra.

And I've read the Ultra itself can bridged, which would give 1 TB VRAM.

I think you can link 4 Mac Studios together, albeit only with TB5 bandwidth. I’s the only solution for fine tuning full models under probably $500k right now and works reasonably well for MoE if you only have a couple people doing the work.

Once Apple addresses their slow matrix GPU performance things are really going to get interesting. I’m going to be devastated if Mac Pro gets m3 because it means 2 years without those massive gains.

nvidia has a huge lead with their interconnect technologies, Apple really should put something special in the Mac Pro if they can swing it, and bring it to the next Studio.

 

[Unregistered 4U]

I think you can link 4 Mac Studios together, albeit only with TB5 bandwidth. I’s the only solution for fine tuning full models under probably $500k right now and works reasonably well for MoE if you only have a couple people doing the work.

Once Apple addresses their slow matrix GPU performance things are really going to get interesting. I’m going to be devastated if Mac Pro gets m3 because it means 2 years without those massive gains.

nvidia has a huge lead with their interconnect technologies, Apple really should put something special in the Mac Pro if they can swing it, and bring it to the next Studio.

There’s economies of scale in it for Nvidia. Apple’s Mac Pro costs what it costs partially because it leans on the R&D efforts from the millions and millions of base processors that sell in iPads and laptops. They stand on those shoulders, and put a few more of those cores on the die. A completely separate parallel effort to make something that doesn’t build on that tech? The prices can only go up from here.

People are still “trained” to think that the capability of high end solutions shouldn’t be available across the entire line. The fact that your average user, with their single threaded tasks, wouldn’t perceive a significant difference, day-to-day, between a Mac Pro and a MacBook Air (due to the single threaded performance being so similar) still seems foreign and wrong to them. Apple has done what AMD/Intel will never be able to do and instead of marveling at that fact, some are pointing at Intel saying, “Why can’t it be more like that?” It could, Apple could just do what AMD/Intel does and ship solutions with FAR more disabled cores than they do today.

 

[innerproduct]

What about mac pro with MPX hidras? Where each hidra is something like a m4 ultra+ ? Energy efficient, silent, connected via something the 400 Gb range instead of just thundebolt. Will it beat nvidia? Probably not but it will be something that could reduce costs in the end.
Oh, and btw, yes AI/ML is a vast field as mentioned. For a small company in CV we get very far on 4090s and similar and we mostly added a single h100 for the memory increase when working on large images. 4090s i än gamer rigs are silent. The h100 needs its own room. A mac studio would be lovely for experimenting and adding in as ”worker” if it just would fully support its hardware when used with torch or mlx.
Tbh, most smaller companies i worked with use gamer cards since long time. Both for AI and viz. Quadros and radeon pros seldom made sense even in visualization if you didn’t have some very nische need like stereo rendering on a cluster with sync needs.

 

[theorist9]

And yes you could stack multiple ultras to get to 2 TB of vram but mem bandwidth on thunderbolt will be limiting factor.

I think you can link 4 Mac Studios together, albeit only with TB5 bandwidth.

That limitation could provide a compelling use case for the next MP, even if it's only equipped with the same Ultra die as the contemporaneous Studio, because of its PCIe slots:

TB5's peak data transfer rate (after overhead) is 80 Gb/s = 10 GB/s (bidirectionally).

By comparison, using an x16 networking card on a MP woud give you these bidirection peak data transfer rates (after accounting for overhead from 242B/256B encoding):

x16 PCIe 5.0: 60.5 GB/s
x16 PCIe 6.0: 121 GB/s

Or they could offer their own proprietary bridging solution for the MP. But PCIe is already part of the MP, so using that saves them on development time & costs.