New Mac Studio with M5 Max and M5 Ultra(apple.com) |
New Mac Studio with M5 Max and M5 Ultra(apple.com) |
I can't hate a direction where Apple becomes more about building great computers rather than trying to force more and more subscriptions. I do wish they would fix many of the long-standing OS and native app problems.
Ugh. I was waiting it out on a very old Mac, but the pricing these days is insane, so now to wait it out for at least two years more and hope for better pricing or just suck it up and pay a lot for less.
Also see: https://apfel.franzai.com
I'm glad they brought the 512GB option back.
Really want to get my hands on a m6 32GB 2TB model but don't really have a need for one.
I know there's tons of marketing language, buzz words and attempts at convincing me of some agenda that isn't super clear without lots of effort in "validating" the slop. I guess its not bad "slop" though if a human put in effort in editing it (imo >50% human curating = not really bad ai slop)
Though I still would prefer I could just get the prompt. What human thoughts, direction and "prompt" went into writing this article? in the same way as we ask for the prompt for AI generated outputs, I would prefer it for human generated output too. For writing at the least. I could have saved time, got the purity of the argument, and got more clear information. I wonder if we can get a future where humans just express their intent with each other and stop trying to hide our agenda; I want a world we can trust each other greater and interpret and act on our goals without the noise of trying to impress or market to each other & the additional words that go into that.
It’s more that it’s a very parallel architecture than fast.
The LPDDR5X under the hood is slower than what you’d find in a graphics card vram of years ago
That’s why you get consumer macs with 512gb while GPU makers are reluctant to give you more than 16gb unless you pay dearly. It’s not the same kind of mem
So not as flexible as apple's unified memory.
We make a lot of price/performance compromises for having an attached screen and keyboard on our computer. That was what got me started.
Then I remembered the days of having to go to a special corner of the house to use a computer, vs now when I have a computer with me all the time. In my bag, on the sofa, on the train. Hell, I'm writing this on the work MBP while waiting for an appointment.
And you know what, I think I got more done when I went and sat in a corner of the house all those years ago. I set up an area for "computer work", and it worked really well.
I have a home office, but it's a jumble of cables going into docking stations and all sorts of weird stuff. I think if I streamline it and turn it into a proper "computer room", I might get some of that mojo back. I might even convince my partner that surrendering the home office and having a corner of the den might be good - she can watch TV while I tinker. And I won't be balancing a laptop on my knee and trying to do two things at once.
And the price/performance thing comes back in. Hmm.
And that, from mental load standpoint, is not healthy for most folks.
Never ended up happening. I didn't get much done at home at all and what I did get done was mostly on the couch; not a great environment for serious work. (I have an office where I do for-profit work so it's not an issue, but still, I like my side projects too)
Eventually I had enough decommissioned computer parts that I could assemble them and re-commission them into a working desktop. So I now have a desktop again. And I actually end up sitting at the desk working on stuff on the desktop in a way I rarely did with the laptop.
I can tailscale into my home network, but what next? VNC/TeamViewer/Remote Desktop setups are all mediocre imho.
Screen is small and is only one. Ergonomics is entirely messed up. Either your screen is too low, or your keyboard is too high. Keyboards are non-ergonomic and have to be made with compromises due to height limits. Touchpad instead of mouse/trackball is compromise for many - and also stuck at one position.
And yet they have somehow spread despite number of people going on business trips not really increasing.
It’s a whole “thing” when I go to the computer now. And frankly, it’s made it way more fun to use. It’s kind of like enjoying the process of listening to vinyl rather than pulling up a song on Spotify
HPSS works well via 5G, and Tailscale is incredible for the setup as much as the hardware.
Edit: I ended up getting a 15" M5 Air but honestly trying it on a 13" M4 Air was actually superior because when all you care is mobility (since the Studio does the heavy work) its nice just going around with almost no weight / bulk.
https://support.apple.com/en-my/guide/remote-desktop/apdf8e0...
E.g. what most major tech companies have - a laptop that is for VSCode via ssh/web browsing, and a beefy Linux dev box you ssh into for everything else?
It's way cheaper, and what I use at home too - a lot cheaper than a Mac studio for everything, especially with RAM and storage.
Just Tailscale into the Studio from iPad Pro 13" with magic keyboard and Kit Knox's rootshell:
https://github.com/kitknox/rootshell
Note that the iPad Pro can also drive a 4K second screen if you like, and most anything else a Macbook with a single port could drive.
Why not Macbook Air? Because the iPad Pro is also a tablet, touch screen, and 5G...
And I'm also gonna ride my M1Max Studio until it dies (hopefully it will outlive me xD)
I was always cobbling together ad hoc network access, to get from one machine to another. Access to my Mac Mini while traveling was a pain. Bringing it with me is ridiculous, and network access was a PITA.
Tailscale is wonderful magic. Free (for my usage), so easy to set up, and now no matter where my various computers are, they are all accessible trivially via a single ssh connection.
So get a beefed up desktop Mac, set up Tailscale, and then use any random laptop, anywhere, to use it headless.
They gave me a nice MBP for my new job. I tried doing heavy work on it locally, it was fine for that, and yet it still ended up being a light terminal into an EC2 instance, partially because their stuff is on AWS and latency is way lower within that. My personal mini is running too.
Everything is a container or VM now, and none of it runs locally for me.
I have a desktop (older intel) with giant monitors and a keyboard for when I sit at the desk. I have the laptop for when I travel, go out or just want to work from the couch.
When I do my next upgrade to "better hardware" I'm not migrating a machine, rather I'm migrating the containers. My workflow is such that if I loose one of the boxes I sit at to a cup of coffee I really wont care other than the financial loss of a new laptop or keyboard.
The biggest win in all this was dumping the off the shelf firewall/router and moving to Opnsense. Wireguard vpn lets me route all my traffic through home for all my devices (and what is now a growing home lab).
There are scopes of work that this setup would not work for. I would not want to be a video editor with this set up, it's not ideal if you want to play AAA games. But for what I do, it is pretty ideal.
The issue (apart from my Mac mini still having the old, bigger form factor back then) is that this requires a full shutdown (obviously), and that is more friction than opening a closed laptop lid. It just takes a moment for every app and background service etc. to settle back in after, and depending on how your brain works, you might not like that a whole lot.
It absolutely is a cool feeling to carry a pretty mighty desktop in the backpack, though.
It will get worse with the MacBook ultra rocking a M5 too.
Apple’s marketing department is gonna kill me.
Funny that after the price started to increase last year (I think October/November), there was a window of a few months where Apple prices stayed the same as they were before, thus making apple prices actually good when compared to the rest of the market.
It was a unique opportunity to have acquired a 512G M3 ultra for $10k.
I'm currently SSH-ing into my workstation from my M4 Air with 24 gb ram and it's ideal for this flow. Slack/editors/clients/browsers/etc. easily gobble up over 16 GB. I have no more dev tools/compilers/source on my client machines, everything is dockered up on remote workstation in isolated VMs (too much supply-chaining).
My only downside to using a pro is not having 120/4k HDMI port on air and my dock won't support it with Apple (it does with windows)
I’m also the kind of person who close tabs and I like to work on one task at a time.
(Edit: saying this as someone who just paid for a new 128GB M5 Pro Studio -- but that's for LLM's, not regular work)
I bought a Macbook Air in the interm waiting to see what the Macbook Ultra looks like, but honestly—it's the best form factor ever. I love it! Even though it's only like a pound heavier, the Pro feels like a monster.
Thinking about getting a Mini or Studio with more horsepower to stay at home.
Any other apps or suggestions? I'm still not exactly sure what working this way looks like.
I'm waiting this out.
I decided to get Mac Studio M4 Max, also all maxed out config and the cooling is so much better that I can run local LLMs like Gemma 3/4, gpt-oss 120b all day long without any heat issues or any audible fan noise. So for my use case it was the right decision. I subsequently added 15'' M5 Max MacBook Pro all maxed out to my collection and even though it is slightly faster on LLM inference (I get 100 tokens/s with Gemma 4 27b model), you just can't run LLMs longer than a few minutes. It starts overheating and gets really loud.
Not exactly "future proof" for >1T parameter models but good for targeting specific lower-parameter models, or if you can rely on pipeline parallelism and run a cluster.
Wow, "Local AI" mentioned in the subheading above the fold - it's really awesome to see Apple leaning into this use case and I think it will definitely pay off for them going forward. Fingers crossed Apple is able to put some engineering effort towards shipping with one of the frontier open weight models included and optimized exactly for the machine.
Tend to be better than Apple AI.
I actually wouldn't want a serious model shipping 'on disk'. Models release so often, it's going to get outdated quickly. LMStudio is trivial to set up.
For a non-quantized Deepseek V4 flash on an ultra, I would estimate about 1000+ tokens per second prefill and 50+ tokens per second on generation. This is actually quite usable and near parity to cloud.
They mention "adds the GPU Neural Accelerators." which, if exploitable for LLM loads, would probably help the prefill a lot
"Storage performance is up to twice as fast, with a next-generation SSD architecture built on PCIe Gen 6..."
This is the first personal computer I've noticed that has PCIe Gen 6 storage. I've only seen enterprise PCIe Gen 6 SSDs up until now. Gen 5 SSDs in consumer devices already have high temperatures and thermal throttling, so I'm worried about how Apple's implementation will perform (I know they don't use off-the-shelf SSDs anymore, but I'd imagine the temps would still be a problem).
Personally, I can't justify dropping the dosh for a 512GB M5 Ultra, but I would be able to justify it to myself if I could get 1TB of memory, because it'd guarantee the flexibility with local models I currently am missing. Seems a huge miss to not offer this... for a price.
Apple also launched the base M6 today with a 32GB RAM limit, suggesting 512GB may remain the maximum for Ultra chips for some time. Since these Ultra chips combine 16 base chips:
32GB × 16 = 512GB
I am in Europe, and the Mac Studio M5 Ultra GPU 64 cores with 96GB RAM is up to 6.649,00 €. Ouch.
> A fully configured IBM Personal Computer AT (Model 5170) with expanded memory and storage cost around $5,795 to $6,000 at its launch in August 1984, which equals roughly $18,600 to $19,300 in 2026 USD.
It would be significantly cheaper to fly to a tariff-free country and buy there.
There are no configurations even close to running something comparable to frontier model variants, they're simply far too large, but something like full precision Qwen 35b or DeepSeek 70b at 50+ t/s is well within available configuration, and potential for plenty of room for large context sizes.
I'm using Flash heavily, and I would describe it as nearly as intelligent as Sonnet-class in agentic coding, but more usable. Less world knowledge of course, and definitely a bit less intelligent; but not _that_ much.
On usability: Takes less handholding, less likely to make unsolicited refactors or whatever, and the writing style is readable.
It's not great at super-long-horizon goals as the Claude 5 models are; but if you have a good harness, you can get around that.
Maybe $17k for a 512gig system that can do 1.2TB/s seems like a pretty good deal for a small office.
Isn't 170GB/s slow for bandwidth?
Compared to something like VRAM it's slow.
1.2TB/s memory bandwidth unlocks a lot with 256GB unified, and agentic AI is pretty good at optimising performance.
For comparison, to get 256GB with NVIDIA, you’re looking at a DIY workstation build (need pcie lanes), and like $70k?
The spark’s ~250gb/s bandwidth doesn’t really count here.
The comparison should be against renting in the cloud for the duration of your task for training and research or using pay-per-api-call providers for general inference instead of buying your own hardware (and paying the electricity and cooling bills on top), because let’s face it, the models you want to use are probably the same ones available on inference providers (but, yes, some are more trustworthy than others).
Speaking as someone that does ML/AI research, you are essentially paying a huge premium for being able to just run your Python script at any time without setting up a deployment script and harness to run the job remotely, while your hardware sits essentially idle the rest of the time.
The only way to make the math work is if you rent your hardware in the background for inference while you’re not using it in anger, but despite all the startups and promises that has never become as streamlined as mining bitcoins or shitcoins used to be and they don’t pay out as much as they say they would. Renting your hardware for training is another option but doing that is a lot more involved, options are fewer and farther in between, you won’t get as much utilization out of it, and doesn’t let you feasibly abort running tasks at a moment’s notice.
1-2$ / hour.
I'm not regretting my setup at home as it got paid by my company which makes sense here, but paying for electricity is quite high and makes already 0.3$/hour alone.
I would argue, the most interesting use case for running it at home is some personal agent which you want to run 24/7.
256GB model is $10k and the 512GB version will probably be double
Seems like miscalculation. If they had their own fab for RAM, they could completely corner the market today.
Most people at Apple have already realized that their processors are already too powerful for regular users - heck, as a developer my M2 Pro with 32 GB RAM is more than enough for me.
Regular users don’t care about local AI either. So, they will probably extract as much money as possible during AI gold rush, but then we will most likely see Apple
a. Making their software worse (god forbid, forced updates)
b. Making their hardware impossible to repair (as they almost accomplished this already) and easier to break.
Plus introducing features like RDMA over thunderbolt, which is critical for distributed inference/training/etc. On the software side, Apple is investing heaps.
It's still ridiculous they don't support expandable NVMes, but the memory being soldered makes sense, you need it for 1.2TB/s bandwidth.
They are still selling high-margin hardware. Apple loves selling high-margin hardware.
EDIT: AMD too, it's not limited to Nvidia, nice.
They're eating nvidia's lunch.
I may consider a M6 Mac Mini as a stop-gap whilst waiting out RAM Apocalypse to be over. Basically abandoning any ambitions of AI sovereignty and riding out subsidised LLM pricing for the next couple of years.
I don't actually own a car and my startup is bootstrapped and our salaries are modest. But the one thing we spend on is laptops. I have M4 max pro with 48GB. That thing was on the expensive side (~4.5Kish). But it delivers a lot of value and I spend most hours I'm awake using it. I like fast builds. I like that I can try out open source AI models. And I like just having the option to run those.
We actually lease them and mine costs something like 105 euro/month. Including Apple Care. I don't need a Mac Studio but I could see some roles where that would not be a crazy expense. Even the tricked out version that basically only costs the same as a very modest car.
But +4000$ for an additional 128GB of ram is simply milking the customers, as they know they will have many of them.
Having owned 3 MacBook Pros since 2008, the decision to make my next computer be a Mac Studio came down to (1) MacBook thermal throttling that slows down CPUs when it starts to overheat and (2) easier upgrade of Mac Studio SSD with after-market storage module whereas the MacBook requires more complicated disassembly and hot air gun to dislodge the surface mounted SSDs.
I have a brand new M5 Pro MacBook Pro I don't like it when the fans turn on. The Mac Studio will be faster and quieter for the same workloads.
Turns out that even with the fans at full blast, the batteries didn’t like being so hot over prolonged periods.
I’m happy with my M4 mini now, with a separate windows pc for gaming.
I was using a VM setup on my MBP but it felt like a huge waste, having to leave a laptop on 24/7 when all it did was run Claude Code inside VMs.
I likely will stick with a Macbook Air 15" for next purchase, and beef up my "Claude Server" down the road.
Thus, my current workhorse is a Studio.
These prices make that decision so bittersweet. I feel so shut out from being able to use AI as economic productivity without being able to take on debt for a local-inf capable mac studio
Is that the new phrase for social media? I hope it is...
This changes when you get into connotations that aren’t available in the laptop form factor, but with RAM prices the way they are those configurations are more than I want to spend on a local machine right now.
For running local LLMs the high memory Mac options always look appealing, but the processing speed (prefill) is so much slower than GPUs that it hurts. For situations where you have no rush and can let something work in the background for 24 hours it can some times be ignored, but the speeds I get from a real GPU setup are so much faster that I never use the larger models on a high memory Mac any more.
Most, if not all, of our current work happens on remote cloud vms. Now I'm stuck with carrying a 3KG monstrosity to work.Every day.
Absolutely no positives compared to my ThinkPad that weighed less than half in my previous job.
Remote development is "good enough" these days. With VS Code, Development containers etc... Having a light weight, portable laptop is so much nicer than a laptop that you can't even rest on your lap for long duration.
The only thing you need to be mindful of using light weight laptops is not having enough RAM to fit all your browser tabs..
I also suspect that trying to stream larger LLM models from disk hammered the USB 4 connection too much - lead to system hangs, so now stream from the internal SSD instead.
But in the end I ended op buying a Lenovo Legion and put Linux on it.
Laptops are so fast these days that I didn't want to be bothered with setting up connectivity to a remote desktop.
But if your laptop never leaves your desk I think a desktop computer is a great option. Relatively cheaper and easier to maintain and upgrade.
Fir kinda the first time in my life I don’t really have development tools on my personal laptop. I ghostty and openvpn client installed.
I have a large remote linux workstation (2x 8c/16t xeon cpus, 256gb ram, 2x8tb spinning rust disk) and i have my tools over there (along with some VMs).
It works surprisingly well.
Also, the macbook neo is a surprisingly capable little machine.
There's definitely an option for everyone if you want to go down the "fixed supercomputer, remote terminal" route.
Makes managing both backups and handling failure scenarios involving loss or unauthorised access to the laptop less of a hassle as well.
It will be new Mac with more GPU cores for ai, right?
It would sit in my basement server room providing local inference.
However, after retiring, I realized I never undocked my MBP.
So I got an M4Pro Mini, and I've been thrilled. If I ever get to where I travel a lot, again, I'll get a laptop, but I don't see a need, right now.
While they're theoretically at parity, in actual use, they're not.
I personally own an M3 ultra, an M1 max as laptops, but my desktop is a Ryzen desktop I built in 2022 and it was a third in price of the ultra for more power.
Would you get away with a Mini?
So a combination of a powerful desktop and a "cheap" laptop might indeed be attractive.
Computers are never "future proof".
A 8 years old graphics card can still play modern games. Ten years ago playing a modern game on hardware that old would've been unthinkable.
And with how the market is right now, we'll be stuck on the current "reference level" of hardware for a while longer.
It was future proof but not really because it struggled a lot in its final years.
i'm still using an old i7 3770k @ 4.8ghz with 16gb (ddr3) ram running linux for random tasks like executing tests. obviously, power consumption is higher.
my main machine is a MBP M1 Max which i use for everything. i also have my main linux desktop workstation that has a 5950x with 128gb (ddr4) ram.
i'll probably get 2-5 years out of my MBP, and my AMD workstation will probably be good for another 5-10 years.
i'm not a gamer, but i have a 3080. i'm sure my 5950x will still be good for gaming in 10 years if paired with a modern GPU.
Upgradeable components however could go a loooong stretch towards that goal. It can't be that hard to follow a common form factor for at least the housing across two or three generations to allow a reuse of everything but the main PCB.
My thinkpad t14g1 with 16gb ram somehow still relevant today.
A NVIDIA RTX 6000, 96 GB at 1.7 TB/s, is 13 grand.
This 256 GB at 1.2 TB/s Mac is extremely competitive, it will be sold out everywhere.
But my next machine for LLMs will probably be the Mac Mini (with 64GB vRAM access) for low-$2300s – which is what I paid for my MC-Prebuilt (and am still very happy with – Ubuntu is great).
After decades of installing and forgetting various shades of linux-distro, Ubuntu has kept me "using" the computer as a tool, instead of "just tinkering with it").
But, that doesn't make it a good deal. It just means the Apple tax doesn't apply when stacked up against AI machines and with memory prices being so out of whack. I'm still planning to wait until the RAMpocalypse ends before I buy any more hardware.
46, actually. 46. I am pretty certain that the explanation lies completely elsewhere...
(overuse of overstatements, for the sake of manipulation, perhaps? something that makes their saying unreliable this way - e.g. they forget to mention the other end of the range, or perhaps the mean, completely .... except when it is good for them, e.g. prices ... that's borderline to lying this way...)
I want a computer dammit, not an appliance.
It's just branding - they don't want to sell you a generic "computer" - they want to sell you a Mac. It's really not all that different from a computer. Might look a little nicer[1]. Uses a custom processor. But you can write software for it like a computer. And it can display things on a monitor like a computer.
They do have wrist, pocket, and tablet versions of computers, but those might be considered more like appliances.
1 - taste being subjective and all
(Scroll-oriented web pages see me running away fast. Not only Apple, which is a champion in it, but articles in general. That make me a passive spectator eating what they want me to eat, the way they want me to. Rather than a free discoverer. I hate being pushed into something. I like freedom.)
Staggering?? I can count on no hands the number of times that a fact or figure has caused me to stagger.
$100/mo for ten years is $12K. $200/mo is $24K. When you look at the valuations of the companies providing these tokens, you don't necessarily see these rates going down.
Your argument against privacy is so obviously bad I don't think it needs a response.
Today, most companies and individuals (including me) are still going to be buying tokens from the cloud. But for many it makes sense to go local today, and it seems pretty clear the day where we turn the corner isn't that far away. E.g., suppose the RAMpocalyse eases in 3 years and Apple releases the M7 ultra, which has actually been designed for local inference (and other players are making their moves as well). I suspect it will become something of a no-brainer to do most inference locally, perhaps supplemented by lighter usage of per-token plans from the cloud for access to specialized models.
There is definitely classes of data that you legally can't share over "fucking gmail, whatsapp and instagram" and you can't send to commodity token providers either.
Taking liberties to save money at the expense of regulatory duties is a good way to get hit with a massive fine if you work in financial or health services.
This is a product aimed at niche professionals doing paid work. If you don't need that just move on to something else.
And no, I do not use gmail, whatsapp or instagram :-)
for me, it’s significantly more about the process than it is about penny pinching. it’s why im not an accountant and why i often despise bean counters.
my dad has soooo much money in tools alone for working on cars, faaar more than he’d ever spend at a mechanics. add in the costs he spends buying hobby cars and real estate/upkeep for the garage to work on his cars and it adds up quick. people spend a lot on their hobbies.
look at how much woodworkers spend on tools. look at how much homelabbers spend. photographers, boat/sailing wonks, horse owners, and on and on and on.
add in the above, then the value of the privacy aspects, idealism aspects, potential career knowledge, etc… and it’s obvious that local models will do well.
Another example as a developer, one particular large project took an hour to build on a spinning disk, on a SSD it takes 2mins
So, yeah, it's really noticable improvement
So while have having this massive almost-symmetric fibre pipe is cool on paper, I haven't felt a huge need to install 2.5G gear all over the house.
data processing, LLMs, model loading, MoE loading, etc, etc relies on very fast storage to keep your GPU saturated.
I think someone made incorrect conclusions and everyone just copied them.
"Storage performance is up to twice as fast, with a next-generation SSD architecture, delivering industry-leading read and write speeds for rapid project loading, file transfers, and loading huge LLMs."
Here's the original, archived press release from August 25th (look under "Blazing-Fast Storage and Pro Connectivity"):
https://web.archive.org/web/20260825130329/https://www.apple...
And here's the first modified version that the Wayback Machine has, from August 26th:
https://web.archive.org/web/20260826020313/https://www.apple...
So it seems like Apple made a mistake in this case. I agree with you that it will likely be PCIe Gen 5 - that's much more logical.
Basically, Apple gets to cheat because they shove everyone onto the same IC & make the OS that runs on this SOC.
I have a bit of paranoia/anxiety about AI, but it's not what most people are concerned with. I understand the limits of these tools very well, and still find them extremely useful. What concerns me is that it's going to become difficult to impossible in the future to run local models which have near-SOTA capabilities in a way in which you can exercise full control of the model. I see the writing on the wall, and its more than worth it for me to invest early to ensure my own capabilities. I am very much not a fan of our "you'll own nothing and be happy" directionality for the world, and I am (at least currently) privileged to have the means to slow that decline for my own self.
Only reason to buy this if you want to own your compute.
Experimentation and inference are all going to be cheaper on the cloud
My ideal Mac Studio would have an NVMe slot or two for additional internal storage, and those slots could be accessed without opening the case. But Apple won't do that, as convenient internal storage upgrades would likely lower their profits.
In many ways, the Mac Studio is the spiritual successor to the trash can Mac Pro. Apple Silicon's architecture solved the thermal problems that limited the trash can Mac Pro.
Apple wants to sell you a new system every couple of years, they don't want you to upgrade. They want you to throw it away the system completely and replace it (and they claim they are environmentally sustainable, btw) when in the PC world you would replace RAM, CPU, GPU, motherboard, maybe the SSD, but surely you can keep the power supply for decades, keep cooling fans, not talking about the chassis that could last 50 years because it's a just metal box.
They are a rip off these products, and I don't understand people that buys them.
BTW even when Apple did produce towers like Mac Pro they did so in a purposely non-standard way, such that you couldn't, for example, replace the PSU with a standard ATX one, the only standard components were really RAM, HDD, CD-ROM drive and in Intel models the CPU, and of course PCI expansion boards (sort of, because to work GPUs needed to have an Apple specific firmware in the old PowerPC models). You couldn't for example upgrade the motherboard, if you wanted to change the CPU to a new generation you also needed to buy a entirely new Mac with a new chassis, new PSU, etc even if what you had were perfectly fine.
We no longer need CD/DVD/floppy drives, storage has shrunk/moved to the cloud. The only thing that's really grown inside a pc case is the video card, and most of these ITX cases are built specifically around fitting popular cards.
Even folks primarily focused on gaming are probably thinking that a full ATX case is a lot of wasted space.
Maybe it's just me, but I think ATX full and mid towers are going the way of the dinosaur.
Even standard ATX cases are kind of small if you want to build a full custom loop and need space for radiators, pumps, reservoirs, tubes, etc. My O11 XL is admittedly enormous, but it is under my desk and does not occupy any space that wouldn't otherwise remain unused.
As for storage, I categorically refuse to store things in the cloud. My files belong at home on my hardware, not on someone else's computer in the "cloud."
They were obsolete a decade ago, but priced out by artificially expensive "SFF PC" cases, fans, and power supplies. PC building completely transitioned to fashionable cargo cult in the late 2010's, with the rise of twitch.
You can't upgrade the processor because it's soldered. You can't upgrade the ram because it's soldered. You can't upgrade the SSD because it's soldered / bonded to the CPU. You could maybe have some PCIe slots, but not many drivers for macOS, so what's the point?
Yes, it was different in the old days, but Apple is ever more a closed hardware architecture with few options. If you want choices, Apple is not for you.
This is the exactly reason.
If your computing needs line up, it's a very serviceable approach.
I haven't added my iPad to the Tailnet yet but i reckon that it could become a very comfortable and productive device for me.
Its update policy can result in this sort of regressions in a middle of a release. One thing that's nicer thing about macOS is that a released version doesn't regress as much (even if they are buggier at release; but in that case you can just wait until a later point release to upgrade).
But the couch is just so comfy.
I know AMD has announced the replacement, and it will be faster, but I’m not sure when it’ll ship.
Also, you can cluster strix halo if you need more than 128GB of ram.
Real-world matmul is about 4x higher on a DGX Spark than a M3 Ultra.
I meant the only downside of not using a pro for my use-case is that I don't have a HDMI port on device with high refresh rate. I have a decent dock that can give me 4k/60hz HDMI but not 120hz refresh for macs (it works on windows). It's a minor thing but having 120hz is nice when I'm docked at home.
Despite being outdated in terms of compute, it stills let me run very good recent models locally, with Deepseek V4 Flash 0731 being the greatest one right now, and hopefully Qwen 3.8 Flash will also fit well when it is released tomorrow!
The M1 ultra definitely leaves to be desired in terms of its token speeds, but I think 20 tps generation and ~200 tps prompt processing (which is what I get with DSv4 flash), is already enough to do a lot of serious work when you combine with the decent prompt caching provided by llama.cpp.
I want a 2026 pickup truck. The dealer also sells a 2027 hatchback.
I am not shopping for a hatchback.
If it was more readily available on retailers such as Amazon I would even recommend 32gb ram air for local workflows - if you're doing development these days you're just going to need it - especially with LLM workflows and parallelizing work. The price difference between 16 and 24 is not even close to not justify picking it for a work machine.
Thing is having more cores/faster cores might make your workflows 1-2x faster but hitting swap is going to be like >10x degradation
Extremely high bandwidth is great for copying data, but not as relevant for walking chains of pointers, where latency/cache/TLB entries are more important.
So just saying "high bandwidth == better" is true when other variables are the same, but they rarely are, especially in comparison to x86-64 offerings.
All of these are independent that it is a SoC with on-package DRAM.
The maths to me was basically equivalent to prepaying for 242 days of runpod pricing for the same GPU; and I reckon I'd be able to get 6+ years of use out of this card with 96GB.
Plus there's the resell value -- it's actually appreciated by ~50% since I bought it.
Plus I do really enjoy that it's 100% local. I wouldn't feel comfortable giving my agents this much information if inference wasn't 100% local.
I wouldn't get another one, I wouldn't have as much value, but one is definitely paying off for me on the financial side.
... and a lot more social media I'd like to admit.
If you’re running a big task or small, everything scales to the appropriate size, regardless of the hardware sitting on your lap or under your desk.
At least that’s how I think of it.
The funniest recurring thing when working at Google was new hires getting baited into taking a Chromebook, then not being able to switch to a Mac for like 2 years. Our team made sure new people didn't fall for that.
Sorry for not being clearer. I don't like the MacBook's noise when the fans turn on.
The Mac Studio has bigger heat sinks to delay the need for thermal management -- and if its fan does need to turn on, the bigger size means it's still silent instead of the high-pitched whooshing noise the tiny fans make in the MacBook.
And if not, would you consider a raspberry pi a computer? or your iphone a computer?
I don't play games, but it still chugs along perfectly fine and for what I use it for I rarely think that it's too slow.
To me there is no reason to update a PC if you don't use Windows or MacOS that forces you to purchase a new hardware to do the same things. Just look at Windows 11, full of useless AI features, weights a ton, and in the end it's probably faster a PC from 25 years ago with Windows XP.
I love taking meetings while walking but since Covid everyone has their camera on making sedentary meetings mandatory.
I expect it to stay a mediocre gaming PC for the next 3 years maybe 5 years.
16GB RAM RTX 3060 TI (8 GB VRAM)
I also like 4K with 90 to 120fps which leads me most of the time to use almost minimal settings.
Look at steam hardware reviews. My setup is quite the default thing. Games optimize for it
Good for inference; however if you like to train, data format support and effective performance is limited (M5 Pro). Some hardware features are not exposed or extremely slow.
You’ll be fine for inference, but pales in comparison to what a RTX 6000 Pro can do for compute/matmuls/training.
I’ll be selling my M4 MBA soon, I genuinely use the Neo more. Huge difference in typing experience.
Great repairability is a plus. It was super easy, and actually fun to open. Felt like unboxing an Apple product. Applied the thermal paste mod for $10 which works excellently; I’ve had it shortly after launch.
And I love the notchless display, even if I wished the color gamut was a bit better.
https://ofox.ai/blog/qwen-3-8-27b-run-locally-vram-gguf-2026...
iMacs are great for a lot of use cases, but my image of the typical HN user would prefer to keep the monitor separate.
Are your referring to Taalas / chatjimmy ?
I definitely think we'll see an ASIC-like approach in the future, especially for embedded small models where it may require minimal silicon area and can result in near-realtime performance. But at the frontier, I don't think this is a solved problem and will continue to have model weight churn that will advantage more flexible general-purpose hardware.
Not that many of us, actually. Only Montana, New Hampshire, Oregon, and some parts of Delaware and Alaska have no sales tax. https://commons.wikimedia.org/wiki/File:Sales_tax_by_county....
https://tax.idaho.gov/taxes/sales-use/use-tax/online-guide/
This is difficult to enforce, but it seems that "taking advantage" is strictly speaking just tax fraud.
At that point just rent proper GPUs in the cloud, you'd have way more power and pay only what you use for.
I'm not optimizing for $$$ spent today and agree that this isn't the cheapest solution to a specific capability problem, but I don't super mind paying for a cool piece of compute that I can do some fun (and educational) stuff with.
Would be nice if someone knowledgeable about electrical engineering and manufacturing processes could lay out some valid reasons for manufacturers to integrate RAM onto the motherboard.
The SO DIMM Ram slot, which is designed, idk, 30 years ago? is not really capable of handling the frequency we are targeting (close to 10GT/s)
Well it might be an idea to keep the layout of the mainboard and connectors the same.
That way, instead of having to upgrade the whole machine, all it would need is a new mainboard. Framework for example managed to pull that off, and in mobile at that, where constraints are much worse than for a desktop computer.
It's not the same thing though. On the M-series, CPU and GPU share a unified memory architecture and ram is much more tightly coupled to get it to go faster. A closer example would be the Framework desktop, actually, where memory is also soldered in for the same reason.
It’s a very non-Apple thing to do, but it’d be pretty awesome if they did.
RAM production is completely sold out for 2027[1] which means the prices are locked in until after then.
It takes about 2 years from the time ground if broken for a new fab to be built and producing RAM.
There were some new fabs announced between February and April this year by both the Korean and Chinese manufactures, so that new capacity might start having an impact in 2028 in the most optimistic scenario.
Samsung says supply will remain tight in 2028[2], and Micron says "tight beyond 2027"
The best hope is that new (Chinese) players overbuild fab capacity and supply outstrips demand. That isn't likely, but perhaps in the late 2028-2029 timeframe could happen.
[1] https://www.techpowerup.com/351344/memory-makers-seal-2027-d...
[2] https://www.tweaktown.com/news/112966/memory-shortages-will-...
[3] https://s25.q4cdn.com/621799436/files/doc_events/2026/06/Q3-...
Producers (Samsung, SK Hynix etc.) will not say "prices expected to drop" or "demand expected to drop" even if it was true because then consumers would start delaying purchased.
The big buyers (OpenAI, hyperscalers etc.) have no incentive to say "supply expected to start opening up" because that would imply their growth trajectory is flattening; also a huge part of their moat now is just deployed RAM.
Otherwise you'll have to wait to see if the AI circular financing club collapses- if you still have a job, there should be deals to be had...
Efficiency is improving, both in hardware and in software and in intelligence density (smaller models can effectively do more of the AI work that needs doing), so I think the pure data center plays will falter. If there isn't some other business attached, they're never going to recoup their investment. Anthropic and OpenAI are buying all the compute they can find right now, but efficiency gains, especially those coming out of Chinese labs where they must be more efficient to compete, will make it less and less of a problem.
I mean, think about the hardware we use for AI. It's basically an accident. GPUs were not designed for AI (though they are becoming more focused on AI). The specialized AI hardware industry is just ramping up.
So, we're still early in the curve for how efficient both the hardware and software can be at performing these tasks, and given the effectiveness of recent very small models (e.g. DeepSeek V4 Flash 0731 and Qwen 3.8 27B), I just don't see a long future for giant data centers built around billions of dollars worth of last years graphics cards. As with the crypto mining operations, at some point, it becomes more expensive to run the hardware than it makes in revenue. And, as with the crypto mining operations, when the money dries up, the hardware hits eBay and prices drop.
There's just less space, so it's harder to get things in and out. And you have a smaller cooler because there's less space, at least for air cooling.
I haven't used itx with video cards, but that's hugely constrained...
Water cooled but it does get a bit loud when I play games, silent for regular usage.
Obviously depends on your team culture, but try it :)
If you just want to run Qwen 3.8 27B and Deepseek v4 Flash in perpetuity and that's it, there are a lot of solutions that will work and this is a fairly user friendly one.
Lastly, I want to clarify that prefill on an x86 CPU is drastically slower than on an M5 Ultra GPU.
It feels like PCIe is a bit of a boat anchor here. There's a SATA->NVMe style transition waiting in the wings to make this all so much better. We really need post-PCIe GPUs. CXL with it's very small low latency flits. This is an "almost certainly not" but I wonder if you could mix PCIe and CXL so you could have the GPU memory expose vmeme as a bunch of CXL.mem pools but still have an otherwise pretty normal GPU. It seems madness that UALink went all in on GPU-to-GPU with no affordances for connecting to host computers.
¹ drivers are responsible to get the car into a low earth orbit themselves and need to be familiar with the orbital maneuvers needed to accelerate the car to 900mph
Years ago I realized running a 100W PC all the time was REALLY expensive, so I switched to an old linux laptop at 30W (about $4/month). That helped, but I moved to the m1 mac specifically for the power efficiency. It does all the things my linux server did, and it does them at 6W (75¢/month).
I've found no competitor with similar performance that can run in such a low power footprint. M4 mini's are way better at power/performance, and I imagine their price is about to come way down since the M6 mini just got announced.
Software-wise, it's different, but mostly equivalent. Homebrew or macports has a similar software inventory to debian. And Apple's container framework is a welcome improvement over colima for running most container workloads.
Macs are totally “fine” for light server duty… as is just about any computer of the last decade+. The CPUs are beasts, the disks are screaming fast.
The operating system itself may not be ideal at serving but you can just run Docker/Orbstack if you need to do something especially Linux-y.
I’d put the question back on you — what are scenarios where an Apple Silicon Mac wouldn’t cut it as a light server for one person or a handful of people? About the only scenario that comes to mind is scenarios where you expect to utilize it so heavily that the fans are running for many hours a day. At some point those are either gonna wear out or just ingest so much dust that the machine runs hotter and needs a deep clean. But even that is largely mitigated by just pointing an external fan at it.
When memory prices come back down, I'll be down to the Apple Store. But, it doesn't make sense to buy hardware right now.
- 120v input plug
- not rack-mounted
- has a video out port
Putting aside the fact it is a marketing label, "Pro" usually means "designed for work" while "Consumer" (in this context) means "doesn't need a special environment".
In computing the distinction is primarily noise, power and cooling requirements.
If a computer is designed to use home power and is quiet enough to use without annoying people and doesn't require specialist cooling then it is a consumer device, even if it is used for work.
Hence why they had to make up the "Studio" brand for the workstation market, because they'd already fully removed any meaning from "Professional"
If you haven't tried Screen Sharing since they deprecated VNC and switched to their proprietary H.264-based protocol, its worth trying. Even YouTube videos play just fine with no noticeable lag.
This open source software was built for lag-free gaming over a network, but I've found they're great for desktop usage too. I frequently use it to remote into my Windows tower. Feels about as responsive as if I was sitting at the desk using the tower directly.
Unfortunately, it relied on Display PostScript and Quartz née Display PDF isn't architected to allow that sort of remote display/access on a per application level.
It worked relatively well, but wasn't perfect. It was as close to what you're describing as I've seen, though. This was a few years ago; something like this might work even better now.
It’s pretty good.
To be clear, it’s not ever supposed to be a primary. It just becomes one from time to time when I make it full screen and am not paying attention.
Are you using Apple displays or did they fix it?
Last year when I was browsing the only recentish Apple silicon capable of driving that was the M3 Ultra.
Care to guess the approximate price of the MBP I bought earlier this year?
This new Studio? Can't find a config under $5k I'd bother with. But for the MBPs that number still mostly tracks for the average Pro user. (I buy large and run it into the ground so long I mistake the ground for the computer's remains.)
My still-being-used 2012 MBP (which cost me about $3K) says, “hi”.
And, as you point out, the new computers I want blow Dvorak’s hypothesis out of the water. Never would I have guessed 30 years ago that Dvorak would be wrong the other direction on price.
What is changing is that there genuine demand for more capabilities disproportionate to the cost decrease curve. Fab demand and supply constraints have slowed or even reversed some cost decreases - but that is still getting absorbed by the overall systems costs when you are looking at things like laptops. If you all you want is the last decades demand to browse the web and use office - things are cheaper than ever.
But for the mac pro, they would say something very obscure.
"100-120V AC or 200-240V AC (wide-range power supply input voltage)" and the maximum current is "12A (low-voltage range) or 6A (high-voltage range)".
They're trying NOT to say it had a 1440 watt power supply.
It isn't, though. That rating is for circuit sizing. Power supplies are rated by their output power. If you assume an 89% efficiency for the power supply, and assume 1.0 PFC, and assume that it must be able to supply its rated power at the critical corner (low voltage), and assume zero margin on the current rating, then this is a 1068 W power supply. More likely its design spec is around 1000 W; it certainly can't be 1440 W.
There is some consolation there
It's not like you have to be on a plane for the portability to be useful.
Still, being able to take you laptop to a meeting to either project, or to look up related material to catch up and clarify before making a point in a meeting etc. is invaluable (no, working at a laptop during a meeting doesn't mean you're not paying attention, it often means you're cross-referencing discussions with other data live).
A laptop is a distinct tool from a desktop and if you try to use it like a desktop, I agree it is a terrible substitute. Personally I find laptops vastly more ergonomic than a desktop, but I never program with my feet on the ground.
I've found my smartphone's the best device for that. I wrote an entire programming language using my phone.
Been trying to make a handheld cyberdeck to replace the phone with something better, but it's still a long way from being real.
The mini sips a truly tiny amount of power. I wonder what the smallest/lightest UPS could could bundle it with would be.
But at this point.. you’ve got a shit laptop.
The shutdown thing didn't bother me, and in fact the Mini was lighter than a Macbook Air at the time. So all in all it was pretty convenient.
Before local LLMs became a thing, I had completely lost interest in buying anything but the cheapest laptop. It felt like "personal computing" was a solved problem. But...it actually isn't, and this is exciting.
I bought a refurbished M3 Max a couple of years ago just to try things out, I get 90 toks/s which is good enough for a lot of tasks. But ya, when DRAM prices come back down I'll definitely look for a better local rig.
Set up wireguard in my router and now I've got remote access into my LAN. My laptop has essentially become a Linux server. I just ssh in from the phone and write code. Can also easily launch and use codex.
RTX 6000 wins in performance, if your model can fit into the VRAM.
There are very obvious and clear advantages to a Mac Studio. It's an entire system for one and you're getting a world class CPU as well.
[citation needed]. I have personally specced out and built an nvidia GPU-based machine which after some optimization, handily beat the Mac Studio in terms of tokens/watt for LLM inference with most models. This was in the M2 Ultra era, and I haven't run the numbers for the later generations, but nvidia's cards have gotten faster just as Apple's CPUs/GPUs have, so I would guess that it's still possible to do.
> RTX 6000 wins in performance, if your model can fit into the VRAM.
"if your model can fit into the VRAM" can be true for the Mac as well.
> There are very obvious and clear advantages to a Mac Studio.
There are certain advantages for sure, depending on your use case. They may _seem_ to be obvious, but as evidenced above, I believe that many people overestimate the Mac's superiority on the metrics you cite when comparing a Mac vs. a dedicated GPU for LLM inference.
It is much more likely for your model to fit in large unified memory of a Mac than the smaller more limited memory of a GPU. Even going with two 5090s, you now have to shard your model and that is a PITA.
But it turns out that MoE is the solution both for running models on macs of limited computer power means (not as fast as GPUs), and on multiple GPUs that require sharding the model.
[citation needed].
RTX 6000 wins in performance, if your model can fit into the VRAM.
I can run agents using deepseek v4 flash or Qwen 3.8 on my m3 ultra and it will be lukewarm and the fan will eventually start blowing softly.
Yes, the Mac might get lukewarm, but it will take 2-3+ times longer to do the same task.
Also while the pre fill performance sucks, the Mac isn’t that slow and can use much better model compared to a similarly priced Nvidia workstation so it’s not really taking much longer in practice.
It’s taking longer than in the cloud for sure. At least a local computer uses the local energy grid that is pretty clean and not fossil energy.
If you can show me a model for which a Mac is faster than the RTX 6000 then I’ll be happy to update or retract my statement.
A Macbook Pro M4 or M5 drives TWO of these at once, well, for equivalent of quad 4K configured as dual super ultrawides:
https://www.samsung.com/us/monitors/gaming/57-inch-odyssey-n...
Per Apple support page, MBP with M5 Max handles:
Two external displays
Two displays up to a native resolution of 8K (7680 x 4320) at 60Hz, 5K (5120 x 2880) at 120Hz, or 4K (3840 x 2160) at 240Hz over Thunderbolt or HDMI
Three external displays
Two displays up to a native resolution of 6K (6144 x 3456) at 60Hz or 4K (3840 x 2160) at 144Hz; plus one display up to a native resolution of 8K (7680 x 4320) at 60Hz, 5K (5120 x 2880) at 120Hz or 4K (3840 x 2160) at 240Hz over Thunderbolt or HDMI
Four external displays
Four displays up to a native resolution of 6K (6144 x 3456) at 60Hz or 4K (3840 x 2160) at 144Hz over Thunderbolt or HDMI
https://support.apple.com/en-us/101571
PS. Use TB4 or TB5 dock from Cal-Digit (match the TB version of your mac), or this or one of the rebranded versions of the same OEM: https://www.owc.com/solutions/thunderbolt-dock, or an Ivanky Fusion dock if you want to go nuts with up to SIX displays, Quad 6K@60Hz and Dual 4K@60Hz:
https://www.amazon.com/Thunderbolt-Monitor-Docking-Station-A...
(Apparently the laptops do a suspend to disk depending on battery remaining, so the OS does support it. It'd be nice to have it available! The disk space required could become annoying if you only bought the 1 TB SSD for your fancy 512 GB Mac though.)
> Also while the pre fill performance sucks, the Mac isn’t that slow and can use much better model compared to a similarly priced Nvidia workstation so it’s not really taking much longer in practice.
This is exactly backwards; assuming you mean that the Mac has more RAM so you can use a larger model, the mac is going to be _even slower_ since the fastest Mac isn’t as fast as the average nVidia setup.
My point is that comparing apples to apples (no pun intended), an nVidia setup is both faster and (possibly with some tweaking) more efficient than a Mac.
> Mac studio wins in memory capacity, price, perf/watt and value.
I bristle at general statements like this when it obviously depends on the specific Mac and GPU in question. But yes, comparing a maxed out M5 Ultra with an RTX 6000, the Mac has much more memory.
> Even going with two 5090s, you now have to shard your model and that is a PITA.
Every modern tool does this for you automatically. It is absolutely not a pain in the least (e.g. llama.cpp ships with pipeline parallelism enabled by default).
If you are using multiple GPUs, MoE is basically going to be your only workable choice unless you can leverage pipeline parallelism (only half your GPUs can work on a prompt at a time, so you need to process prompts back to back in a pipeline setup, and they better be doing similar things because your vram is limited).
I’ve got a 4060 ti 16gb, and I’m thinking about getting another. I previously specced out a cluster using multiple 3090s. At the time, the 3090s were going for $700 on eBay. They’re more than that now, but there’s no need to spend $4k per GPU at all.
This is a good video to watch.
Yes, I agree.
> What about [...] performance per watt which I’ve now mentioned ~~four~~ five times and you’ve ignored the previous ~~three~~ four times?
The video was not focused on performance per watt at all, and the best attempt that the video makes at measuring performance per watt actually shows the opposite: that the 5090 system is more efficient than the Mac.
When he's running the 27B model, at about 5:02, he shows that the Mac Studio is pulling 251.5W, and the PC is pulling 315.3W.
Then he shows the results:
Mac: 27.62 tok/s PC: 40.92 tok/s
Doing the math, we arrive at 0.11 tokens/W for the Mac, and 0.13 tokens/W for the PC. The PC is about 20% more efficient.
The only other time he even shows the power usage at all is near the beginning, running a 4B model which is trivially small for both systems.
So once again, I ask, what are the sources for your performance/watt claim?