[HN Gopher] Apple unveils new Mac Studio
___________________________________________________________________
Apple unveils new Mac Studio
Author : ksec
Score : 273 points
Date : 2025-03-05 14:00 UTC (9 hours ago)
(HTM) web link (www.apple.com)
(TXT) w3m dump (www.apple.com)
| bananapub wrote:
| in case people are unaware, this is more exciting than other
| random computer updates, since the Mac Studio is probably the
| best system in the world for running LLMs, as it can come with >>
| 100GB of ram and _that 's all accessible to the graphics/neural
| accelerator_ at high speed.
|
| this new one comes with up to 512GB of unified RAM!
| bryanlarsen wrote:
| Or you could wait a couple months and get a Strix Halo based
| desktop from HP, Framework, Asus or GMK for > 100GB of unified
| memory.
| noelwelsh wrote:
| Doesn't Strix Halo top out at 192GB RAM? The new M3 Ultra
| seems a lot more powerful on paper.
| Scramblejams wrote:
| I believe it tops out at 128, of which 96 can be used as
| GPU VRAM. Hoping AMD will open the memory floodgates on the
| next rev.
| kccqzy wrote:
| I'm not aware that Strix Halo has shipped on desktop. I
| thought it was only shipped on laptops and tablets? And
| Framework announced a desktop but did not ship it yet.
| bryanlarsen wrote:
| It hasn't shipped anywhere yet. You can find previews of
| Strix Halo on laptops and tablets, but you can't buy them
| yet.
| short_sells_poo wrote:
| It has terrible compute performance...
| blobbers wrote:
| Better than cloud API calls?
| jazzyjackson wrote:
| Have you used a 100GB model on a Mac studio? Tokens per second
| is single digit, I didn't find it usable at all, found myself
| going back to cloud APIs where 3000$ goes a much longer way
|
| I'm looking forward to trying Nvidia's little set top box if it
| actually ships, should have higher memory bandwidth, but still
| Ill probably set up a system where I email a query with
| attachments and just let DeepSeek email me back once it's
| finished with reasoning at 10T/s
| clonky wrote:
| It might blow your mind that you can run a quantized
| DeepSeek-R1 (671B) at over 15 t/s on an M2 Ultra 192GB and
| still get around 9000 context.
| lenerdenator wrote:
| > 512GB of RAM
|
| Keep these things the hell away from the people who develop
| Chrome and desktop JS apps.
| whizzter wrote:
| In 2025 the question isn't "will it run crysis", it's "will it
| run a simple CRUD app".
| esafak wrote:
| Will it run Electron?
| bloomingkales wrote:
| Gonna pile on:
|
| At this point we may need TSMC to make a specialized chip
| to run Electron.
| zitterbewegung wrote:
| It has been done for Java [1] and as we make smaller and
| smaller chips who knows.
|
| [1] https://en.wikipedia.org/wiki/Jazelle
| unilynx wrote:
| ARM already has them. https://developer.arm.com/documenta
| tion/ddi0602/2022-03/SIMD...
| therein wrote:
| This was discussed before and interesting but apparently
| the name of that instruction is misleading. Someone had
| chimed in and talked about how having Javascript in its
| name is entirely unnecessary as that exact same floating
| point representation is commonly used outside Javascript
| as well.
|
| If you disassemble some armv8 binaries that aren't
| dealing with Javascript, you do see still see FJCVTZS.
| vvillena wrote:
| There are already specialized instructions in the Apple
| Silicon chips. IIRC there's something tailored for the
| Objective-C runtime, and something useful for Javascript
| runtimes.
| favorited wrote:
| Uncontended acquire-release atomic operations are
| basically free on Apple Silicon, which synergizes with
| the Objective-C (and Swift!) runtimes, where every
| retain/release is an atomic increment/decrement.
|
| https://web.archive.org/web/20201119143547/https://twitte
| r.c...
| belter wrote:
| You need an AWS Region for that...
| Maken wrote:
| Will it run Discord?
| gjsman-1000 wrote:
| In the future, we'll decide HTML, CSS, and JS are too much of
| an inconsistent burden; so every website will bundle their
| own renderers into a <canvas> tag running off a WASM blob.
| Accessibility will be figured out later - just like it was
| for the early JavaScript frontends.
|
| I am looking forward to the HTML Frameworks explosion. You
| thought there were too many JS options? Imagine when anyone
| can fork HTML.
| jsheard wrote:
| Why stop there? LLMs will free us from the shackles of
| having to ship actual code, instead we'll ship natural
| language descriptions and JIT them at runtime. It may use
| orders of magnitude more resources and only work half of
| the time but imagine the _Developer Velocity(tm)_
| gjsman-1000 wrote:
| In fact, with AI becoming more powerful, the <canvas> tag
| might soon become even more viable; because nobody will
| need ARIA tags or similar to tell them what's on screen.
| The AI screen reader will look at the website as a whole
| and talk to the user. With accessibility no longer
| required, and with any UI being just a dumb framebuffer,
| we'll finally see perfect chaos.
| lynx97 wrote:
| And blind people will be the first test subjects for the
| "we see everything you read" project. Sweet. A small
| enough group that has no way out. Besides, after the
| initial giveaways, imagine the revenue if you can charge
| for every single pageview.
| Zekio wrote:
| can't wait for all the imaginary features
| tempodox wrote:
| You can hallucinate them right now already. Just ask
| WebGPT.
| hnthrow90348765 wrote:
| The LLM created code will then be consumed by my AI agent
| which will rewrite the application to filter all of the
| bullshit and be fit for my minimalist preferences like a
| Reader Mode for CRUD apps.
| swiftcoder wrote:
| You jest, but isn't this Web Components? Or alternately,
| Flutter
| gjsman-1000 wrote:
| Web Components was too verbose and nobody could figure it
| out. Flutter is just the beginning of the newest scheme
| by RAM manufacturers to bloat our RAM usage. We've
| stagnated at 8GB on midrange computers for too long.
| prisenco wrote:
| Web Components aren't that bad, but they could definitely
| use a DX makeover.
|
| For simple components, I much prefer them to firing up
| the React ecosystem.
| asdajksah2123 wrote:
| > every website will bundle their own renderers into a
| <canvas> tag running off a WASM blob
|
| Isn't that Flutter?
| threatofrain wrote:
| Soon it'll be all 3D content anyway... the old world of a
| graph of documents is going away. The web breathed a sigh
| of relief when Apple's Vision Pro bombed.
| catapart wrote:
| Not that I intend to scale this in any way, but I'm working
| on an in-game UI rendered on the canvas, and I am thinking
| I might be able to hack something together based on this
| youtuber's library and excellent explainer video[0]. The
| thought had definitely occurred to me that if someone
| wanted to really roll up their sleeves and maintain a js
| port of the library, it would provide a translate-able UI
| from native C to native JS and back. At least, I can
| imagine a vite/webpack-like cli that reads the C
| implementation and spits out a js implementation.
|
| Of course, I could also imagine one that reads the C and
| provides the equivalent html/css/js. And others might scoff
| "why not just compile the whole C app into wasm", which
| would certainly be plenty performant in a lot of cases. So
| I guess I don't know why it isn't already being done, and
| that usually means I don't know enough about the problems
| to have any clue what it would actually take to make such
| things.
|
| In any case, I'm also looking forward to a quantum leap in
| web app UI! I'm not quite as optimistic that it's ever
| going to happen, I guess, but I can see a lot of benefit,
| if it did.
|
| [0]https://www.youtube.com/watch?v=by9lQvpvMIc
| peatmoss wrote:
| The state of web deployment in 2025 is the universe
| punishing me for calling java applets and other java web
| deployment tech "heavyweight" back in the day.
| fumar wrote:
| What would we dev areas are important to you?
| fumar wrote:
| I'm thinking about this space now. Ideally, I want a new
| browser like platform with stricter security properties
| than browsers but better rendering out of the box
| capabilities.
| lynx97 wrote:
| <canvas> is already a middle finger in the direction of
| accessibility. You don't need wasm to put blind people at
| an extra disadvantage. SVG Accessibility anyone? No? What a
| surprise. Classical web accessibility has basically ended.
| We (blind people) are only using sites which are
| sufficiently old to be still usable.
| rikroots wrote:
| I'm genuinely trying to do something about <canvas>
| element accessibility. Whether it's enough ...? Probably
| not. But if I can do the work to try and show that
| <canvas> elements can be made more accessible, then
| there's no excuse for developers working on far more
| popular JS canvas libraries from making an attempt to
| better my efforts.
|
| I do strongly agree that <canvas> elements should not be
| used to replace HTML/CSS! My personal web hierarchy is 1.
| HTML/CSS/images; 2. Add (accessibility-friendly) JS if
| some fancy interaction is useful; 3. More complex - try
| SVG/CSS; 4. use <canvas> only if nothing else meets the
| project requirements.
|
| https://github.com/KaliedaRik/Scrawl-canvas
| cellularmitosis wrote:
| Speaking of CRUD, would Apple's on-chip memory have
| significant advantages for running Postgres vs a threadripper
| with a mobo full of ram?
|
| It seems like vertical scaling has fallen out of fashion
| lately, and I'm curious if this might be the new high-water
| mark for "everything in one giant DB".
| ffsm8 wrote:
| I'm not sure how this would impact the server market in any
| way considering that epyc thread ripper has supported 4 TB
| for over 5 yrs now.
|
| Is it the usual Apple distortion effect where fanboys just
| can't help themselves?
|
| It's definitely a sizeable amount of RAM though, and
| definitely enough to run the majority of websites running
| on the web. But so would a budget Linux server costing
| maybe 100-200 bucks per month.
| Moto7451 wrote:
| The question is about embedded DRAM, not trying to put a
| Mac in the data center. I am unaware of an apples to
| apples comparison here, but on the same Intel and AMD
| platform there can be a performance increase associated
| with embedded high speed LPDDR5 vs something on an
| SODIMM, which is why CAMM is being developed for that
| space.
|
| I would be interested as well in what an on chip memory
| bank would do for an EPYC or similar system since exotic
| high performance systems are fun even if all I'll ever
| touch at this point is commodity stuff on AWS and GCP.
| ffsm8 wrote:
| He edited his comment. The previous version did reference
| the 512 GB being so big that it'd be a game changer for
| servers.
| jmb99 wrote:
| Yeah, 512GB was a game changer for servers... with
| DDR3...
|
| And that wasn't even where it topped out, there were
| servers supporting 6TB of DDR3 in a single machine. DDR4
| had at least 12TB in a single (quad-CPU) machine that I
| know of (not sure if there were any 96*256GB DDR4
| configs). These days, if money's no object, there exist
| boards supporting 24TB of DDR5. I think even some quad-
| CPU DDR2-era SKUs could do 1.5TB. 512GB is nothing.
|
| (Not directly in response to you, just adding context.)
| hot_gril wrote:
| There are already big servers designed for huge single
| databases, for example the 8-socket Xeon types. Tbh I don't
| understand exactly why RAM is such a concern, but these
| machines have TBs of it.
| Waterluvian wrote:
| Something I've been surprised to find over the years working at
| software companies is just how many C++ writing, Linux using
| senior engineers there are who simply do not understand how
| allocation works and what HTOP is actually telling them.
|
| I really think a sizable chunk of people in the "omg my RAM!"
| camp are basing it on vibes, backed up by a misread of reported
| usage.
|
| This reminds a long time ago when I was trying to figure out
| why the heck my Intel Mac was allocating all my RAM and most of
| my swap to Preview or Chess.
| klik99 wrote:
| I'm surprised how many people bring up "Erm that memory is
| actually not being used" as if there isn't plenty of knock on
| effects from how memory pressure is actually done. For
| example, if I keep chrome open long enough my builds slowly
| use less and less threads because the build thinks there's
| less memory available, so I have to periodically close chrome
| and reopen and restore last session.
|
| It's true reported memory allocation does not equal actual
| memory used and that's very clever of everyone who brings it
| up, but it does actually cause real annoyances.
| NoMoreNicksLeft wrote:
| >so I have to periodically close chrome and reopen and
| restore last session.
|
| I thought ublock was forced out of Chrome months ago... how
| are you people still using it? I switched back to Firefox a
| couple years ago already, even if it's occasionally
| painful.
| klik99 wrote:
| Yeah I actually use Firefox almost exclusively, I only
| use Chrome when websites mysteriously don't work in
| firefox, 99% of the it's lack of firefox support.
|
| But in my example I was thinking of a particular 2-month
| stretch where this kept biting me and I was using Chrome
| at that point. In terms of memory usage, Firefox is no
| better though (at one point it was, but not any more).
|
| Now I'm afraid of saying "memory usage" lest someone pops
| out to comment "that's not how memory works" like whack-
| a-mole.
| recursive wrote:
| Don't visit sites with ads.
| cmrdporcupine wrote:
| Top-ish utilities should just be preconfigured to only show
| RSS unless you absolutely need to know what's virt. A lot of
| griping would diminish.
|
| There are many specialized allocation patterns -- especially
| for larger system things like DBs, virtual machine / runtimes
| etc. -- that will mmap large regions and then only actually
| use part of it. Many angry fingers get pointed often without
| justification.
| klik99 wrote:
| I think the griping occurs when things actually slow down,
| or system perceived available memory cripples resources.
| Maybe they point the finger to the wrong place due to
| misreading virtual, but I doubt people are getting angry
| when their system is running smoothly.
|
| And this attitude "oh memory usage problems are a
| misreading of top" promotes poor memory management hygiene
| - and I think there's a strong argument that's all good in
| server applications / controlled environments but for
| desktop environments this attitude causes all sorts of
| knock on effects.
| masfuerte wrote:
| Maybe we're all idiots who look at some irrelevant numbers
| and declare the sky is falling. Or maybe we notice everything
| running really slowly because the computer is constantly
| paging.
| kees99 wrote:
| Chrome has to run on chromebooks, quite a few of which are
| still-supported models with 4GB of non-upgradeable RAM.
| ant6n wrote:
| These chromebooks won't run chrome, they'll meander it.
| lenerdenator wrote:
| I wouldn't even call it meandering.
|
| Know that scene from one episode of Aqua Teen Hunger Force
| where George Lowe (RIP) is a police officer and has his
| feet amputated, so he drags himself while pursuing a
| suspect?
|
| Yeah. It does that.
| superjan wrote:
| So that means it can run with 4GB. Is there a way to block it
| from using more?
| Mistletoe wrote:
| Who do you think buys these? :)
| doublerabbit wrote:
| Renderfarms Animation Studios
|
| We had some hefty rigs at the last studio I worked at.
| nicce wrote:
| Are these really cost effective for that usecase?
| whywhywhywhy wrote:
| Not really. Smaller scenes you'd use nvidia GPU, larger
| scenes you'd probably save money doing normal servers.
| yjftsjthsd-h wrote:
| You run a render farm made of macs?
| doublerabbit wrote:
| Did. Until I locked all the artists out of their work
| accidentally when I went one lunch break. Screwed up a Fw
| config.
|
| The old xeon stations were power houses.
| layer8 wrote:
| They should make a "webdev" edition with like 4 GB.
| Hamuko wrote:
| The Mac Mini M2.
| DwnVoteHoneyPot wrote:
| Raspberry Pi 4GB
| singularity2001 wrote:
| 512GB only available on M3
| reustle wrote:
| That's almost the full deepseek r1!
| seunosewa wrote:
| Almost is a painful word in this case. Imagine if it could
| actually run R1. They'd make so much money. Falling short by
| a few dozen GB is such a shame.
| asah wrote:
| $10k and up
| WorldWideWebb wrote:
| So they wouldn't put the power button on the back of the latest
| Mini, but they did on the Studio? That's frustrating (yes, minor
| nit).
| jonnrb wrote:
| Power buttons are for power users. lol
| bigtex wrote:
| You are holding it wrong - Steve Jobs
| zitterbewegung wrote:
| This was always part of the original design of the Mac Studio
| so they have never changed the design. This is a spec bump.
| FloatArtifact wrote:
| They didn't increase the memory bandwidth. You can get the same
| memory bandwidth, which is available on the M2 Studio. Yes, yes,
| of course you can get 512 gigabytes of uRAM for 10 grand.
|
| The the question is if a llm will run with usable performance at
| that scale? The point is there's diminishing returns despite
| having enough uRAM with the same amount of memory bandwidth even
| with increased processing speed of the new chip m3 for AI.
| wkat4242 wrote:
| True, I also noticed that bigger models run slower at the same
| memory bandwidth (makes sense).
| memhole wrote:
| Yeah, I don't think RAM is the bottleneck. Which is
| unfortunate. It feels like a missed opportunity for them. I
| think Apple partly became popular because it enabled creatives
| and developers.
| throw-qqqqq wrote:
| > I don't think RAM is the bottleneck
|
| Not the size/amount, but the memory bandwidth usually is.
| espadrine wrote:
| > _if a llm will run with usable performance at that scale?_
|
| Yes.
|
| The reason: MoE. They are able to run at a good speed because
| they don't load all of the weights into the GPU cores.
|
| For instance, DeepSeek R1 uses 404 GB in Q4 quantization[0],
| containing 256 experts of which 8 are routed to[1] (very
| roughly 13 GB per forward pass). With a memory bandwidth of 800
| GB/s[3], the Mac Studio will be able to output 800/13 = 62
| tokens per second.
|
| [0]: https://ollama.com/library/deepseek-r1
|
| [1]: https://arxiv.org/pdf/2412.19437
|
| [2]: https://www.apple.com/newsroom/2025/03/apple-unveils-new-
| mac...
| fullstackchris wrote:
| You seem like you know what you are talking about... mind if
| I ask what your thoughts on quantization are? Its unclear to
| me if quantization affects quality... I feel like I've heard
| yes and no arguments
| sosuke wrote:
| I don't know what I'm talking about but when I first asked
| your question this https://gist.github.com/Artefact2/b5f810
| 600771265fc1e3944228... helped start me on a path to
| understanding. I think.
|
| But if you don't already know the question your asking is
| not at all something I could distill down into a sentence
| or to that would make sense to a lay-person. Even then I
| know I couldn't distill it at all sorry.
|
| Edit: I found this link I referenced above on quantized
| models by bartowski on huggingface https://huggingface.co/b
| artowski/Qwen2.5-Coder-14B-GGUF#whic...
| espadrine wrote:
| There is no question that quantization degrades quality.
| The GGUF R1 uses Q4_K_M, which, on Llama-3-8B, increases
| the perplexity by 0.18[0]. Many plots show increasing
| degradation as you quantize more[1].
|
| That said, it is possible to train a model in a
| quantization-aware way[2][3], which improves the quality a
| bit, although not higher than the raw model.
|
| Also, a loss in quality may not be perceptible in a
| specific use-case. Famously LMArena.ai tested Llama 3.1
| 405B with bf16 and fp8, and the latter was only 2 Elo
| points below, well within measurement error.
|
| [0]: https://github.com/ggml-
| org/llama.cpp/blob/master/examples/q...
|
| [1]: https://github.com/ggml-
| org/llama.cpp/discussions/5063#discu...
|
| [2]: https://pytorch.org/blog/quantization-aware-training/
|
| [3]: https://mistral.ai/news/ministraux
| _aavaa_ wrote:
| This doesn't sound correct.
|
| You don't know which expert you'll need for each layer, so
| you either keep them all loaded in memory or stream them from
| disk
| kgwgk wrote:
| Note that 404 < 512
| espadrine wrote:
| In RAM, yes. But if you compute an activation, you need to
| load the weights from RAM to the GPU core.
| _aavaa_ wrote:
| Got you, yeah I misread you commend the first time around
| jazzyjackson wrote:
| I returned an M2 Max studio with 96GB RAM, unquantized llama
| 70B 3.1 was dog slow, not an interactive pace. I'm interested
| in offline LLM but couldn't see how it was going to produce
| $3,000 ROI.
| FloatArtifact wrote:
| It would be really cool if there was awebsite "we there yet"
| for reasonable offline AI.
|
| It could track different hardware configurations and
| reasonably standardized benchmark performance per model. I
| know there's benchmarks buried in github Llama repository.
| robbomacrae wrote:
| There seems to be a LOT of interest in such a site in the
| comments here. There seem to be multiple IP issues with
| sharing your code repo with an online service so I feel a
| lot of folks are waiting for the hardware to make this
| possible.
|
| We need a SWE-bench for open source LLM's and for each
| model to have 3Dmark like benchmarks on various hardware
| setups.
|
| I did find this which seems very helpful but is missing the
| latest models and hardware options.
| https://kamilstanuch.github.io/LLM-token-generation-
| simulato...
| slama wrote:
| The M3 Ultra is the only configuration that supports 512GB and
| it has memory bandwidth of 819GB/s.
| shrx wrote:
| Why don't they provide performance comparisons between the two
| chips offered, M3 Ultra and M4 Max?
| relium wrote:
| It'll likely be very workload dependent. The M4 Max will
| probably do a little better in single threaded tasks like
| browser benchmarks and the M3 Ultra will do better in things
| like video transcoding and 3D rendering.
| shrx wrote:
| Yes but I'd still like to know what tradeoffs I am making
| when deciding to get one or the other option. Right now it's
| all hand-wavy.
| cyberlimerence wrote:
| What model can you run realistically with 512 GB of unified
| memory ? I'm curious who is even the market for such an offering.
| wkat4242 wrote:
| DeepSeek R1 for one, quantised but not too cripplingly.
| numpad0 wrote:
| the full R1 takes >512GB and the 1.52bit takes >128GB. So
| enough for agent + app to realize a fully autonomous
| monolithic AGI humanoid head, potentially, but then it'll be
| compute limited...
| wkat4242 wrote:
| Yeah I was thinking more about q6_0 or so. The q4_K_M is
| 404GB so you can still push it a bit higher than that.
| Obviously the 1.52 bit doesn't make sense.
|
| I'm never going to pay 10k for that though. Hopefully
| cheaper hardware options are coming soon.
| saganus wrote:
| I assume they are getting ready for the next year or two of LLM
| development.
|
| Maybe there's not much market right now, but who knows if
| DeepkSeek R3 or whatever will need something like this.
|
| It would be awesome to be able to have a high-performance
| local-only coding assistant for example (or any other LLM
| apllication for that matter).
| mlboss wrote:
| The future is local AI agents running on your desktop 24x7.
| This and NVIDIA Digits will be the hardware to do that.
| petesergeant wrote:
| > featuring M4 Max and new M3 Ultra
|
| I hate that this naming shit has gotten so bad
| blobbers wrote:
| Agree. This ship is adrift.
|
| Remember when it was just a MacBook, or air, or pro and it had
| a year.
| whynotminot wrote:
| It's really not that complicated guys.
|
| It's definitely a little odd to have M3 Ultra > M4 Max, but I
| feel like anyone complaining about this must have never
| bought any other manufacturers' wares in their lives. Obtuse
| complication is kind of the norm in this industry.
| rsynnott wrote:
| Oh, but things were far worse back then, in terms of knowing
| what you were getting. For instance, let's say you bought a
| 13" MacBook Pro in 2016. Do you have a dual core or quad core
| processor? Depends on whether you have a touchbar!
|
| (For reasons best known to themselves, Apple made two
| completely different 13" MBPs that year, both new, with the
| loathed butterfly keyboard, weighing a different amount, with
| different processors, and the same name.)
| pram wrote:
| Literally not even the same thing you're comparing. There
| were more processor and graphics options. Want an i5, i7, or
| i9 and what about a RX 580, 5300, 5500, 5700, or 5700XT?
| caconym_ wrote:
| Has it? These names are basically (architecture version, amount
| of compute) tuples. Much better than what AMD, Intel, Nvidia,
| etc. are currently doing.
| InitialLastName wrote:
| Which is a bigger amount of compute, "max" or "ultra"?
| caconym_ wrote:
| This question is trivially answered by visiting Apple's
| website, which is---to my point---not generally true for
| their competition. If you have some further point to make,
| I recommend stating it more clearly so we can avoid wasting
| time here.
| alexjplant wrote:
| Why is the "Ultra" more powerful than the "Max"? I would
| expect "Max" to mean "maximum" in this context but it
| seems to mean "directly subordinate to that which is the
| maximum". This is pretty obviously a point of confusion.
| Just because other CPU manufacturers do goofy stuff with
| naming doesn't mean that Apple is exempt from criticism
| for doing something so obviously bereft of common sense.
| caconym_ wrote:
| Well then, it's a good thing I didn't say anywhere that
| they should be exempt from criticism. Incidentally, it's
| not something I believe. But I do think they are still
| doing a much better job at naming than other CPU
| manufacturers, despite the obnoxious Pro/Max/Ultra stuff.
|
| I suppose I commented here because I think people are
| letting their subjective distaste for those terms sway
| their opinion of a superior naming scheme.
| brianmurphy wrote:
| I would expect the M4 being a newer generation of the
| chip to be faster than a M3, but apparently that's not
| what Apple did.
| wpm wrote:
| Why would you expect that? Also, the M4 _is_ faster than
| the M3, but that doesn 't mean Apple couldn't or didn't
| want to fuse two of them together for an M4 Ultra.
|
| Is an Intel 10700K faster than a 12400F? The generations
| are different but the chips have vastly different
| capabilities and features.
|
| M4 is the _generation_. The modifier modifies the
| generation. M4 Pro is an M4 with some extra pizzaz. M4
| Max is an M4 with lots of extra pizzaz.
| ulbu wrote:
| ultra means beyond. so it can make sense if you choose
| to. but the choice does suck, imo.
| kllrnohj wrote:
| > This question is trivially answered by visiting Apple's
| website, which is---to my point---not generally true for
| their competition.
|
| wtf are you talking about? Intel, Nvidia, and AMD all
| absolutely have complete specs for their products readily
| available on their respective websites. Much, much more
| complete ones than Apple does as well.
| caconym_ wrote:
| I guarantee you that a quick scan of any of those
| companies' websites will not equip one with a useful
| general understanding of the naming scheme they use for
| their products, in the sense that one can see a product
| name and immediately know where it falls in their lineup
| and what workloads it's meant to handle. That is what
| "the fuck" I am talking about.
| hot_gril wrote:
| https://www.intel.com/content/www/us/en/processors/proces
| sor... is pretty clear
| hot_gril wrote:
| Same with Intel and others, but it's still annoying.
| seanmcdirmid wrote:
| An Ultra is two Maxes. A Max is two Pros.
| perfmode wrote:
| Ultra. Ultra consists of two Max chips operating as one.
| drdaeman wrote:
| The bit I found confusing is that I don't immediately
| understand what's more performant in various scenarios, M4
| Max or M3 Ultra. The former has higher architecture version
| but lower amount of compute, the latter is previous arch but
| more compute.
| caconym_ wrote:
| Fair, but it's not like other CPU/GPU manufacturer naming
| schemes give you that either. At least Apple's scheme
| clearly tells you e.g. "previous arch but more compute".
| hot_gril wrote:
| Different tiers within the same generation are much easier
| to compare than across gens. Apple could benchmark them all
| and name them accordingly, but even that's misleading
| because it can be workload-dependent.
| dialup_sounds wrote:
| If you think that's bad, have you looked at Intel and AMD's
| naming scheme lately?
| Insanity wrote:
| Or.. USB naming schemes..
| kps wrote:
| Consider, if you will, ARM7 and ARMv7.
| deadbabe wrote:
| So DeepSeek R1 running locally at like 4 tokens per second? Okay
| blobbers wrote:
| The previous ranking article said the M3 ultra was the most
| powerful chip ever.
|
| Mac ecosystem is starting to feel like the PC world. Just give me
| 3 options. Cheap, good, and expensive. Having to decide how many
| dedicated graphic cores for a teenagers laptop is impossible.
| bee_rider wrote:
| For chips, Ultra and Max are like their workstation chips or
| something, right? It seems expected that they should be a
| little more differentiated, they are specialist, aren't they?
| erickhill wrote:
| The way I think about it is if I buy a Max chip I'm getting
| the performance of the generation that will be released a
| year later now in the current form factor, and then some.
|
| For example, I got the M1 Max when it was new. A year later
| the M2 came out. Spec-wise, the M1 Max was still a bit better
| than the M2 Pro in many regards. To me, getting a Max buys
| you some future proofing if you or your company can afford it
| (and you need that kind of performance). I use the Max with a
| lot of video work, and it's been fantastic.
| fckgw wrote:
| They have that. On laptops they have the M4, M4 Pro and M4 Max.
| Cheap, good and expensive.
| eyelidlessness wrote:
| My buying strategy has been the same since they started
| soldering RAM: buy the lowest spec CPU/GPU they offer with the
| amount of RAM I will need (which all but once has always been
| the maximum RAM they offer, which unfortunately usually means
| also buying the max CPU/GPU).
| zitterbewegung wrote:
| If you are in college or school getting a MacBook Air would be
| best and the size of your screen is going to have a higher
| impact (13 to 15 inch) than the Dedicated Graphics cores. Would
| advise not getting an MacBook Pro.
| hot_gril wrote:
| For the teen's laptop, you can simply get the base model. Even
| a base M1 is more than fast enough.
| vr46 wrote:
| Apple's CPU nomenclature is getting more deliberately confusing
| than BMW's.
| el_benhameen wrote:
| Open AI's naming conventions are still the best-in-worst-in-
| class, though.
| vr46 wrote:
| When you say "conventions", do you actually mean, "insanity"?
| cxie wrote:
| The M3 Ultra with 512GB unified memory is a monster for AI
| development, while the new M4 MacBook Air makes AI features
| accessible to the mainstream. Apple's approach of building their
| own silicon shows massive dividends - they can optimize the
| entire stack from hardware to software for specific workloads
| like AI inference
| ssijak wrote:
| Is this AI written? :)
| mort96 wrote:
| If only it was built for workloads with value instead, I'm sick
| and tired of being unable to buy hardware that doesn't claim to
| have been made specifically for bullshit
| naikrovek wrote:
| AI isn't everything that it was hyped to be when Copilot
| first came out, but AI does indeed have valid uses. It's like
| any other tool in a toolkit.
| mort96 wrote:
| I disagree.
| hot_gril wrote:
| What is your workload that this machine helps with?
| geerlingguy wrote:
| The buzz is all around AI and unified memory... but after editing
| 4K content on an M4 mini (versus my M1 Max Mac Studio), I've
| realized the few-generations-newer media processing in the M4 is
| a huge boost over the M1.
|
| Coupled with the CPU just having more oomph, I ordered an M4 Max
| with 64 GB of RAM for my video/photo editing; I may be able to
| export H.265 content at 4K/high settings with greater-than-
| realtime performance...
|
| I'm a little sad that the AI narratives have taken over all
| discussion of mid-tier workstation-ish builds now.
| TylerE wrote:
| The one thing that has me a bit bummed with this is that the
| Ultra, which I had planned to upgrade to, is only an M3 not an
| M4. Bit disappointing after waiting this long.
| tromp wrote:
| Not all that disappointing considering that most of the
| performance improvement in M4 seems to come from increased
| power consumption. In some applications, M4 performs worse
| per watt than M3.
| Uehreka wrote:
| Yeah but if you're buying an Ultra you're probably more
| concerned with raw performance than perf-per-watt. These
| aren't exactly used in laptops.
| cosmic_cheese wrote:
| It feels a bit like we entered the "consumer grade workstation"
| era a while back when AMD started selling 16-core CPUs that
| will happily socket into run of the mill consumer motherboards
| and that continued with the higher end M-series SoCs.
|
| It really is cool to see. It's nice that that kind of
| horsepower isn't limited to the likes proper "big iron" like it
| once was and can even be reasonably be packaged into a laptop
| that is decent at being mobile and not an ungainly portable-on-
| a-technicality behemoth.
| perfmode wrote:
| Is the M4's media processing superior to the M3? Would the M3
| Ultra not perform as well on video editing?
| schainks wrote:
| The M3 Ultra is two M3 chips together on one die. In
| aggregate they should outperform an m4 max by quite a bit.
| perfmode wrote:
| I meant single-core performance.
| 2OEH8eoCRo0 wrote:
| I'm surprised you use Macs since you usually lean toward more
| open HW and FOSS.
| xnx wrote:
| Since Apple has always positioned itself as a tool for creatives,
| is it likely that the Mac Studio my be a good tool for AI video
| generation using open weight models like Alibaba Wan?
| fluidcruft wrote:
| At this point, having the power button not be on the bottom is a
| major selling point for me vs the annoying-as-hell mini.
| zie wrote:
| My thinking is, why would you ever turn it off? They go to
| sleep and wake up great and barely even sip power when on, let
| alone when asleep.
| fluidcruft wrote:
| It's a lab computer. You can tell people not to shut it off,
| but it's still always turned off when I try to use it. Could
| be being shut down via ITs management tools/policies for all
| I know.
| electriclove wrote:
| I've had the new Mini for a few months and can't recall having
| to use the power button.
|
| How often are you using the power button on your Mini? What is
| your use case?
| mort96 wrote:
| How do you turn it on?
|
| If you never turn off your computer, it makes sense that you
| never use the power button. But some people do turn their
| computers off, and for us, it's really useful to be able to
| turn them on again.
| dylan604 wrote:
| I'm still on a wired USB full sized keyboard from at least
| a decade ago, but didn't the newer keyboards see the return
| of the power button? Did I dream that?
| mort96 wrote:
| I did some quick googling before answering, and from what
| I could find, people are generally saying that you can't
| power on the Mac Mini in other ways than by pushing its
| power button.
|
| Even if you can power it on using a wired keyboard
| though, I'm certain that you can imagine people who
| prefer wireless keyboards but also turn their computer
| off.
| dylan604 wrote:
| I could have worded that more clearly. I wasn't disputing
| power button on the bottom is odd as much as thinking
| that Apple brought the keyboard power button back. Maybe
| it was the TouchID on the keyboard, but on their laptops
| that is also the power button, so possibly just an
| assumption on my part.
| mort96 wrote:
| Oh, no I think you were clear enough, it's probably me
| that wasn't clear. I tried to find evidence that people
| were able to power on their Mac Mini in ways other than
| via the power button on the machine, such as a button on
| a keyboard. I couldn't find that, everyone just said that
| the physical power button on the machine is the only way.
| 7e wrote:
| An idle Mac doesn't use much power. Why are you turning it
| off?
| mort96 wrote:
| I like to turn my desktop off at night. I don't need a
| better reason than that.
| randcraw wrote:
| I like my computer to be secure when I'm not using it.
| Powered down is secure.
| TylerE wrote:
| Why? Sleep/suspend on macs is incredibly good, and power
| usage rounds to zero.
| mort96 wrote:
| Because I like to turn off my desktop at night. I like to
| come back to a fresh start. I sometimes reboot just to
| get a clean slate.
| JumpCrisscross wrote:
| > _Because I like to turn off my desktop at night_
|
| Put simply, more people like the aesthetic of no visible
| power button than like the aesthetic of daily rebooting
| their computer.
|
| If I were you, and I really couldn't let go of that, I'd
| put the Mac in sleep and have it scripted to restart at
| _e.g._ 6AM each day. You get the best of both worlds.
| Feel like you have a "fresh" Mac every morning. Let it do
| its updates and whatnot behind the scenes.
| alpaca128 wrote:
| The power button was already invisible when it was on the
| back.
| fluidcruft wrote:
| It's a shared computer in a hospital used for research data
| management. Basically, every time I walk up to it to use it,
| it's turned off.
|
| Maybe Apple should remove power off from the UI menus if
| they're claiming it uses less energy to leave it on.
|
| (I'm dubious of that claim people are repeating here, but
| what the hell do I know I'm just a physicist. Reality
| distortion isn't my thing.)
| marci wrote:
| If you have a laptop, do you turn it off or just close the
| lid?
|
| The mini is probably less power hungry than the macbooks
| (less components). I have some wifi 5/ac routers that
| consume more power at idle (nothing connected to them) than
| apple laptops.
| dewey wrote:
| > How often are you using the power button on your Mini? What
| is your use case?
|
| Every single day, not by choice but because it's constantly
| waking up in sleep mode to do maintenance task then
| overheating and shutting down again. Something about macOS
| and Bluetooth devices not playing nice.
| yborg wrote:
| I think another tier needs to be added to the Maslow pyramid
| for this particular class of complaint. I have had to reboot
| the M4 Mini on my desk a number of times now and it takes less
| than 3 seconds to lift the corner an inch and depress the
| switch.
| singularity2001 wrote:
| 256GB, 512GB only available with M3 Ultra not M4 WTH
| layer8 wrote:
| The M3 Ultra is a higher-spec chip than the M4 Max.
| TylerE wrote:
| But this is the first time a machine offered with an Ultra
| has been a lower generation chip than the base Max.
| layer8 wrote:
| It may not be the last time:
| https://arstechnica.com/apple/2025/03/apple-
| announces-m3-ult...
| fadedsignal wrote:
| Why do they always compare their new-gen CPUs with M1??
| Insanity wrote:
| Guess they have usage statistics to know a lot of their users
| are still on M1.
|
| My partner and I are both still on M1 (our personal machines)
| and don't really see the need to upgrade.
| dylan604 wrote:
| Are you a lawyer? Why do you only ask questions to which you
| already know the answer?
|
| Snark aside in case you're seriously asking, it's a PR thing.
| Generation to generation might not show much differences in
| direct comparisons that makes the crowd ooh and ahh. The M1
| chip was the first Apple Silicone chip, so going back to their
| first one for basis of comparison provides for more oohs and
| ahhs. The charts look pretty this way too.
| altairprime wrote:
| For people who upgrade annually, comparisons don't matter.
|
| For people who are making a forced purchase, comparisons don't
| matter.
|
| For people who are content with what they've got, comparing
| with the oldest-popular market segment offers a clear statement
| of improvement and helps long-term users calibrate when to make
| their next purchase.
| TylerE wrote:
| Well, in this case it's totally relevant as the extant studios
| are all M1. They never got M2 or M3.
| mbrubeck wrote:
| M2 Mac Studios were released in June 2023.
| https://support.apple.com/en-us/111835
| OnionBlender wrote:
| Are there any good sites or channels that do performance
| comparisons of new Mac hardware?
| slt2021 wrote:
| It mentions AI, but Macs dont support CUDA.
|
| What people are using for LLMa on macs, is it ggml ?
| sampton wrote:
| ggml and pytorch support mps. aside from bleeding edge most ml
| workloads can run on mps these days.
| bri3d wrote:
| GGML has a Metal (Apple's GPU interface layer) backend, yes,
| using MPS (Metal Performance Shaders), which are pre-baked
| shaders provided by Apple in a way similar to cuDNN. This is
| probably the most popular method for large-scale inference with
| modern bleeding-edge models.
|
| There's also Apple CoreML, which is sort of like ONNX in that
| it provides a limited set of primitives but if you can compile
| your model into its format, it does good low-power edge
| inference using custom hardware (Neural Engine).
|
| Apple also provide PyTorch with MPS, as well as a bunch of
| research libraries for training / development (axlearn, which
| is built on JAX/XLA, for example).
|
| They also have a custom framework, Accelerate, which provides
| the usual linear algebra primitives using a custom matrix ISA
| (AMX), and on top of that, MLX, which is like fancy accelerated
| numpy with both Metal and AMX backends (and slower CPU backends
| for NEON and AVX).
|
| Overall, there's a lot you can do with AI on Apple Silicon.
| Apple are clearly investing heavily in the space and the tools
| are pretty good.
| erickhill wrote:
| People with the 2023 lattice Mac Pro be like... can I upgrade
| mine?
| silvestrov wrote:
| The SSD prices are insane.
|
| $400 to go from 1TB to 2GB.
|
| $307/TB to go from 1TB to 16TB.
|
| That is 3 times the Amazon prices:
| https://diskprices.com/?locale=us&condition=new&capacity=4-&...
| rsynnott wrote:
| Given that it's a desktop, most people should just get it with
| the default size and get an external thunderbolt NVMe disk.
| Only if you need >Thunderbolt 5 speeds (ie 80 Gbit/sec) do you
| really need the internal drive, and most NVMe is slower than
| this in any case.
| WWLink wrote:
| Or if you don't want ugly ass external boxes cluttering up
| your desk.
|
| I don't get why they couldn't be arsed to stuff a few m.2
| slots in there. They could keep the main nand their weird
| soldered on BS with the firmware stuffed in a special
| partition if they want. Just give us more room!
| outime wrote:
| You seriously don't get why?
| rsynnott wrote:
| https://en.wikipedia.org/wiki/Mac_Pro#Apple_silicon_(2023),
| kinda. The most ultra-niche of Apple's products.
| hot_gril wrote:
| And they still sell this, with an M2. Why?
| staplung wrote:
| I did this recently with a new Mac Mini that I set up. MacOS
| recently added the ability to locate the home directories on
| any volume. There's a somewhat hidden feature too that if you
| drag the Applications directory onto an external drive it
| will move selected apps there (the larger ones like Pages,
| etc.); combine that with the option in the App Store to keep
| large downloads on a separate disk.
|
| So far it's been working quite well with the exception that
| VSCode does not seem to understand how to update itself if
| you keep it in the external Applications folder: every time
| it tries to update itself it just deletes itself instead.
| Moved it back into the /Applications folder and it's been
| fine.
| rsynnott wrote:
| You can just use it as a boot drive, IIRC.
| wpm wrote:
| > MacOS recently added the ability to locate the home
| directories on any volume
|
| Mac OS has always been able to do this.
| zuhsetaqi wrote:
| Instead of dragging them over you should create a link.
| This way it's the same as before for Applications like VS
| Code.
| rafram wrote:
| Last I tried this, Spotlight didn't play well with
| symlinked application bundles.
| moralestapia wrote:
| This is also quite convenient when you buy a new laptop and
| just unplug/plug and that's it, you have everything.
| naikrovek wrote:
| the Studio doesn't use nvme but it does put its storage on a
| removable card. The mac mini does as well. So you don't have to
| pay Apple for the storage you want. There are places which sell
| storage upgrades for the Mini and the M1 Studio, and they, of
| course, are cheaper than what Apple charges for the upgrade
| when you buy the machine. dosdude1 on youtube has some videos
| of this exact upgrade, and a bit of googling will help you find
| vendors. I am assuming that this M3 and M4 Studio will be the
| same, but that's not a guarantee.
| canucker2016 wrote:
| One can upgrade the SSD storage for a M1/M2 Mac Studio through
| a third party for a lot less money than what Apple requires at
| purchase time.
|
| I'd expect an upgrade route for the new Mac Studio will appear.
|
| Here's one YouTube video showing an upgrade to 8TB of SSD
| storage. see https://www.youtube.com/watch?v=HDFCurB3-0Q
| ohgr wrote:
| Yeah they really need to get that under control. It's a
| complete rip off at this point.
|
| I don't mind them charging say $50 "Apple premium" for the fact
| it's a proprietary board and needs firmware loading onto the
| flash but the multiplicative pricing is bullshit price gouging
| and nothing more.
| flenserboy wrote:
| This seems like a much better deal than a blinged-out Mini.
| Interesting choices this generation.
| ein0p wrote:
| So basically the entire stash of M4 Ultra got used up in Apple
| Private Cloud, I guess.
| hart_russell wrote:
| $14,000 fully configured by the way
| sdf4j wrote:
| The first paragraph that talks about the OS itself is depressing:
|
| >macOS Sequoia completes the new Mac Studio experience with a
| host of exciting features, including iPhone Mirroring, which
| allows users to wirelessly interact with their iPhone, its apps,
| and notifications directly from their Mac.
|
| So that's their highlight for a pro workstation user.
| Nevermark wrote:
| Just be glad they didn't focus on movies, music and cute apps.
| Macs seems to be the only product line that continues to semi-
| dodge Apple's myopic media/services/social kiosk lens they now
| view all their other product lines through.
|
| If that sounds too negative, compare their current vision for
| their products with Steve Jobs old vision of "a bicycle for the
| mind". iOS-type devices are very useful, but unleashing new
| potential, enabling generational software innovation, just
| isn't their thing.
|
| (The Vision Pro is "just" another kiosk product for now, but it
| is hard to tell. The Mac support suggests they MIGHT get it.
| They should bifurcate:
|
| 1. A "Vision" can be the lower cost iOS type device, cool apps
| and movies product. Virtual Mac screen.
|
| 2. A future "Vision Pro that is a complete Mac replacement, the
| new high end Apple device, filled out spacial user interface
| for real work, etc. No development sandbox, Mx Ultra, top end
| resolution and field of view, raise the price, raise the price
| again, please. It could even do the reverse kind of support,
| power external screens that continued working like first class
| virtual screens, when you needed to share into the real world.
|
| The Vision Pro should become a maximum powered post-Mac device.
| Not another Mac satellite. Its user interface possibilities go
| far beyond what Mac/physical screens will ever do. The new
| nuclear powered bicycle for the mind. But I greatly fear they
| want to box it in, "iPad" everything, even the Mac someday.)
| rafram wrote:
| They use a similar line on the MacBook Air page. If you're
| buying an (up to) $13,000 Mac, hopefully you already understand
| macOS and its features, I guess.
| makz wrote:
| What LLMs are people running on this kind of setups?
| parsimo2010 wrote:
| I found this part about PCIe expansions interesting: "For those
| who rely on PCIe expansion cards for their workflows, Thunderbolt
| 5 allows users to connect an external expansion chassis with
| higher bandwidth and lower latency."
|
| (Maybe this is a feature that Apple has supported for a while,
| but I am unaware) Does this mean they will be officially
| supporting all PCIe devices like GPUs? Or do they only mean
| certain PCIe components like SSD expansions and network
| interfaces?
| shellac wrote:
| This has been a thing for quite a while. Thunderbolt is, in
| part, PCIe over serial (now USB C). There have been GPUs in
| external boxes, so it is undoubtedly possible, but I don't
| think they have many users.
|
| Edit: BlackMagic was what I was thinking of
| https://support.apple.com/en-gb/102103. 'Requires intel
| processor'
| jimnotgym wrote:
| Isn't it time Apple moved forward with its design aesthetic?
| Apple devices always used to look cutting edge. Now they just
| look samey
| bfrog wrote:
| They've reached peak design perhaps. The milled alloy bodies do
| have negatives but not many.
| jimnotgym wrote:
| That's an interesting take. A design outside of fashion. I
| can't see that being true myself. I want to see something as
| bold as an iMac again.
| speedylight wrote:
| If only it didn't cost as much as a used car to get max ram
| capacity I'd get one. It's so ridiculously small for how much
| power it has.
___________________________________________________________________
(page generated 2025-03-05 23:01 UTC)