10 grand for 256GB memory. Likely double that for 512GB, but won't be available or finalized until October. Thunderbolt 5 is highest bandwidth external IO available at 120Gb/s. 1.2TB/s claimed max internal memory bandwidth.
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.
I recently got a refurbished Strix Halo 128GB desktop for a steal ($2200) and I have a feeling it's going to be my last PC purchase for the next 3-5 years. Who knows how long this RAMpocalypse will last. My Lunar Lake 32GB laptop is already a couple years old, and it's not a speed demon but it ticks all the laptop boxes (portable, great battery).
About 20 years ago my dad bought me a $5k computer, it was future proof for about "5 years" before we had to upgrade its internal parts (more memory, new graphics card).
It was future proof but not really because it struggled a lot in its final years.
He had that option,l. However buying apple hardware todqy you are stuck and need to replace all of it even if for example the CPU is plenty fast but you need more memory or GPU power.
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.
Think it would have same memory bandwidth if the RAM was upgradeable?
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.
It isn't integrated into the motherboard, it's integrated onto the same package as the CPU/GPU which allows for better signal integrity and higher speeds. You can get somewhat close to the same speeds while modular with tech like LPCAMM2, but there are some pretty difficult challenges to overcome to close the gap completely. As an example, the Framework Laptop 13 Pro CPUs support up to 9600MT/s (same as M5), and Micron sells LPCAMM2 modules that can run at 8533MT/s, but the 13 Pro only officially supports 7467MT/s.
If it was upgradable, then yes, spending more on top of it every year would make it future proof, but that's not the point. It's that spending 10 grand doesn't get you a future proof computer today.
> The way the Apple M-series does ram that might be difficult to pull off.
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.
> 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.
That actually would be interesting - yes, the computer itself is basically a PCB, but it’s also wrapped in a couple pounds of aluminum, a power supply, cooling fans, and a few other things that don’t need to be consumables. Upgrade a mini or a studio by swapping the new board into the old case - yeah, you’re not saving much money, but you also don’t need to throw out the entire rest of the case, and you can ship the main board in the space of a couple CD cases.
It’s a very non-Apple thing to do, but it’d be pretty awesome if they did.
Same thing, I still have at home a desktop I've built when I was 17 (now I'm 30!) and still runs fine for day to day tasks, of course with Linux on it. It has 8Gb of RAM (DDR3), 512Gb SSD (not original of course because back then they did cost an astronomic price) and an i7 CPU that I've overclocked, an NVIDIA GPU with 2Gb of memory that still plays older games fine.
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.
The relevant comparison isn't one mac studio to one RTX 6000, it's a 24 channel DDR5 system, which also has ~1.2TB/s of memory bandwidth (or more when Xeon 6 compatible 8800mt/s memory becomes widely available), vastly higher prefill due to more CPU horsepower, orders of magnitude faster networking, can hook into GPU accelerators, can be upgraded etc. A baseline 384GB system from eg Puget is ~30K vs ~12K for the 256GB Mac Studio and you do get value for the money.
Prefill is gated by GPU compute, not CPU compute. A Epyc/Xeon CPU will never beat an M5 GPU in prefill. If you want something that can match M5 Ultra in prefill, you'd need to add beefy Nvidia GPU. However, the problem is that the GPU memory is separate from the 384GB system memory. And here lies that advantage of Apple Silicon. It's unified memory and accessible by the CPU and GPU.
It can be more than "slightly", particularly if the model you're interested in (or will be interested in in 6 months) doesn't fit on the mac studio. You also need to account for eg storing 10TB of random checkpoints, load time when experimenting, and so on. When you start actually needing throughput these are all capability gaps in practical use, not just x% benchmark differences.
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.
A 3x price difference means you can get 3 256GB mac studios which you can connect through thunderbolt and with RDMA a total of 768GB ram with compute/memory bandwidth scaling basically linearly (with a small overhead cost).
Excellent post. Heck yes. And with MRDIMMs coming, we're going to get another >50% boost in throughput per channel real soon, with a massive uptick in max capacity (4x).
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.
How's the compute side now, I wonder? Because while the Ultras have impressive memory bandwidth for inference, processing prompts still takes a dog's age on my M3 Ultra. I heard the M5 makes some strides forward in this area, though, and the M7 in particular promises to go a lot further.
M5 is excellent, they’ve finally gotten their own tensor cores.
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've been using Macs as my primary machine for over three decades – but in the past year a Microcenter-Prebuilt (with a 5070Ti) has helped make Ubuntu my preferable computing environment (I have three Silicon Macs with similar access to 16GB vRAM – the 5070Ti "smokes" them in compute).
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").
Except the RTX 6000 will run circles around the Mac studio in just about every way. Memory bandwidth is literally the only spec where Apple is competitive, and while high memory bandwidth is necessary for LLMs to perform well, many people strangely don't understand that memory bandwidth alone is not sufficient.
> Mac studio wins in memory capacity, price, perf/watt and value.
[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.
> "if your model can fit into the VRAM" can be true for the Mac as well.
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.
> 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.
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).
Sharding a dense model using Tensor Parallelism (TP) across dual RTX 5090s has a significantly worse performance penalty over PCIe than sharding an MoE model.
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).
Have you ever actually set up a multi-GPU system for inference? Based on my experience you are drastically overstating the problem. Both tensor and pipeline parallelism (without NVLink) produce a machine which is faster than any Mac on the planet, which is what we’re discussing here. Yes, each has pros and cons, and neither scales perfectly linearly. But it works great regardless.
You can parallelize inference with much cheaper GPUs as well!
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.
You made claims about performance per watt and other metrics which were completely unsubstantiated and aren't backed up by the line you quoted. That's what I was asking for citations about.
Yes, I agree. What about the other metrics, including performance per watt which I’ve now mentioned four times and you’ve ignored the previous three times?
Glad you agree. I was just confused why you were questioning it. The big biggest advantage for Apple Silicon is that you can get much more VRAM per $ over Nvidia cards. No controversy.
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?
I’ve run the numbers on this, and an optimized nVidia build can meet or beat Apple platforms in terms of tokens per watt, which is the most important efficiency metric if what you care about is using the least amount of energy to generate a given response.
Yes, the Mac might get lukewarm, but it will take 2-3+ times longer to do the same task.
Yes I agree that at full speed with parallel workloads, the tokens per watt are better using Nvidia GPU servers.
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.
Okay, so you said “Energy efficiency is in another league with the Mac Studio”; I assume you meant that it was more efficient than an nVidia GPU setup. But when pressed, you agreed that’s not actually the case. So it’s not actually that much different in terms of energy efficiency.
> 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.
I was not speaking about specific models so I didn’t feel the need to cite any. Not sure why the backhanded insult was necessary. You need multiple Mac Studios to run the largest models as well, so neither is a one-size-fits-all device.
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.
That's in the same ballpark as two 128GB AI machines like the Asus GX10 or DGX Spark or Strix Halo. And, it seems very likely to perform better than either of those for inference. And, 256GB brings some pretty good models into play.
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.
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.
No one making statements has an incentive to tell the truth though (in fact, they're all incentivised to keep proclaiming the RAMpocalypse will never end).
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.
There was an article last week that a Chinese RAM manufacturer was planning to add new fabs to be able to ramp up. More or less simultaneously there was also an article on Apple considering switching to using china sourced RAM for China destined devices.
There are signs that the money faucet is being turned down. Various investments that were announced have been quietly canceled or reduced in scale. I don't think it'll happen soon, but it seems unlikely to be more than a couple years. If I were a betting man, 12-18 months seems right. The AI companies that can make enough money will survive, the ones running on investor cash and debt, won't.
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 are some news about AI/LLM progress rate flattening out: People uses the cheaper model more than better model. Some AI startup in Chinese lose half of value.
Yes, if I was going to throw away ten grand on a computer that can run models that are much worse than what I can rent from a variety of providers for a few bucks a month, I'd buy the Apple.
I didn't mention DeepSeek, though with the caching effectiveness in Reasonix, you can get DeepSeek for pretty danged cheap directly from deepseek.com, even at the higher prices. I was referring to the many subscriptions available that will sell you quite a lot of usage for $20/month. Given how slow local inference is, you're probably not going to ever get your money back out of it if you just use it for inference (or training/fine-tuning, since you can rent GPUs for a buck or two an hour and ten grand is a lot of hours).
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.
I don't think the $20/month packages come with API access at all.
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.
In computing "Pro" is not the opposite of "Consumer".
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.
Apple's use of "Professional" is just a fancy alias for high end or premium, it in no way indicates anything about being "for professionals" (most extreme example: "professional" iPhone models)
Hence why they had to make up the "Studio" brand for the workstation market, because they'd already fully removed any meaning from "Professional"
It's unclear to me how bandwidth scales with multiple connections. Many-to-many does not seem ideal. Daisy chaining would be fine for straight pipeline work. There doesn't seem to be an equivalent of a ethernet switch for thunderbolt 5 though.
It will always be wasteful to have a GPU sleeping next to you 99% of the time. Something like OpenRouter is the solution imho, for me at least. I realize that that some people care about privacy more though, and I respect that.
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.