512GB unified memory is absolutely wild for AI stuff! Compared to how many NVIDIA GPUs you would need, the pricing looks almost reasonable.
A server with 512GB of high-bandwidth GPU addressable RAM in a server is probably a six figure expenditure. If memory is your constrain, this is absolutely the server for you. (sorry, should have specified that the NPU and GPU cores need to access that ram and have reasonable performance). I specified it above, but people didn't read that :-)
What is the memory bandwidth to the CPU cores? Is it competitive with 8-channel DDR5 servers for non-GPU compute?
How do people feel about the value of the M3 Ultra vs. the M4 Max for general computing, assuming that you max out the RAM on the M4 version of the Studio?
The kinds of workloads that could truly leverage the M2 Ultra over the M2 Max were vanishingly small. When comparing the M3 Ultra to the M4 Max, that number gets even smaller, because the M4 Max will have ~15% higher single core perf. The insane memory available on M3 Ultra is its only interesting capability, but its still not big enough to run the series of largest open source LLMs.
Hot take: You can tie yourself into six knots trying to spin a yarn about why the M3 Ultra spec is super awesome for some AI use-case, meanwhile you could buy a Mac Mini and like 200 million GPT-4o tokens for the cost of this machine that can't even run R1.
Since no one specifically answered your question yet, yes, you should be able to get usable performance. A Q4_K_M GGUF of DeepSeek-R1 is 404GB. This is a 671B MoE that "only" has 37B activations per pass. You'd probably expect in the ballpark of 20-30 tok/s (depends on how much actually MBW can be utilized) for text generation. From my napkin math, the M3 Ultra TFLOPs is still relatively low (around 43 FP16 TFLOPs?),…
I actually think it’s not a coincidence and they specifically built this M3 Ultra for DeepSeek R1 4-bit. They also highlight in their press release that they tested it with 600B class LLMs (DeepSeek R1 without referring to it by name). And they specifically did not stop at 256 GB RAM to make this happen. Maybe I’m reading too much into it.
What people who did not actually work with this stuff in practice don't realize is the above statement only holds for batch size 1, sequence size 1. For processing the prompt you will need to read all the weights (which isn't a problem, because prefill is compute-bound, which, in turn is a problem on a weak machine like this Mac or an "EPYC build" someone else mentioned). Even for inference, batch size greater than 1…
Sure, nuance. This is why Apple makes so much fucking money: people will craft the wildest narratives about how they’re going to use this thing. It’s part of the aesthetics of spending $10,000. For every person who wants a solution to the problem of running a 400b+ parameter neural network, there are 19 who actually want an exciting experience of buying something, which is what Apple really makes. It has more in comm…
Pretty much. In addition, PyTorch on the Mac is abysmally bad. As is Jax. Idk why Apple doesn't implement proper support, seems important. There's MLX which is pretty good, but you can't really port the entire ecosystem of other packages to MLX this far along in the game. Apple's best bet to credibly sell this as "AI hardware" is to make PyTorch support on the Mac excellent. Right now, as far as AI workloads are concerned, this is only suitable for Ollama.
True. But with Project Digits supposedly around the corner, which supposedly costs $3,000 and supports ConnectX and runs Blackwell; what's the over-under on just buying two of those at about half the price of one maxed M3 Ultra Mac Studio?
And how much VRAM will Project Digits have?
128gb each, so two would have 256gb.
Its half that of a max spec Mac Studio, but also half the price and eight times faster memory speed. Realistically which open source LLMs does 512gb over 256gb of memory unlock? My understanding is that the true bleeding edge ones like R1 won't even handle 512gb well, especially with the anemic memory speed.
Two questions for the fellow HNers: 1. What are various average joe (as opposed to researchers, etc.) use cases for running powerful AI models locally vs. just using cloud AI. Privacy of course is a benefit, but it by itself may not justify upgrades for an average user. Or are we expecting that new innovation will lead to much more proliferation of AI and use cases that will make running locally more feasible? 2. Wit…
Risk of shut-out.
I'm in Hong Kong, I can't even subscribe to OpenAI or Claude directly, though granted this doesn't so much apply to the already "open" models
128gb each, so two would have 256gb. Its half that of a max spec Mac Studio, but also half the price and eight times faster memory speed. Realistically which open source LLMs does 512gb over 256gb of memory unlock? My understanding is that the true bleeding edge ones like R1 won't even handle 512gb well, especially with the anemic memory speed.
We really should see what happens when Project Digits is finally released. Also, I would love in NVIDIA decided to get in the CPU/GPU + unified memory space.
I can't imagine the M3 Ultra doing well on a model that loads into ~500G, but they should be a blast on 70b models (well, twice as fast as my M3 Max at least) or even a heavily quantized 400b model.
Computers these days - the more appealing, exciting, cooler desirable, the higher the price, into the stratosphere. $9499 What ever happening to competition in computing? Computing hardware competition used to be cut throat, drop dead, knife fight, last man standing brutally competitive. Now it's just a massive gold rush cash grab.
You're thinking about this as if the average Joe is interested in this. They're not. The tech folk salivating in this discussion are a rounding error when it comes to computing. The vast majority has no need for this.