I like to think back to 2011 and paraphrase what people were saying: "Is anyone seriously using gpu hardware to write nl translation software at the moment?"

"No, we should be use cheap commodity abundantly available cpus and orchestrate then behind cloud magic to write our nl translation apps"

or maybe "no we should build purpose built high performance computing hardware to write our nl translation apps"

Or perhaps in the early 70s "is anyone seriously considering personal computer hardware to ...". "no, we should just buy IBM mainframes ..."

I don't know. Im probably super biased. I like the idea of all this training work breaking the shackles of cloud/mainframe/servers/off-end-user-device and migrating to run on peoples devices. It feels "democratic".

I remember having lunch with a speech recognition researcher who was using GPUs to train DNNs to do speech recognition in 2011. It really was thought of as niche back then. But the writing was on the wall I guess in the results they were getting.

AMD didn't read the wall, unfortunately.

I don't think of examples really apply, because it's more a question of being on "cutting edge" vs personal hardware.

For example, running a local model and access to the features of a larger more capable/cloud model are two completely different features therefore there is no "no we should do x instead".

I'd imagine that a dumber local model runs and defers to cloud model when it needs to/if user has allowed it to go to cloud. Apple could not compete on "our models run locally privacy is a bankable feature" alone imo, TikTok install base has shown us enough that users prefer content/features over privacy, they'll definitely still need SoA cloud based models to compete.

Apple are. Their “Personal Voice” feature fine tunes a voice model on device using recordings of your own voice.

An older example is the “Hey Siri” model, which is fine tuned to your specific voice.

But with regards to on device training, I don’t think anyone is seriously looking at training a model from scratch on device, that doesn’t make much sense. But taking models and fine tuning them to specific users makes a whole ton of sense, and an obvious approach to producing “personal” AI assistants.

[1] https://support.apple.com/en-us/104993

They already do some “simple” training on device. The example I can think of is photo recognition in the photo library. It likely builds on something else but being able to identify which phase is your grandma versus your neighbor is not done in Apple‘s cloud. It’s done when your devices are idle and plugged into power.

A few years ago it wasn’t shared between devices so each device had to do it themselves. I don’t know if it’s shared at this point.

I agree you’re not going to be training an LLM or anything. But smaller tasks limited and scope may prove a good fit.

Not really (I work on AI/ML Infrastructure at a well known tech company and talk regularly w/ our peer companies).

That said, inference on apple products is a different story. There's definitely interest in inference on the edge. So far though, nearly everyone is still opting for inference in the cloud for two reasons:

1. There's a lot of extra work involved in getting ML/AI models ready for mobile inference. And this work is different for iOS vs. Android 2. You're limited on which exact device models will run the thing optimally. Most of your customers won't necessarily have that. So you need some kind of fallback. 3. You're limited on what kind of models you can actually run. You have way more flexibility running inference in the cloud.

A cloud solution I looked at a few years ago could be replicated (poorly) in your browser today. In my mind the question has become one of determining when my model is useful enough to detach from the cloud, not whether that should happen.

Inference on the edge is a lot like JS - just drop a crap ton of data to the front end, and let it render.

Power for power, any thoughts on what mobile inference looks like vs doing it in the cloud?

Mobile can be more efficient. But you're making big tradeoffs. You are very limited in what you can actually run on-device. And ultimately you're also screwing over your user's battery life, etc.

I’ve found it to be pretty terrible compared to CUDA, especially with Huggingface transformers. There’s no technical reason why it has to be terrible there though. Apple should fix that.

Yeah. It’s good with YOLO and Dino though. My M2 Max can compute Dino embeddings faster than a T4 (which is the GPU in AWS’s g4dn instance type).

MLX will probably be even faster than that, if the model is already ported. Faster startup time too. That’s my main pet peeve though: there’s no technical reason why PyTorch couldn’t be just as good. It’s just underfunding and neglect

t4's are like 6 years old

And there is a lot of work being done with mlx.

Yes, it can be more cost effective for smaller businesses to do all their work on Mac Studios, versus having a dedicated Nvidia rig plus Apple or Linux hardware for your workstation.

Honestly, you can train basic models just fine on M-Series Max MacBook Pros.

A decked out Mac Studio is like $7k for far less GPU power. I find that highly unlikely.

A non-decked out Mac Studio is a hell of a machine for $1999.

Do you also compare cars by looking at only the super expensive limited editions, with every single option box ticked?

I'd also point out that said 3 year old $1999 Mac Studio that I'm typing this on already runs ML models usefully, maybe 40-50% of the old 3000-series Nvidia machine it replaces, while using literally less than 10% of the power and making a tiny tiny fraction of the noise.

Oh, and it was cheaper. And not running Windows.

They are talking about training models, though. Run is a bit ambiguous, is that also what you mean?

No.

For training the Macs do have some interesting advantages due to the unified memory. The GPU cores have access to all of system RAM (and also the system RAM is ridiculously fast - 400GB/sec when DDR4 is barely 30GB/sec, which has a lot of little fringe benefits of it's own, part of why the Studio feels like an even more powerful machine than it actually is. It's just super snappy and responsive, even under heavy load.)

The largest consumer NVidia card has 22GB of useable RAM.

The $1999 Mac has 32GB, and for $400 more you get 64GB.

$3200 gets you 96GB, and more GPU cores. You can hit the system max of 192GB for $5500 on an Ultra, albeit it with the lessor GPU.

Even the recently announced 6000-series AI-oriented NVidia cards max out at 48GB.

My understanding is a that a lot of enthusiasts are using Macs for training because for certain things having more RAM is just enabling.

The huge amount of optimizations available on Nvidia and not available on Apple make the reduced VRAM worth it, because even the most bloated of foundation models will have some magical 0.1bit quantization technique be invented by a turbo-nerd which only works on Nvidia.

I keep hearing this meme of Mac's being a big deal in LLM training, but I have seen zero evidence of it, and I am deeply immersed in the world of LLM training, including training from scratch.

Stop trying to meme apple M chips as AI accelerators. I'll believe it when unsloth starts to support a single non-nvidia chip.

Yeah, and I think people forget all the time that inference (usually batch_size=1) is memory bandwidth bound, but training (usually batch_size=large) is usually compute bound. And people use enormous batch sizes for training.

And while the Mac Studio has a lot of memory bandwidth compared to most desktops CPUs, it isn't comparable to consumer GPUs (the 3090 has a bandwidth of ~936GBps) let alone those with HBM.

I really don't hear about anyone training on anything besides NVIDIA GPUs. There are too many useful features like mixed-precision training, and don't even get me started on software issues.

If you work for a company willing to shell out sure there are better options.

But for individual developers it’s an interesting proposition.

And a bigger question is: what if you already have (or were going to buy) a Mac? You prefer them or maybe are developing for Apple platforms.

Upping the chip or memory could easily be cheaper than getting a PC rig that’s faster for training. That may be worth it to you.

Not everyone is starting from zero or wants the fastest possible performance money can buy ignoring all other factors.

Agreed. Although inference is good enough on the Mac, there is no way I am training on them at all.

It's just more efficient to offload training to cloud Nvidia GPUs

But you get access to a very large amount of RAM for that price.

Don't attack me, I'm not disagreeing with you that an nVidia GPU is far superior at that price point.

I simply want to point out that these folks don't really care about that. They want a Mac for more reasons than "performance per watt/dollar" and if it's "good enough", they'll pay that Apple tax.

Yes, yes, I know, it's frustrating and they could get better Linux + GPU goodness with an nVidia PC running Ubuntu/Arch/Debian, but macOS is painless for the average science AI/ML training person to set up and work with. There are also known enterprise OS management solutions that business folks will happily sign off on.

Also, $7000 is chump change in the land of "can I get this AI/ML dev to just get to work on my GPT model I'm using to convince some VC's to give me $25-500 million?"

tldr; they're gonna buy a Mac cause it's a Mac and they want a Mac and their business uses Mac's. No amount of "but my nVidia GPU = better" is ever going to convince them otherwise as long as there is a "sort of" reasonable price point inside Apple's ecosystem.

What Linux setup do you recommend for 128GB of GPU memory?

Not all of us who own small businesses are out here speccing AMD Ryzen 9s and RTX 4090s for workstations.

You can't lug around a desktop workstation.

> a dedicated Nvidia rig

I am honestly shocked Nvidia has been allowed to maintain their moat with cuda. It seems like AMD would have a ton to gain just spending a couple million a year to implement all the relevant ML libraries with a non-cuda back-end.

AMD doesn’t really seem inclined toward building developer ecosystems in general.

Intel seems like they could have some interesting stuff in the annoyingly named “OneAPI” suite but I ran it on my iGPU so I have no idea if it is actually good. It was easy to use, though!

There are quite a few back and forth X/Twitter storms in teacups between George Hotz / tinygrad and the AMD management about opening up the firmware for custom ML integrations to replace CUDA but last I checked they were running into walls

I don't understand why you would need custom firmware. It seems like you could go a long way just implementing back-ends for popular ML libraries in openCL / compute shaders

smaller businesses have no business having a dedicated GPU rig of any kind

I don't think this is what you meant but it matches the spec: federated learning is being used by Apple to train models for various applications and some of that happens on device (iphones/ipads) with your personal data before its hashed and sent up to the mothership model anonymously.

https://www.technologyreview.com/2019/12/11/131629/apple-ai-...

Does one need to train an AI model on specific hardware, or can a model be trained in one place and then used somewhere else? Seems like Apple could just run their fine tuned model called Siri on each device. Seems to me like asking for training on Apple devices is missing the strategy. Unless of course, it's just for purely scientific $reasons like "why install Doom on the toaster?" vs doing it for a purpose.

It doesn’t require specific hardware; you can train a neural net with pencil and paper if you have enough time. Of course, some pieces of hardware are more efficient than others for this.

Yes, there’re a handful of apps that use the neural engine to fine tune models to their data.

Isn't Apple hardware too expensive to make that worthwhile?

For business-scale model work, sure.

But you can get an M2 Ultra with 192GB of UMA for $6k or so. It's very hard to get that much GPU memory at all, let alone at that price. Of course the GPU processing power is anemic compared to a DGX Station 100 cluster, but the mac is $143,000 less.

You want to buy a bunch of new equipment to do training? Yeah Mac’s aren’t going to make sense.

You want your developers to be able to do training locally and they already use Macs? Maybe an upgrade would make business sense. Even if you have beefy servers or the cloud for large jobs.