Honestly imo this is just proof apple will win in the end. Eventually a phone will be able to run a model good enough to do most things and it then is game over.
Even not being able to significantly update a model that is burned on a chip the performance gains are immense. You also don't need the latest chip fabs to make them drastically reducing the cost.
It's extremely expensive to build that and you'll be at least two major model generations behind before you even get your first wafers back. By the time you got your production run ready to go and packaged for market nobody's going to care.
Once we end up going something like 24 months between major advances and capabilities for these models then I can start to see asics for a model being possible.
Yeeeees but the models are in some sense doubling in performance every 4 months, so I expect this to happen in serious quantities approximately when the economic bubble bursts and investors are no longer willing to pay for training.
(Based on widespread news reporting of the existing impact on US electricity markets, I expect this around the end of this year; but with regards to news reporting I am aware of the Gell-Mann amnesia effect, so if this is as much BS as the water issue turned out to be…)
For the Hyperscalers...and Oracle...cant wait for the day...
I'm running this stuff at home on my desktop and using it through an app on my phone. 60-140TPS depending on model / use case.
It's more than fast enough to even maintain voice conversation.
Quantised models running overnight go most of the way for non-coding tasks.
https://www.amd.com/en/developer/resources/technical-article...
At 7 months of claude code subscription per node, the cluster pays for itself in 28 months. On a 5 year (60 month) depreciation schedule, you can buy two of those clusters for basically break even, so you get two concurrent request streams (each of which can batch, etc).
The next generation hardware has already been announced, and should ship roughly two Moore’s law doublings later. It’s likely its steady state price is <= $1400 USD (2024), and it is faster.
So, once the bubble pops (because the financial machinations eventually will come to an abrupt halt), and the labs stop buying hardware for data centers, local inference will be extremely practical and cheaper than a subscription.
My main question is, when that happens, will UNIX Surplus be selling inference servers for pennies on the dollar (like after the dotcom crash), or are the power requirements too exotic for home use?
That's why I pointed out the next generation is coming soon. Also, the AMD docs aren't using quantization (as far as I can tell, I only skimmed), which gives a speedup roughly linear in the compression ratio. Algorithms for that continue to improve, so expect a lossless factor of 2-8x on DRAM and throughput, at least.
If you take four Framework Desktop 128 GB Strix Halo motherboards at a previous low and currently-unattainable price, and ignore the cost of storage, power, networking, and cases, and then take an older model that's just about competitive with Opus 4.5, and then you quantise that model down to Q2_K_XL and ignore the performance degradation below Opus 4.5 this will cause, and then you compare the cost of this with the highest tier of Claude subscription that includes much newer models which are far more able than your Opus 4.5 benchmark, then you might have some sort of break-even inside a year.
Yes, there's a new AMD hardware generation coming, but it's going to be as heinously expensive as the current one has become, and it's still got relatively low memory bandwidth. Yes, there are already newer models than Kimi 2.5, but with the limitations of this cluster (e.g. still needing heavily quantised models to be feasible) you'll get incremental improvements at best.
I'm keen to be supportive and bullish about open/home models, but I worry that this is such a stretch, and the two options in your comparison are so incomparable, you'll turn far more people off than you convince.
I don't see how these are related.
The accuracy and capabilities of your model are directly related to its size. You need a lot of memory for that.
It will be decades before we get enough useful memory in a phone form factor at a price point people can afford it before something like a frontier model now is useful on the phone.
Now, you can run some models on your phone today.
Either way, Apple is using Google today. That could change, but Google isn't exactly getting out of the TPU business and they've been doing it a long time.
Also, some of you live in a very weird Apple bubble. Apple is not so relevant outside the US.
The United States if it dares to (I think they will try) but isn’t going to be able to stuff AI models back into the bottle open source is the future and when it comes to AI models yes you’ll be able to customize it to your specifications locally but the genie is out of the bottle. The bull out of the barn and is running down the road.
If United States insist on trying to lock the doors, censor, sanction, the rest of the world will just design and engineer around the United States. Trying to put up a wall, will damaged the United States more particularly with the current performance of Taco. None of the other countries are going to follow the United States not with the current administration they will hedge their bets.
Apple, is using Google now but that will change because the world is probably going down the open path, it’s looking like there was no real rush and no reason to spend so much money on something that’s going to be a commodity in the end, the only hold up is hardware and if it wasn’t for this current memory fiasco, many more people would have access to the hardware that they need to run models locally.
Already running models locally without apple hardware, and have an encrypted vpn tunnel from my mobile devices back to my desktop over the internet. The time is now.
1. You must be willing to be resourceful. 2. Be willing to learn, do the hard things. 3. Accept the tradeoffs.