Model-on-Chip is coming. GPU are for general computing but have a huge bottle neck for doing model inference.

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.

I don't know what you mean by "economical", but it has been "economical" to run this stuff yourself for the last 3 years.

1. You must be willing to be resourceful. 2. Be willing to learn, do the hard things. 3. Accept the tradeoffs.

Pre bubble prices (~= “we stop building data centers with subsidized credit / circular loans / hidden debt”), a 128GB halo strix ran for $1400, and 200-ish watts. Four of those in a cluster will run a 1T parameter frontier model:

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?

>> do most things and it then is game over.

For the Hyperscalers...and Oracle...cant wait for the day...

> it really isn’t economical to run this stuff yourself

Quantised models running overnight go most of the way for non-coding tasks.

> Sadly until china scales production of hardware it really isn’t economical to run this stuff yourself.

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.

> 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.

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.

[dead]