09-06-2024, 12:39 PM
Musk's new project, dubbed Colossus and built by his new AI startup, xAI, has been created to train the latest version of the GROK language model, known as GROK-3. It's powered by no fewer than 100,000 Nvidia H100 GPUs.
If that's not enough for you, in an X post Musk says Colossus will double in power "in a few months" thanks to the addition of another 50,000 H200 Nvidia chips, which each pack roughly twice the AI acceleration performance of an H100 GPU.
It's not clear how much this is all costing Musk and xAI. Estimates of pricing for Nvidia's H100 GPUs vary from $20,000 to as much at $90,000 a pop. Presumably, Musk managed to get a comparatively decent deal buying 100,000 of the things in one go.
But even at the lower estimate, you're looking at $2 billion for the Nvidia chips for phase one, let alone building the datacenter, all the relevant infrastructure, staffing up, and doing all the work involved in setting up training for an advanced LLM. Oh, and whatever those other 50,000 H200 are costing on top as a little light frosting.
Indeed, it was only a few weeks ago that xAI launched GROK-2 as an exclusive-access thing for X subscribers. GROK-2 apparently made do with a piffling 15,000 H100 chips for training, the poor deluded little AI dear. And yet by some measures, GROK-2 ranks second and only behind ChatGPT-4o in the LLM league tables.
So, even the first phase will be six to seven times more powerful than GROK-2, only to supposedly double in power a few months later. Clearly, Musk has his sights set on building the most powerful LLM out there.
https://www.yahoo.com/tech/elon-musks-monster-wakes-xai-135759910.html
If that's not enough for you, in an X post Musk says Colossus will double in power "in a few months" thanks to the addition of another 50,000 H200 Nvidia chips, which each pack roughly twice the AI acceleration performance of an H100 GPU.
It's not clear how much this is all costing Musk and xAI. Estimates of pricing for Nvidia's H100 GPUs vary from $20,000 to as much at $90,000 a pop. Presumably, Musk managed to get a comparatively decent deal buying 100,000 of the things in one go.
But even at the lower estimate, you're looking at $2 billion for the Nvidia chips for phase one, let alone building the datacenter, all the relevant infrastructure, staffing up, and doing all the work involved in setting up training for an advanced LLM. Oh, and whatever those other 50,000 H200 are costing on top as a little light frosting.
Indeed, it was only a few weeks ago that xAI launched GROK-2 as an exclusive-access thing for X subscribers. GROK-2 apparently made do with a piffling 15,000 H100 chips for training, the poor deluded little AI dear. And yet by some measures, GROK-2 ranks second and only behind ChatGPT-4o in the LLM league tables.
So, even the first phase will be six to seven times more powerful than GROK-2, only to supposedly double in power a few months later. Clearly, Musk has his sights set on building the most powerful LLM out there.
https://www.yahoo.com/tech/elon-musks-monster-wakes-xai-135759910.html
