Colossus AI Supercluster with over 100k Nvidia H100 GPUs
twitter.com
twitter.com
Gamers laid the foundations for GPGPU taking off, we're now getting left in the AI dust.
Oh well, plenty of fun indie games being made.
It’s a very expensive, tragic, joke.
They're neat, they have uses, but they're not replacements for anything. Even Whisper cannot replace a human transcriptionist as it will just make up and insert random lines that are not present in the source audio.
I assume native english speakers do better (?) but we speak english (with accents) and whisper has no issues at all.
A failing traditional TTS can be spotted by glancing through a transcript. A failing Whisper can only be identified by thorough comparison, with the failures being far more impactful and important to spot.
I hope we get to see low power, local LLMs with good performances rather than continuing building huge energy sinks like this one in times where we should be trying to lower our footprint.
> Now we can use supercomputers to "simulate" the workflows of 95% of the office drones worldwide *(us included)*. I think that's a MASSIVE paradigm changing leap forward.
This is not what it is doing. And if it was, why would I, as an "office drone" celebrate this...?
I hope they use it for more than that.
Because the way Twitter is going at the moment they are likely to be massively fined and/or banned by the EU. They have been continually warned to sort out the misinformation and by every definition it has gotten far worse since Musk took over.
Was always surprised this purchase wasn't made by Tesla given they have actual use cases for it.
Because nobody in management tends to notice or care when white-collar workers get stripped of most or even all their job responsibilities.
There's a whole genre of Reddit post that goes something like "I only do a few real hours of work per week" or "I was hired but never told to do anything" or "my team was laid off but they forgot to fire me for five years, so I just came in and goofed off all day."
(I think this is because many types of white-collar worker are secretly valued more for their knowledge and edge-case intelligence — i.e. their ability to solve emergency problems in a pinch, or randomly save the company a million dollars one day — than they are for consistent, productive output. For many such workers, "regular job duties" are just thing management uses to keep them busy so that they don't get bored and quit!)
For a healthy company, "AI taking your job" actually means "AI taking all the stressful and boring parts of your job", leaving you to do just the fun parts. It's like the people who do data entry who secretly automate 99% of their work, leaving them to just "manage" a script — but someone else is stepping up to create the "script" for you to "manage."
https://www.servethehome.com/inside-100000-nvidia-gpu-xai-co...
No, and please dont spread incorrect information just because "there's a meme."
The cluster has 100,000 H100's. Each H100 takes about 700 watts. Call it 1kW per H100 for simplicity, and assume power is about $0.10/kWh (they probably built this somewhere power is cheap, so it's likely cheaper than that).
100,000kW * $0.10/kWh = $10,000/hour or $33 power cost for 12 seconds.
A possibly more interesting, and true, statistic is it costs about a quarter million dollars a day to power up this cluster.
hah hah, train some more.
Meta is using more than 100k Nvidia H100 AI GPUs to train Llama-4
Really? Not glycol? [0]
0. https://www.towerwater.com/how-does-a-glycol-cooling-system-....