Nvidia CEO Jensen Huang knows exactly how long it takes most companies to build an AI data center using Nvidia GPUs. He describes how his Nvidia team worked with xAI and Elon Musk to build a complete AI Data Center with 100,000 Nvidia H100 GPUs in 19 days. Elon Musk has said it took 122 days from start to finish.
Jensen said this would normally take 4 years. It would take 3 years for the planning and getting the site and permits and then it would take one year to build it and get it working and everyone trained. This means XAI took about 103 days for the planning and other non-GPU planning and other tasks.
The removal of most of the 3 year planning and preparation requires leveraging Elon’s Tesla team and processes they have worked out for factories and the prior setup of 20,000 GPUs used for Grok 2.
Debugging, optimizing, and ensuring sufficient power supply for these massive clusters can take additional time. The actual time to reach full operational capacity may be longer than the initial deployment timeframe.
The Colossus supercomputer uses at least 150 MW of power, as 100,000 H100 GPUs use 70 Megawatts. They have been using 14 diesel generators to power ithe Memphis supercomputer. The power will need to be increased to feed all 100,000 H100 GPUs.
Jensen Huang on Elon Musk and xAI building the fastest supercomputer on the planet in 19 days
“From the moment of concept to building a massive factory, liquid-cooled, energized, permitted, in the short time that was done. That is superhuman. And as far as I know, there’s only… pic.twitter.com/k57ye2YlkQ
— Startup Archive (@StartupArchive_) October 16, 2024
Tesla, SpaceX AI for Testing and Processes Have Been Used at xAI and Tesla AI Data Centers
Joe Justice describes that Tesla has a stack of scripts that run multiple times a second. Changes can be imagined, designed, produced and tested and put into production in the same day.
Everything is optimized for Pace of Innovation. Tesla has built instant testing into its cars even as they are being built.
One of the first things installed is the computer and monitor in the car. When other parts are added, the software runs tests to verify that it has been installed correctly and is meeting standards.
This has likely been added to the xAI AI data center installation, build and testing processes.
Tesla can completely design, build and test and certify a car in one hour. Other car companies take a year. If xAI has converted the installation, building and testing of a 100k GPU cluster from a 365 days process into a 19 day process and shortened the planing and preparation from 3 years into 100 days then this could be an unbeatable level of speed for xAI and for Tesla AI.
AI Data centers need to be built and upgraded as fast as the chips can be made available.
Faster xAI and Tesla AI Data Center Construction Could Be A Decisive Advantage
xAI will be making an expansion that will double the size of the Colossus data center in next 4-5 months.
xAI will build another even larger 300k B200 system in the summer. This will be about 12X the compute of the current system.
So three iterations of data center builds in the roughly 120 day timeframes. If no other company can match this then xAI would build a 2+ year lead.
Elon and his team's "superhuman" engineering speed continues to amaze. Piecing together xAI’s rapid progress: ⌛️Jul 2023: Founded ⌛️Nov 2023: Grok 1 released ⌛️Aug 2024: Grok 2 released, contracting 24,000 H100 GPUs from $ORCL (Elon noted 15,000 were used for training). It’s… https://t.co/q4RI3NO6bl pic.twitter.com/UBYqGZYjsP — Freda Duan (@FredaDuan) October 17, 2024 “> Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology. Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels. A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
Seems we are currently in the learning phase of LLMs yet. The models haven’t converged into a good enough AGI yet. But it will eventually happen and then, the focus will be in more efficient ways to implement de NN.
Like non-MatMul algorithms, ASICs, photonic crystals and crazier ideas yet.
Of course, if we assume AIs will be the ones training and improving their future versions, the algorithmic stability might never arrive. But I think that at some point we will actually reach a plateau of “good enough” for most real world uses and purposes.
Even if we still have self improving AIs going into ASI somewhere, the good-enough models could go into a somewhat fixed state and serve for most commercial uses not needing super-intelligence.
Wonder how long it will take him to design an equivalent datacenter module that can fit into starship and be powered by solar arrays? Communications via StarLink. I see that as a good solution for the future demand for massive AI computation and its equally massive power requirement. Gonna be a LOT of AI brains floating out there eventually.
I don’t like those diesel generators – couldn’t he have used his solar array tech instead? He seems to be repudiating his earlier goals to replace fossil fuel usage via solar with battery storage and electric vehicles.
cooling is a huge issue in space, the size of the radiators would dwarf the starship if it was purposed as a spacefaring datacenter.
Yes.
But at the end of the day, it’s the quality and desirability(!) of the end product – not the speed at which One can make the factories and machines and supply lines to end-result that product.
Until Tesla starts offering PHEVs. Until xAI can offer functionality that differentiates them from the AI crowd beyond: faster. faster. faster. Until TeslaBots start effectively replacing minimum wage and low-skill labor in all sectors – inside and outside. Until SpaceX accesses a market that has actual long-term growth beyond science and satellites…
they will All grab quick and easy market share in a blind-side maneuver and then fall back to mediocrity (or worse) when competitors feel out the real market future. It’s the ‘I’m Elon and I get bored easily’ syndrome.
It’s always appealing to be first and then Walmart-ize or Amazon-ize an industry by promising more, faster, cheaper — but this is cutting edge technology, not construction materials.
The refined competition is mounting… Toyota + Boston for new human-skill Bots; everyone for PHEVs; the usual suspects for AI; no-one for Space; new and better battery super-gigafactory in NV for LiS, previous SpaceX engineer putting together a Cargo Airship industry (Airship Industries) that will likely challenge long-haul (autonomous) trucking across the US….
It’s one of kind of CEO that builds an industry and another type that sustains and grows an Industry.. my 2c… to be continued…
I was almost following you. Then you said this “Until SpaceX accesses a market that has actual long-term growth beyond science and satellites…” and talked about PHEV. Starlink is connecting the entire world with internet. The upside potential for profit is a good portion of the worlds population. PHEV? 99.2% of trips are under 100 miles. Why would you add the complexity of gas motor? Who would buy it? The best selling phev is a Jeep wrangler and Honda Civics sell triple the numbers. You need to reevaluate your ideas.
Your right. But only to a psychological point. I don’t know were you get the figure; 99.2%, but your correct that most trips “any one day” are less then 100miles. Hey, I love EV’s, but since you can’t “juice-up a car” as fast as you can “gas up one”, that will always be a problem, until our technology improves. Questions?