Elon Musk gives an overview of the challenges one might face when trying to use NVIDIA GPUs. #elonmusk #interview #viral

By admin

23 thoughts on “Elon Musk Just Said It Is INCREDIBLY DIFFICULT To Use NVIDIA GPU’s For AI Training | #shorts”
  1. For people wondering, NVIDEA GPU’s are used to run LLM’s (large language models). GPU’s excel at mathematical floating point calculations which are required to perform training of the models. Basically they ingest all sorts of content, words, or images , to create the models that are used to describe things. Imagine feeding it 100,000 images of a variety of dogs so it can learn what dogs look like. The processors are all interconnected. They speed up and slow down during the learning process as different commands are executed. When they change speed the do so in a coordinated fashion (hence the symphony). They use more power when they speed up and less power when the slow down. These speed changes happen rapidly. And since you are dealing with a 50000 node cluster, the power changes are significant. I’m not sure how they address those power demand fluctuations. Perhaps with large scale capacitors in a UPS? 50000 nodes throttling up and down could amount to 5 megawatt or greater power fluctuations in very short periods of time.

  2. And of course as any gamer will tell you, GPUs can massively spike way beyond their specified power draw for a fraction of a second when under load. Fun times in a cluster.

  3. Elon, thanks for everything you do for humanity. I wish you would consider Kennedy instead of Trump. I love Trump, but i think he will be too easily swayed by Party insiders like las time. Not as badly – but still. Kennedy & Shanahan are better managere. Did you see her interview with Weinstein the physics guy? Amazing. A very smart lady and sees the whole chessboard, not just one move combination.

  4. Isn't that a problem that a capacitor is designed to solve? Why not use very large capacitors to continuously draw large amounts of power (this should be calculated depending on the amount of GPU usage you project to use), and the gpu power draw should come from the capacitor banks. Any excess energy can be redirected to the grid after training.

  5. Good luck with Texas’ electric grid which is notoriously unreliable because it is completely independent from the rest of the country and can’t access power from the national grid when the State’s power is tapped out.

Leave a Reply

Your email address will not be published. Required fields are marked *