Source: This interview between Emad Mostaque (CEO of StabilityAI) and Peter Diamandis. The relevant bit is at 4min53s [0].
It was published on April 6th and Emad claims that "in 6 months" (early October?) we are going to see GPU tech that will allow much larger scale (for large companies) and democratize training AI systems (for smaller companies).
His followup claim is that NVidia has solved the scaling problem that was limiting large GPU clusters to ~1000 GPUs. He says while GPT-3 took three months to train, this new supercomputer hardware will enable training 4 models of GPT-3's size per day.
However I couldn't transcribe the tech's name properly, nor could I find any announcement or literature around it. What are some sources where one could learn more about this?
It was published on April 6th and Emad claims that "in 6 months" (early October?) we are going to see GPU tech that will allow much larger scale (for large companies) and democratize training AI systems (for smaller companies).
His followup claim is that NVidia has solved the scaling problem that was limiting large GPU clusters to ~1000 GPUs. He says while GPT-3 took three months to train, this new supercomputer hardware will enable training 4 models of GPT-3's size per day.
However I couldn't transcribe the tech's name properly, nor could I find any announcement or literature around it. What are some sources where one could learn more about this?
[0]: https://www.youtube.com/watch?v=SKoYhcC3HrM&t=293s