Nvidia’s AI-focused keynote at Computex 2021

Nvidia had the honor of opening the Computex with its keynote this year and it was noticeable that it was the opening keynote after a Computex hiatus of three years. In front of the packed room, an almost elated Jensen Huang, the CEO of Nvidia, presented what his company has been working on recently and what it will come up with in the near future. Given recent developments, it will come as little surprise that AI played a leading role in this. Virtually everything Nvidia showed was aimed at facilitating or exploiting AI.

The consumer market played only a minor role in this. Not surprising, because Nvidia had of course already lost its powder in that area with the introduction of the RTX 4060 Ti less than a week earlier. We as consumers are of course noticing all the developments that were presented as almost revolution, in everything from gaming to software use, and maybe even our health and the way we work.

Consumers

After Jensen emphasized how fast Nvidia’s GPUs and especially ray tracing have become, the 4060 Ti was shown again. That turned out to be the prelude to Nvidia ACE, a framework for creating AI characters for games. Colleague Daan wrote an extensive news item about this. In short, generative AI is used to create characters in games that you should be able to interact with in a natural way.

After this gaming foray, Jensen focused on accelerated computing and generative AI. According to him, accelerated computing and generative AI come at exactly the right time, as traditional scaling and Moore’s Law slow down. With the right workloads, a GPU cluster is much more efficient than computing the same on traditional CPUs. Calculating a large language model, the building block of ChatGPT, for example, is much faster with 10 million dollars in GPUs than the same amount for CPUs. Also according to other metrics, such as energy consumption, GPUs are much more efficient than CPUs. Finally, with just a fraction of the investment, $400,000, you can perform the same calculations on the llm as with those $10 million of CPUs.