The Future of AI is Here - Highlights from NVIDIA GTC 2024 Keynote
Introduction
In a remarkable scene reminiscent of a rock concert rather than a tech conference, NVIDIA’s GTC 2024 transformed expectations. The star? Not a typical celebrity, but Jensen Huang, NVIDIA’s founder and CEO, clad in his signature leather jacket. Amidst the backdrop of NVIDIA’s meteoric rise to a $2 trillion titan, Huang has become an emblem of tech royalty, rivaling the stardom of icons in completely different fields. As NVIDIA reclaims the in-person stage after five years, the palpable energy and packed venue underscore the seismic shifts AI is driving across every conceivable sector. Huang’s jest that “this isn’t a concert, but a developers’ conference” couldn’t mask the fervor of his ‘fans’—a testament to the profound impact NVIDIA’s computing prowess and AI innovations are set to unleash. In an age where AI’s potential is boundless, NVIDIA’s GTC 2024 stands as a pivotal moment, heralding a future where technology and human creativity converge in unprecedented ways.
Blackwell Era
NVIDIA Blackwell Platform Arrives to Power a New Era of Computing.
The larger volume does not sufficiently demonstrate the improved performance resulting from the advanced process. For a more illustrative example, consider OpenAI’s latest GPT-4-1.8T parameter scale, which can complete 2000 Blackwell units in just 90 days of training. In contrast, the same workload would require 8000 GPU to be completed on the Hopper GPU. By utilizing the advanced process, costs associated with electricity and token generation can be reduced to a quarter of the original cost, while energy consumption and training efficiency are increased by 25 times.
They don’t offer a single piece of hardware, they offer a complete solution.
To improve hardware sales, excellent software is indispensable. The potential to disrupt the market often lies in the earliest software. Therefore, as early as thirty years ago, NVIDIA heavily invested in the software ecosystem while selling hardware.
"The more you buy, the more you save. You don't need to understand strategy or technology, that's NVIDIA."
From 黄仁勋
On top of the GB200, NVIDIA is also offering a solution to paralleling tens of thousands of GB200s into a true supercomputer via NVIDIA Quantum InfiniBand networking technology, a system capable of stringing together hundreds of thousands of Blackwell GPUs as a single unit to train the top generative AI macromodels on the planet. The NVIDIA DGX SuperPOD powered by NVIDIA GB200 Grace Blackwell Superchips — for processing trillion-parameter models with constant uptime for superscale generative AI training and inference workloads.
“In the future, data centers are going to be thought of … as AI factories,” Huang said. “Their goal in life is to generate revenues, in this case, intelligence.”
NVIDIA NIM
A New Way to Create Software.
Enterprises that wish to utilize large models often require fine-tuning and privatization of their deployment to benefit their business. Currently, this is typically accomplished through cloud vendors or small service providers that deploy the models.
Nvidia has launched its own model customization service, Nvidia Inference Manager (NIM), at the conference to facilitate easy debugging and use of large models wherever Nvidia hardware is available. This is insane.
Their new service also provides tools that help users fine-tune large models, which include multiple microservices.
Omniverse
“We need a simulation engine that represents the world digitally for the robot so that the robot has a gym to go learn how to be a robot,” he said. “We call that virtual world Omniverse.”
They continue to optimize Omniverse while bringing Omniverse to Apple Vision Pro, and the new Omniverse Cloud API allows developers to stream interactive industrial digital twins into VR headsets.
Robotics
Huang just announced Project GR00T, which is a foundation model for humanoid robots. The project is aimed at driving breakthroughs in robotics and embodied AI. How cool is that? They also introduced the Jetson Thor, a new computer designed for humanoid robots. It’s based on the NVIDIA Thor system-on-a-chip and includes significant upgrades to the NVIDIA Isaac robotics platform.
I think this part is the most exciting moment of the entire conference. Year of the humanoid is coming. We are moving step by step towards the world presented in science fiction movies and novels, let’s make it happen.
"The soul of NVDIA — the intersection of computer graphics, physics, artificial intelligence. It all came to bear at this moment."
From 黄仁勋
