Summary
Highlights
Jensen Huang explains how NVIDIA started in the early 90s by recognizing that a small percentage of software code performs most of the processing, and this processing could be done in parallel. This led to the creation of the GPU, initially for video games, which require extensive parallel processing for 3D graphics. This approach was revolutionary, offering significantly faster computation, akin to a 'time machine' for scientific endeavors.
NVIDIA developed CUDA to make the power of GPUs accessible to a wider range of researchers and developers, allowing them to program GPUs using familiar languages. This platform was born out of both desperation to solve internal problems in creating dynamic virtual worlds and external inspiration from researchers using their GPUs for general-purpose acceleration, like medical imaging. The large market for video games ensured high production volume for GPUs, making CUDA a viable and impactful platform.
In 2012, AlexNet, a deep neural network, leveraged NVIDIA GPUs to significantly outperform competitors in image recognition, marking a seismic shift in computing towards AI. This moment demonstrated the potential of training computers with vast amounts of data rather than step-by-step instructions. Huang recognized the profound implications for re-engineering the entire computing stack, leading to the development of AI supercomputers like DGX.
Huang discusses the core beliefs that sustained NVIDIA's commitment to AI despite a decade of significant investment before mainstream adoption. These beliefs center on accelerated computing and the scalable nature of deep learning networks. He emphasizes that AI's ability to learn from various data modalities, essentially digital human experience, opens a universe of problem-solving opportunities, from text-to-text translation to predicting protein structures.
The next decade will focus on the application of AI, particularly in physical AI or robotics. NVIDIA's Omniverse and Cosmos platforms are crucial for training robots in digital worlds, allowing them to learn gravity, friction, and cause-and-effect much faster and more safely than in the real world. Huang envisions a future where 'everything that moves will be robotic,' including humanoid robots, autonomous vehicles, and smart environments, with robots learning in these simulated environments grounded in physical truth.
Huang acknowledges the concerns surrounding AI, such as bias, hallucination, and impersonation. He stresses the importance of engineering for AI safety, ensuring systems function properly and don't harm. He also highlights the fundamental technological limit of energy efficiency, stating that advancing computing relies on making systems more energy-efficient. NVIDIA has achieved significant improvements in this area, increasing AI computing energy efficiency by 10,000 times in eight years.
NVIDIA's approach to hardware design prioritizes flexibility. While specialized chips for certain AI models could be built, Huang believes in creating architectures that enable continuous innovation and new ideas, rather than locking into a single algorithm like the 'transformer.' This philosophy ensures that NVIDIA's hardware can adapt to the rapid evolution of AI research and computer science.
NVIDIA's current significant bets include the fusion of Omniverse and Cosmos for generative world creation in robotics, further development in human robotics, and groundbreaking work in digital biology to understand molecular and cellular languages. They are also investing in climate science for high-resolution regional climate predictions. Huang views GPUs as 'time machines' that allow scientists to foresee and optimize for the best possible future in these critical areas.
Huang advises everyone to learn how to interact with AI, seeing it as a crucial skill for the future. He believes AI will act as a personal tutor and assistant, empowering individuals to become 'superhuman' by reducing the drudgery of tasks and lowering barriers to knowledge and intelligence. He encourages everyone to engage with AI tools like ChatGPT to enhance their professional and personal capabilities.
Huang showcases the new GeForce graphics card, highlighting its AI-powered capabilities in gaming, where AI predicts most pixels to achieve perfect image quality with fewer computations. He also presents a mini version of the DGX AI supercomputer, now accessible at a lower price point, to make AI development available to a broader audience, including students and engineers, enabling them to build their own AI applications.
Huang reflects on NVIDIA's journey, crediting its extraordinary impact to core beliefs and consistent dedication. He emphasizes the company's commitment to making its powerful capabilities accessible to all, from large corporations to individual researchers, across diverse fields. He hopes future generations will recognize NVIDIA's role at the epicenter of transformative changes in digital biology, material sciences, and robotics.