Why Does NVIDIA CEO Jensen Huang Say Moore’s Law Is Over? (Supercut)

TL;DR

Jensen Huang says Moore’s Law is over because CPU scaling has reached its limit, ending automatic improvements in performance and cost without software changes. NVIDIA’s alternative is accelerated computing: domain-specific software rewritten for GPUs and CUDA, producing reported application accelerations of 20, 30, or 50 times. Read on to understand how this shift also moves development from human-coded software to machine learning and systems such as Blackwell.

Transcript

general purpose Computing as we know it has existed for 60 years until now for the last 30 years we've had the benefit of mors law an incredible phenomenon without changing the software the hardware can continue to improve in an architecturally compatible way every single industry has subsequently been built on top of it but we know now that the sc... Read More

Key Insights

• Moore's Law has reached its limit, necessitating new approaches in computing.
• Accelerated computing, particularly through GPUs, is essential for future advancements.
• NVIDIA's CUDA architecture has been pivotal in transforming multiple industries.
• Machine learning has disrupted traditional software development methods.
• NVIDIA's Blackwell system is a breakthrough in handling large-scale data and AI tasks.
• The future of AI involves creating agents that can perform complex tasks autonomously.
• Physical AI requires virtual environments like Omniverse for training and development.
• NVIDIA's technology development is driven by two scaling laws: training and inference.

Questions & Answers

Q: Why does Jensen Huang say Moore’s Law is over?

Huang says CPU scaling has reached its limit, ending the “free ride” in which compatible hardware improved without developers changing their software. Computing performance and costs can no longer be expected to improve while software remains untouched, so applications must adopt a different approach.

Q: What is accelerated computing?

Accelerated computing uses specialized processors and redesigned software to make applications that are impossible or impractical on general-purpose computers feasible. Huang says there is no single processor that accelerates everything, so NVIDIA reinvents the stack from algorithms and architecture through the applications above them.

Q: Why are GPUs important after the end of CPU scaling?

GPUs provide acceleration for workloads that general-purpose CPUs cannot perform efficiently, beginning with real-time computer graphics. They also run neural networks well, making them central to the shift from conventional software development to artificial intelligence.

Q: How much can NVIDIA’s acceleration libraries improve application performance?

Huang says NVIDIA can accelerate applications by 20, 30, or 50 times across its domain-specific libraries. Achieving those gains requires rewriting software and working with industry partners, software developers, customers, and the wider ecosystem.

Q: What industries and applications does NVIDIA CUDA accelerate?

Huang names semiconductor manufacturing, computational lithography, simulation, computer-aided engineering, and quantum computing among the areas CUDA supports. NVIDIA expanded the architecture one domain at a time until its acceleration libraries covered nearly every industry.

Q: How does software 2.0 differ from software 1.0?

In software 1.0, programmers write algorithms in languages such as Python, C, Fortran, Pascal, or C++ to transform inputs into outputs. In software 2.0, a computer studies patterns and relationships in large amounts of observed data, learns the predictive function, and produces a trained model for inference.

Q: What role did NVIDIA’s cuDNN library play in artificial intelligence?

Huang describes cuDNN as one of NVIDIA’s most famous application libraries and says it helped democratize artificial intelligence. It forms part of the broader collection of domain-specific acceleration libraries built on NVIDIA’s computing architecture.

Q: What is NVIDIA Blackwell designed to do?

Blackwell is presented as a massive system designed to study data at enormous scale. Its purpose is to discover patterns and relationships and learn representations from many data types, including words, numbers, images, videos, chemicals, proteins, fluid patterns, and particle physics.

Summary & Key Takeaways

• NVIDIA's CEO Jensen Huang highlights the end of Moore's Law and the rise of accelerated computing. GPUs, unlike traditional CPUs, provide the necessary power for AI advancements. NVIDIA's Blackwell system exemplifies this shift, offering unmatched data processing capabilities.

• The transition from software 1.0 to software 2.0 involves machine learning taking precedence over human coding. This shift allows for more efficient and powerful AI applications, with NVIDIA leading the charge in this technological evolution.

• NVIDIA's advancements in AI, particularly through its Blackwell system, are set to revolutionize industries by enabling the creation of intelligent agents. These agents can autonomously perform tasks, marking a significant step forward in AI capabilities.