Deep Dive into AMD Advancing AI 2026: An Insider Perspective on Hardware, Software, and the Future of AI
A recap of the opening day of AMD Advancing AI 2026 at the Moscone Center in San Francisco, featuring insights from tech industry gurus.
Stock photo for illustration only, not from the actual event
- The AMD Advancing AI 2026 event was held free of charge at the Moscone Center in San Francisco.
- Chris Lattner, creator of LLVM, compared CUDA to GCC as a powerful yet legacy monolith.
- George Hotz, founder of tinygrad, noted that the framework consists of just 25,000 lines of Python code.
- The panel discussion emphasized the importance of a unified ecosystem across hardware, software, and frameworks.
The two-day AMD Advancing AI 2026 seminar at the Moscone Center in San Francisco welcomed attendees to explore, learn, and interact with exhibitors and speakers completely free of charge. This represents a major investment by tech companies in driving their own ecosystems. While shareholders may eventually look for a return on these investments, for attendees, it served as a prime opportunity to seriously explore an artificial intelligence landscape deeply focused on hardware and robotics.
Day one kicked off with a panel discussion featuring three speakers: Chris Lattner, Ramin Hasani, and Hassan Akbari. Chris Lattner, creator of the LLVM compiler infrastructure, discussed the inseparable relationship between hardware and applications. He drew a comparison between NVIDIA's CUDA and GCC, the decades-old open-source compiler that is massive and deeply entrenched like a defensive moat. However, he noted that the strategy to compete isn't to charge head-on into that moat, but rather to use architectures that bypass its limitations via Modular's Mojo language as a portable alternative to CUDA.
The comparison of the compiler and hardware ecosystem's technology "moat" reflects how being locked into a single language or platform poses challenges for developers in the artificial intelligence industry. The emergence of alternatives like the Mojo language or the development of new architectures serves as a key strategy to build flexibility and long-term competitiveness.
Meanwhile, Ramin Hasani from Liquid AI presented conceptual-level insights, emphasizing the importance of solving problems at the algorithmic level before looking at kernel optimizations. This includes leveraging Liquid foundation models and matching the best models to the appropriate hardware, whether CPU, NPU, or GPU. At the same time, Hassan Akbari tied all these perspectives together by stressing the importance of a unified ecosystem bridging frameworks, hardware, and software, which must be connected to maximize collaborative performance rather than optimizing any single component in isolation. He also raised a thought-provoking question regarding the practical cost-efficiency of model scaling.
Another major highlight was the appearance of George Hotz, famous for jailbreaking the iPhone, reverse-engineering the PlayStation 3, and developing the comma.ai self-driving car project. During his presentation, he discussed hardware utilization such as the Radeon RX 7900 XTX graphics card, which despite being on the market for three years still offers great value at its 999 USD price point.
"commoditize the petaflop"
George Hotz
Additionally, George Hotz talked about tinygrad, an open-source framework for neural networks and deep learning libraries, stating that the system consists of just 25,000 lines of pure Python code, with the core code sitting at around 9,000 lines. This reflects a streamlined and straightforward approach to technology development.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment