New AI Driven Semiconductor Market

We propose a new business model RUMS – Rapid and Unified Manufacturing Service – which is suitable for the era of AI with high variety and specialized chips and speed, in contrast to the current mainstream so-called “Fabless-Foundry” model. RUMS promotes close collaboration between design and manufacturing through co-creation with customers, aiming to realize true well-being of humans with innovative and green end products.

Tenstorrent uses open software, chiplets and industry standards like RISC-V, Ethernet, MLIR, etc to build a scalable AI solution from embedded to data center. Our hardware design is modular and easy to understand. Our software stack is inherently multi chip aware with deep integration of sharding, data movement and networking. This talk will describe how we use industry standards and layers to build a scalable and accessible AI computer.

For the past two years, we have seen AI capturing the imagination of the public and institutions alike. But as fast as AI is advancing, we are not extracting its full value. Delivering maximum efficiency and performance AI requires an end-to-end solution that co-optimizes hardware, algorithms and AI models, software, and applications— everything from sand to cloud, including handling multiple types of AI workloads in heterogeneous and distributed environments. In this keynote, we will examine what it takes to build such a system.

We will see how advances in semiconductor technology, chip design, and optimizations for AI workloads lead to chips that make training and running deep learning models less memory intensive, faster, and more efficient. The integration of these chips into the design of AI infrastructure can determine the cost, speed, and efficiency of every stage of the AI lifecycle. Innovations in software and decoding techniques are bringing several times latency improvements. Infrastructure and software co-designed and co-optimized with models and advances to increase the efficiency of algorithms will deliver improvements in speed, energy, compute space, and time efficiency.

A holistic approach that considers algorithms with infrastructure and software up and down a fully integrated stack focused on excelling at critical enterprise workloads has been IBM’s philosophy to building computing systems since its dawn. We did it with IBM Z, a highly specialized full stack system that continues to be the workhorse of enterprise computing and premier transaction processing platform in the world. We are doing it with IBM quantum systems, the industry leading quantum computing platform. And now, we are doing it with generative AI. Join Darío Gil, IBM Senior Vice President and Director of Research, as he unpacks how IBM is building what’s next in generative AI systems to unlock the value of artificial intelligence in the enterprise.