At the Hot Chips 2024 conference, held in Palo Alto, California, IBM has revealed its new accelerator chip Spyre, designed to enhance the artificial intelligence (AI) capabilities in its IBM Z systems. This innovative technology, developed in collaboration with IBM Research, promises to revolutionize AI workload processing in enterprise environments by providing unprecedented performance.
A Quantitative Leap in AI Architecture
The Spyre is the evolution of the AI chip that IBM introduced in 2022 with the IBM z16, which incorporated the Telum microprocessor, an integrated AI accelerator for real-time inferences. The new Spyre chip takes this technology to the next level by incorporating 32 acceleration cores and an advanced architecture with 25.6 billion transistors.
The Spyre utilizes 5 nm process technology and is mounted on a PCIe card. This design allows for card clustering, meaning a single IBM Z machine can be equipped with clusters of Spyre cards to add up to 256 additional acceleration cores.
Optimization and Scalability for Enterprises
With approximately 70% of the world’s transaction value managed by IBM mainframes, the Spyre has the potential to bring generative AI to these mission-critical systems. This chip is designed to help businesses scale their AI capabilities as needed, enabling the deployment of AI models and programs with greater power and efficiency.
The Spyre not only enhances performance compared to standard CPUs but also uses a chip design optimized for AI tasks. Unlike traditional architectures that constantly transfer data between the processing unit and memory, the Spyre allows data to be transferred directly between calculation engines, resulting in energy savings and increased efficiency.
Applications and Future Possibilities
The Spyre will open up new opportunities for AI usage in IBM Z systems. Beyond fraud detection in transactions, a system equipped with a Spyre cluster could use more complex AI models to identify intricate fraud patterns that less advanced models might overlook.
Additionally, the Spyre will enable deeper integration with products like Watsonx, IBM’s AI and data platform. For example, it will allow running the Watsonx Code Assistant to modernize code bases on mainframes much more effectively, facilitating the understanding and updating of application code.
A Promising Future for AI in Mainframes
This development marks just the beginning of what IBM Research envisions for AI on IBM Z. The research team is working to push the current limits of inference and explore effective ways of fine-tuning and even training models directly on mainframes. The vision is to enable organizations to train and deploy models on platforms like Watsonx within their premises, keeping all data secure and in place.
With Spyre, IBM is opening a new chapter in AI processing for enterprises, offering high-powered solutions to tackle complex challenges and ensuring that AI capabilities are within reach of organizations that require maximum security and performance.
Source: IBM