AMD CEO Lisa Su Highlights Global AI Infrastructure Changes Shift to CPU GPU Era and Local Execution of 200 Billion Parameter Models
At AMD's AI Developer Day event held in Shanghai, the Chief Executive Officer of AMD, Lisa Su, discussed about the tremendous changes taking place in the AI infrastructure worldwide. In keynote speeches delivered at the event, it was highlighted that in the large networks nodes the CPU to GPU ratio that stood at 1:4 from 2022 to 2025, has now reached to a 1:1 ratio and is shifting from the ratio of 1:4 to 1:1. According to Su, proliferation of the autonomous AI agents and localized inference tasks is changing the role of CPU.
In the corporate scenario, AI technology is moving from the pilot stage to all device level expansion. According to Su, the industry is stepping into the CPU plus GPU era and software applications are designed to scale both up and down, and are also scaling from the devices in an end to end system up to the cloud. While the GPUs would remain essential throughout the cloud, PCs, robotics, CPU with huge parallel processing ability is required to handle AI agents. According to AMD, its corporate strategy is to supply integrated end to end computing packages which are tailored to meet the needs of particular software with minimum hardware requirement.
AMD has come up with a small desktop with size of palm, running on the Ryzen AI Max processor which is able to execute a 200,000,000,000 parameters AI model offline. This model uses shared memory structure of unified memory which uses 128 GB RAM for CPU, GPU and NPU. Shared memory greatly eliminates the latency when transferring data from video to main memory and also conserves energy. This task required the cluster data center 3 years back but, the new AMD system executes the load on device without internet.
According to future predictions, in the coming 5 years, 5,000,000,000 users will be actively using the AI. Su predicts that, industry has gone beyond the phase of large language models and the next 10 years in Digital Productivity, the usage of multiple digital agents working together on several operations will result in individual user using 5, 10 or 100 digital assistants for carrying out the operations.
