AMD Enhances Machine Learning
December 2016 - At the AMD Tech Summit, CEO Lisa Su introduced a major shift in the data center and computing marketplace. SVP and Chief Architect of the Radeon Tech Group Raja Koduri presented AMD response to these developments in the marketplace. The market has seen the initial trend of the standalone PC moving to the Web Browser Era, we then entered the Smartphone and Apps Era. We are now entering the Immersive Era of computing. This immersive markets cover education, entertainment, gaming, medicine, design, visualization, and journalism.
Today’s computing and user environment generates over 2.5 Quintillion Bytes of data per day, and the majority of it is visual and image data. The premise of the new machine learning application space is that new paradigms in use and generation of data has given rise to our ability to create in far excess of our ability to analyze, understand and make decisions in real time. In many cases, the analysis is being completed after the temporal usefulness of the data. This is creating another issue of a backlog of unanalyzed data that has to be stored, archived and secured.
These conditions are driving the creation of Machine Learning and Machine Intelligence Data Centers that are using a combination of mathematics, software and artificial intelligence techniques (such as Neural Networks) to produce more easily interpretable visual images and pictures from the numerical data. These machine intelligence based data centers are using Heterogeneous Processors (CPUs, GPUs and specialty DSPs), Open Source Software, Open Interconnect, and Open Accelerators to bring this computing power to the users.
To address this need, AMD has launched the Radeon Instinct Initiative. This is a three part solution that is made up of :
- The Radeon Instinct Hardware Platform
- The ROCm Software Platform
- An Optimized Machine Learning / Deep Learning Framework and vertical specific applications
The mathematical base of the machine learning, both inference and training methods, is well suited to GPUs. This is the driver for the Radeon Instinct Accelerator solutions. The current generation GPUs are being used in the MI6 and MI8 5.7TFLOP and 8.2TFLOP single slot plug in cards for inference applications. The MI25 is designed for Training applications and will utilize AMD’s next generation Vega GPU when it is released in the near future. These cards are compatible with the current generation of x86 based server and processor technology, as well as next generation platforms such as AMD’s Zen.
The Radeon Instinct solution is available commercially from providers such as Super Micro with their 1028GQ-TRT server, along with the 100 Teraflop solutions from Inventec K888 G3 with Radeon Instinct and the Inventec Falconwitch with Radeon Instinct model PS1816 400 Teraflop solution.