Advanced Micro Devices (AMD) is increasing its product line by introducing new processors that are specifically designed for use in data centers. The company's CEO, Lisa Su, announced the CPUs at the CES 2023 exhibition, which include Alveo V70 and Instinct MI300.
(https://cdn.wccftech.com/wp-content/uploads/2021/05/AMD-EPYC-Milan-Server-CPUs.jpg)
The Instinct MI300 is an APU (adaptive processing unit) with 146 billion transistors, and it can be integrated into a single box that combines components from a CPU and a GPU. This processor is designed to run high-performance computing and artificial intelligence software. The MI300 is made up of chiplets arranged in a three-dimensional structure as two layers layered on top of one another.
The first layer includes nine CPU and GPU chiplets built using a five-nanometer technology that uses AMD's newest Zen 4 processing architecture as the foundation for the CPU modules. The second layer comprises four additional chiplets that oversee auxiliary functions like data intake and output created using a six-nanometer technique.
In addition to the MI300, AMD introduced another data center processor at the CES event, the Alveo V70. It has an architecture created by its Xilinx subsidiary and is designed for inference, which is the process of putting learned AI programs into use. According to the manufacturer, the AMD Alveo V70 AI Accelerator provides the highest performance and energy efficiency for a variety of AI inference tasks.
The MI300 boasts a 128 GB onboard memory pool that saves the processed data, and AMD claims that sharing memory across the chip's GPU and CPU modules makes them more effective than conventional data center processors. With MI300, the data being processed is kept on each chip separately, eliminating the need for the CPU and GPU modules to make multiple copies of the same data, thereby improving processing speed. AMD also claims that the MI300 can train AI models up to eight times faster than the MI250X graphics card it currently offers and boost AI task performance per watt by five times.
"Genoa" is poised to become the most high-performance processor for general-purpose computing worldwide. With optimized 5 nm technology, it is expected to have up to 96 "Zen 4" cores and support DDR5 and PCIe® 5 technologies. In addition, "Genoa" will include CXL support, providing significant memory expansion capabilities for data center applications. Its launch is scheduled for 2023.
Specifically designed for cloud applications, "Bergamo" is equipped with 128 "Zen 4c" cores optimized for configurations with a large number of cores. It is programmatically compatible with the Zen 4 architecture and features all the same software and security options as "Genoa," in addition to being compatible with its connectors. Deliveries of "Bergamo" are scheduled for the first half of 2023.
The 3D chiplet design and massive transistor count are impressive on paper, but real gains depend heavily on software optimization and ecosystem maturity, areas where AMD historically lags.
The touted 5x perf/watt improvement could be smoke and mirrors without standardized benchmarks. Until we see broad adoption and third-party validation, these claims should be taken with a grain of salt.