At the tail end of AMD’s 75-minute CES keynote, held yesterday evening in Las Vegas and via livestream, CEO Lisa Su shared new details of the forthcoming MI300 datacenter chip and publicly unveiled the silicon. The MI300 (teased earlier this year) is the first to combine a CPU, GPU and memory into a single integrated design, incorporating nine 5nm chiplets that are 3D stacked on top of four 6nm chiplets with 128 gigabytes of HBM3 memory.
The device has more than 146 billion transistors and is “the most complex chip ever built by AMD,” according to Su. It brings together AMD’s next generation CDNA 3 GPU architecture, optimized for HPC and AI performance, with 24 Zen 4 cores, leveraging AMD’s fourth generation Infinity architecture.
The AMD Instinct MI300 accelerator is designed for power-efficient AI training performance and HPC workloads, said Su. The follow-on to the Instinct MI250X GPU – the computational workhorse of the Frontier exascale supercomputer – the MI300 delivers the next step function in performance, “to enable the next generation of HPC and AI,” Su said.
“Bringing together a datacenter CPU, GPU and memory into a single integrated design allows us to share system resources – memory and IO – and it makes it much easier to program,” she added.
According to AMD testing, the MI300 delivers eight times more performance and five times better efficiency than its predecessor, the MI250X. “To put this in perspective, over the holidays, there’s been a lot of talk about ChatGPT, and what you can do with these large language models. [These models] take months to train on 1000s of GPUs that consume millions of dollars of electricity. MI300 can reduce the time to train these models from months to weeks, with dramatically lower energy costs. And more importantly, it can also support much, much larger models that can be used for even more advanced and more powerful AI services in the future,” said Su.
The MI300 is currently in AMD’s labs. It will be sampling soon and is slated to ship in the second half of 2023. The chips will power El Capitan, which is projected to be the world’s most powerful supercomputer (providing more than 2 exaflops peak performance) when it is fully deployed in 2023 at the DOE’s Lawrence Livermore National Lab. El Capitan is one of a cadre of three exascale supercomputers being acquired by the U.S. Department of Energy under its CORAL programs for ~$600 million each. Frontier (an HPE/AMD system) was deployed at Oak Ridge National Lab last year and Argonne National Lab is expecting completion of the Aurora supercomputer (an Intel/HPE system) this year.
Su underscored the company’s successes in HPC and the larger datacenter space. “We are very proud of our leadership in HPC,” she said. The world’s fastest supercomputer and the first to surpass the Linpack exascale milestone, Frontier, is powered by AMD Epyc third-generation CPUs and Instinct GPUs, and 75% of the world’s top 20 most energy-efficient supercomputers are powered by AMD.
Beyond goverment HPC, AMD technology is deployed at major cloud providers and across a range of industries. Notably, Epyc CPUs were used to render many of the visual effects for the blockbuster film Avatar: the Way of Water. Artists at Weta FX selected AMD Eypc CPUs out of available solutions, citing superior performance per watt per cabinet across their workloads. “Artists can now see their scene rendered live in front of them, whereas before that would often have to go off and be on a render wall,” said Kimball Thurston, senior engineer, Unity/Weta Digital.
Su closed her keynote with a call for collaboration. “The way we look at it is we as an industry have to come together with our partners and with the ecosystem to really solve some of the world’s most important challenges. I am so happy and honored to be in this industry. This is the best time to be in semiconductors. It’s the best time to be in tech, and I love what we’re doing together.”
Header image: AMD’s Lisa Su debuts the MI300 chip at CES in Las Vegas.