AMD Unveils Multi-chip GPU for Next-gen AI

AMD this week unveiled its third generation of Epyc server processors along with a new Instinct MI200 family of accelerators that offers a major upgrade over the previous Instinct MI100 and which AMD bills as offering “exascale-class” performance.

Faster than the competition

The AMD Instinct Mi250X provides almost five times faster performance than Nvidia’s A100 GPU for double precision (FP64) HPC applications, and more than 380 teraflops of peak theoretical half-precision performance to enable data-driven research.

The A100 is an enterprise-level GPU that is available for purchase or can be accessed in the cloud on platforms such as the Cloudera Data Platform.

Under the hood, the MI200 series accelerator features AMD’s second generation CDNA architecture, as well as new packaging technology that makes it the industry’s first multi-die GPU.

Specifically, the MI200 adopts the “chiplet” approach used so successfully with its Epyc server processors and which gave it a performance edge over Intel’s Xeon chips. In this case, AMD leveraged the design to pack two Graphics Compute Die (GCD) in a single multi-chip module.

Connectivity is achieved through multiple 25Gbps “Infinity Fabric” links for up to 100GBps of bi-directional bandwidth between the GPUs.

AMD unveiled three MI200 series accelerators: The top-end MI250X, the MI250, and the PCIe-based MI210. The first two models adopt the Open Compute Project’s OCP Accelerator Module (OAM) form factor. This allows it to draw as much as 700 watts for a single card, and gives it twice as many data links as can be achieved with a standard PCIe card, notes tech site Anandtech.

“AMD Instinct MI200 accelerators deliver leadership HPC and AI performance, helping scientists make generational leaps in research that can dramatically shorten the time between initial hypothesis and discovery,” said Forrest Norrod, senior vice president and general manager of the Data Center and Embedded Solutions Business Group at AMD.

“With key innovations in architecture, packaging and system design, the AMD Instinct MI200 series accelerators are the most advanced data center GPUs ever, providing exceptional performance for supercomputers and data centers to solve the world’s most complex problems,” he said.

The MI250X and MI250 accelerators are currently shipping, and the MI210 will be available in the future. It is understood that the new Instinct MI200 Accelerators and third-gen EPYC CPUs will be used to power a new supercomputer for the US Department of Energy and is scheduled to come online next year.

Image credit: iStockphoto/hernan4429