News

Data Center HPC Zooms Forward with AMD’s “Worlds Fastest” Accelerators

November 11, 2021 by Jake Hertz

At this year’s Accelerated Data Center Premiere event, AMD announced a new family of accelerators, the Instinct MI200 series, that claims to literally speed up high-performance computing (HPC) in data centers.

The role of the data center and cloud computing is increasing significantly with new applications like AI, and even more recently, “the metaverse” (or AR/VR). 

Hoping to fuel these growing applications, hardware companies have invested heavily to develop new computing resources for HPC in the data center. 

One of the companies at the forefront of this development is AMD, and this week they held their virtual Accelerated Data Center Premiere. At this event, AMD discussed a plethora of new developments that they’ve been working on, the highlight of which is its new AMD Instinct MI200 series accelerators

 

An image of AMD's latest Instinct accelerators. Image used courtesy of AMD

 

In this article, we’ll look deeper at AMD’s new family of HPC accelerators and cover some of the innovations that the family is trying to bring. 

 

The Instinct MI200 Series 

The most notable announcement from AMD’s event earlier this week was releasing a new family of HPC accelerators, the Instinct MI200 Series. According to AMD, the MI200 series benefits from three major areas of innovation. 

First, the new series is based on the new and improved 2nd Gen, CDNA 2 architecture. This architecture works to improve computational capabilities and throughput via improved Matrix Core Technology. 

Building on the 1st Gen architecture, the 2nd Gen enables support for a wider range of data types and applications such as FP64 for scientific computing. Even more granularly, the architecture consists of 880 matrix cores where the compute unit (CU) arrays are split into four “shader engines.”

 

A block diagram for the CDNA architecture which offers up to 8 Infinity Fabric links in the MI200 series.

A block diagram for the CDNA architecture which offers up to 8 Infinity Fabric links in the MI200 series. Image used courtesy of AMD

 

Second, the new architecture leverages AMD’s 3rd Gen Infinity Fabric technology. Infinity Fabric is AMD’s proprietary interconnect architecture that allows for improved communication between multiple blocks of a heterogeneous computing system, including CPUs and GPUs. 

Whereas the previous generation relied on standard PCIe for communication to the host processor and only had three Infinity Fabric links for GPUs, the MI200 Series offers up to eight Infinity Fabric links for connection to the host processor and GPUS. 

Lastly, AMD claims advances in packaging technology have been a huge boon for the MI200 Series. AMD states that the new series introduces the industry’s first multi-die GPU design with 2.5D Elevated Fanout Bridge. 

Altogether, the MI200’s packaging is said to deliver approximately 2x more cores and roughly 3x higher memory bandwidth than previous-gen GPUs.

Now that the general concept of these accelerators is understood, let's take a look at the specs for the MI250X, specifically.

 

Instinct MI250X's Specs

The flagship product in the new MI200 series is the MI250X, which AMD claims to be the world’s fastest HPC and AI. 

Featuring 220 compute units and 14,080 stream processors, the Instinct MI250X is said to deliver 4x improvement over its most competitive GPUs. 

For double-precision FP64 and FP32, the processor tops out at 47.9 TFLOPs, and, thanks to the new FP64 Matrix Core technology, it achieves up to 95.7 TFLOPs double precision (FP64 Matrix) peak theoretical performance. When scaled down to FP16, the GPU reaches 383 TFLOPs peak theoretical performance. 

 

The M1250X performance versus NVIDIA’s A100.

The MI250X performance versus NVIDIA’s A100. Image used courtesy of AMD

 

In terms of memory performance, the MI250X claims to offer the industry’s best aggregate peak theoretical memory bandwidth of 3.2 terabytes per second, a number representing a 2.7x improvement over previous generations. Additionally, this GPU comes with a memory size of 128 GB HBM2e and a memory clock speed reaching up to 1.6 GHz.

 

Big Releases From AMD 

Whenever someone releases a new product that claims to be “the industry’s fastest,” it’s always exciting news. With the new Instinct MI200 GPU family from AMD, the company looks to have positioned itself as an undeniable industry leader in the HPC space. This innovation could be a good place for HPC to get a boost to the next level.