AMD gave its data center processing portfolio an EPYC refresh this week with the launch of its Milan-X CPUs and Instinct MI200 artificial intelligence (AI) accelerators. The chipmaker also teased its fourth-generation Genoa and Bergamo processor families.

Milan-X is based on AMD’s existing EPYC 3 platform, announced this spring, and is its first server processor to implement 3D V-Cache. The technology stacks a static random-access memory module directly on top of the processor die, increasing performance up to 50% in “targeted technical computing workloads,” the company claims.

“These processors have three times the L3 cache compared to standard Milan processors. At the top of the stack, that adds up to 804 megabytes of total cache per socket,” CEO Lisa Su said during her keynote at AMD’s Accelerated Data Center virtual event. “This additional L3 cache relieves memory bandwidth pressure and reduces latency and in turn speeds up application performance dramatically.”

AMD reports Milan-X will be drop-in compatible with existing server platforms.

The new chips are available now in private preview in Microsoft Azure’s high-performance compute (HPC) virtual machines and are expected to see public rollout later this year. Milan-X will be generally available via AMD partners including Dell, Cisco, Lenovo, Hewlett Packard Enterprise (HPE), and Supermicro in early 2022.

“Milan-X shows the company is smartly investing in areas it sees an opportunity to extend an already strong leadership position,” Matt Kimball, senior analyst at Moor Insights and Strategy, wrote in an article for Forbes. “The performance gains alone should drive even greater EPYC adoption.”

The EPYC refresh comes as AMD scores another victory over rival Intel. Meta, which recently changed its name from Facebook, this week announced the development of a new, more power-efficient server platform powered by team red’s EPYC 3 processors.

Alongside Milan-X, AMD launched its second generation of Instinct data center GPUs, which it claims are the world's fastest AI accelerators.

AMD's MI200-series features two GPU dies with a total of 58 billion transistors based on a 6-nanometer manufacturing process. This translates to 220 compute units and 880 matrix cores, a 1.8 times increase over the previous generation MI100.

The GPUs are fed by up to eight on-package, high-bandwidth memory modules for a total of 128 gigabytes of total memory.

“AMD appears to be putting a direct target on the Nvidia A100 GPU,” Kimball wrote. “If one looks through the Top500 and Green500 lists, one will notice that the A100 enjoys a prominent, if not prolific standing in these areas … It seems fair to say that AMD has noticed these numbers and sees an opportunity.”

In fact, AMD's Forrest Norrod, SVP and GM of data center and embedded solutions at AMD, claimed the MI200 is 490% faster than Nvidia's A100 in highly precise floating-point 64 calculations.“This is critical for HPC workloads requiring the highest level of precision, like weather forecasts,” he said.

AMD’s MI200-series chips are available now in a compact original equipment manufacturer form factor similar to Nvidia’s A100 GPU. The GPU will also be available in a slimmed-down PCIe form factor in the near future.

The U.S. Department of Energy’s Oak Ridge National Laboratory plans to use the new GPUs alongside AMD’s EPYC 3 processors in its Frontier supercomputer. Built in collaboration with HPE, the supercomputer is expected to deliver 1.5 exaFLOPs of computing power when it comes online early next year.

“The HPC market is not only lucrative, but also a proving ground for mainstream server adoption,” Kimball said of AMD’s recent success in the HPC space. “Combined with the company’s success in the cloud, the sky could be the limit for EPYC if AMD plays its data center marketing cards right.”

AMD also teased its upcoming EPYC 4 processor family, codenamed Genoa and Bergamo.

Genoa will offer 96 cores based on Taiwan Semiconductor Manufacturing Co. (TSMC)’s 5-nanometer manufacturing process when it launches next year.

“We’ve worked with TSMC to optimize 5 nanometers for higher performance computing, and it offers twice the density, twice the power efficiency, and 1.25 times the performance of the 7 nanometer process we’re using in today’s products,” Su said. “When it’s introduced, we expect Genoa will be the world’s highest performing processor for general purpose computing.”

The CPU is AMD’s first to support Compute Express Link, DDR5 memory, and PCIe Gen 5.0 connectivity.

However, AMD isn’t stopping at 96 cores. Bergamo will feature up to 128 Zen 4c cores and is specifically aimed at cloud-native workloads. The chip is slated for release in early 2023.