Arm talks 40% and 50% better performance from 2 new server chips

Arm Holdings has disclosed details of its two new server-processor designs, Neoverse N2 and Neoverse V1, as well as an updated high-speed mesh to connect its processors.

The two designs were introduced last September but Arm was mum on performance. Now it’s talking numbers.

The Neoverse V1 is designed for scale-up servers, especially high-performance computing (HPC). It supports for Scalable Vector Extension (SVE) and delivers more than a 50% performance increase over the N1 for HPC machine-learning workloads.

It also supports bfloat16, an extension that has become almost standard instruction set for AI and machine learning, along with PCI Express Gen5, DDR5 memory, and HBM memory as well. It will be built on 7nm and eventually 5nm fabrication designs.

“With V1, we’ve targeted the highest performance and widest execution units to take on the highest end of workloads that infrastructure customers are looking to deploy,” Dermot O’Driscoll, vice president of product solutions in Arm’s Infrastructure Line of Business said.

The Neoverse N2 platform uses the new Armv9 architecture that the company recently announced and claims can deliver 40% more single-threaded performance for a variety of workloads vs. the N1, but still retains the same level of power and area efficiency as N1.

“With N2, it’s a slightly more moderate design in terms of performance, but that allows you to build a higher number of them. So when we talk about scale out, we’re talking about customers building in excess of 100 cores per monolithic SOC,” said O’Driscoll.

Arm says it already has several customers lined up for the N2 and V1 as well as N1 wins:

• Marvell announced its OCTEON family of networking solutions based on Neoverse N2 will begin sampling by end of 2021, providing a 3x performance uplift over previous generation OCTEON solutions.
• India’s Ministry of Electronics and Information Technology (MeitY) has announced it will join SiPearl and ETRI in licensing Neoverse V1 for its national exascale HPC project.
• Oracle plans to adopt Ampere Altra CPUs for Oracle Cloud Infrastructure.
• Arm-powered AWS Graviton2 continues to rapidly expand its EC2 footprint with steady growth and regional expansion.
• Alibaba Cloud just tested the upcoming Alibaba Cloud ECS Arm instances using Neoverse N1.
• Tencent is making investments in both hardware testing and on software enablement that will allow them to adopt Neoverse technology for cloud applications.

The third bit of news is the Arm Neoverse CMN-700 mesh interconnect for the Neoverse V1 and N2 platforms. CMN-700 is an evolution from the CMN-600 that is used in data centers, and there are some sizable bumps in capacity.

Its total core support has been increased 4x from 128 cores in the CMN-600 to 512 in the 700, system level cache per die is up 4x from 128MB to 512MB, the memory-device port count has gone from 16 to 40, and support for CXL interconnects has been added.

CXL, or the Compute Express Link protocol, is gaining popularity as an alternative to PCIe because it works with PCIe but also acts as a unifying mesh between processor cores and memory pools. PCIe is good for point-to-point transfers but is not a mesh design.

Arm expects products based on N2 and V1 to hit the market by the end of the year.