A few months after presenting the Neoverse N2 and V1 platforms, Arm highlights their performance and functionality. The chip designer also touts the growing list of processors using the Neoverse family of products to bring specialized processing to the cloud, HPC and beyond.
The next generation of network processors from Marvell’s Octeon family will be based on Neoverse N2, delivering three times the performance of the previous generation. Sampling will start at the end of 2021.
Good news for the ecosystem revolving around Arm’s offerings. As a reminder, Tencent is adopting Neoverse technology for cloud applications, while Alibaba Cloud has tested the next Alibaba Cloud ECS Arm instances, showing a 50% performance improvement of the JDK DragonWell on Neoverse N1. For its part, Oracle adopts Ampere Altra processors, with up to 160 Neoverse N1 cores, for Oracle Cloud Infrastructure and Graviton2 from Amazon Web Services (AWS), based on Arm, extends Arm’s footprint in the cloud.
Two new platforms
When Amazon launched Graviton 2, “what surprised was not just the fact that Arm could make a convincing heart,” says Chris Bergey, one of the Arms officials interviewed by , but also “The fact that AWS can put 64 of these cores and get consistent performance on scale-out workloads.” This consistent performance, he adds, is in large part attributable to Arm’s CMN-600 mesh interconnect.
Arm now presents the CMN-700, a key component for building better performing SoCs based on Neoverse V1 and Neoverse N2. In the meantime, the Neoverse V1, as previously announced, offers a 50% increase over the Neoverse N1 and a four-fold improvement for machine learning workloads. The Neoverse V1 is the largest microarchitecture ever produced by Arm, intended for HPC and exascale computing.
The Neoverse N2 platform, for its part, is the first based on the Armv9 architecture. It offers 40% better single-wire performance than the N1, while maintaining the same level of efficiency in terms of power and surface area as the N1. For high throughput computation, such as in hyperscale clouds, it offers a 1.3 times improvement on NGINX compared to N1. For use cases with power and space constraints, such as edge and 5G use cases, it offers 1.2 times faster DPDK packet processing than N1.