Chip designer Arm Holdings is sharing details about its Neoverse V1 and N2 cores, introducing its new CMN-700 interconnect, and showcasing its partners’ plans to take the Arm Neoverse IP to market. The company says its Neoverse portfolio offers performance advantages for the datacenter, cloud and high-performance computing markets, as well as 5G and edge computing.
“As Moore’s law comes to an end, solution providers are seeking specialized processing. Enabling specialized processing has been a focal point since the inception of our Neoverse line of platforms, and we expect these latest additions to accelerate this trend,” stated Chris Bergey, senior vice president and general manager of Arm’s infrastructure line of business, in a company blog post, published today.
The Ministry of Electronics and Information Technology (MeitY) in India announced it will license the Neoverse V1 platform to support a made-in-India exascale project. MeitY joins two other partners who have made public their intentions to build Arm-based exascale-class HPC SoCs: SiPearl in France (part of the European Processor Initiative) and ETRI in South Korea.
Introduced in September, the Arm Neoverse V1 platform (codenamed Zeus) is the last of the Armv8 architecture line and the first Arm core to support Scalable Vector Extension (SVE). V1 is capable of delivering 50 percent more performance for HPC and machine learning workloads (based on IPC – instructions per clock), and offers a 1.8x improvement for a range of vector workloads over N1, according to Arm.
The Arm Neoverse N2 core (codenamed Perseus) is the newest Neoverse platform and the first to implement the Armv9 architecture. Compared to N1, N2 delivers 40 percent more single-threaded performance, according to Arm, along with a 1.3x improvement for high-throughput and hyperscale workloads.
Neoverse N2 introduces enhanced vector instructions, SVE2 – an Armv9 feature that supports vector processing for machine learning and digital signal processing applications. SVE2 builds on the Scalable Vector Extension (SVE) technology that Arm developed with Fujitsu for Riken’s Fugaku supercomputer, widening the range of use cases and also adding programmability and portability benefits, Arm says.
As explained by Arm, “the SVE design concept enables developers to write and build software once, then run the same binaries on different AArch64 hardware with various SVE vector length implementations. The portability of the binaries means that developers do not have to know the vector length implementation for their system.”
Arm is also announcing the CMN-700 interconnect (CMN stands for Coherent Mesh Network), which the company says is essential to building Neoverse V1 and N2-based SoCs, enabling use cases for multi-chip, memory expansions and accelerators. The follow-on to CMN-600, CMN-700 provides “a step function increase in performance on every vector, from core counts and cache sizes to the number and types of memory and IO devices you can attach,” stated Bergey.
“Through our continued investment in CCIX and CXL, we’re providing more customization options and enabling partner solutions with fast fabric and high core count scalability,” he added.
The CMN-700 can support up to 256 cores per die, pending TDP constraints and reticle limits. “In theory, [256 cores are] possible,” Bergey told reporters in a pre-briefing held last week. “It may make more sense to build multiple chiplets [in some cases], but from a mesh point of view, the support is there.”
MeitY, which promotes advanced technology as the engine for India’s empowerment, revealed its plans to establish exascale computing capability within India using an indigenously designed Arm SoC based on the Arm Neoverse V1 platform.
“The Arm Neoverse V1 platform…combines high-end CPU and vector performance with strong power efficiency, [giving] our SoC and System designers the freedom to optimize CPU, memory and IO capabilities for the unique requirements of exascale computing,” said Hemant Darbari, director general, Centre for Development of Advanced Computing (C-DAC).
The MeitY project joins other major national efforts to develop homegrown technologies for the exascale era, including SiPearl in France (part of the European Processor Initiative) and ETRI in South Korea. All three projects are licensing the Neoverse V1 IP, which supports the SVE instructions that helped make the Fugaku supercomputer the fastest in the world.
“In pursuit of advancing ABCA – Artificial Intelligence, Big Data, Cloud Computing and IoT – ETRI is developing the K-AB21 ‘artificial brain’ CPU based in part on the Arm Neoverse V1 core,” said Dr. Youngsu Kwon, director of the AI processor research department at ETRI. “Chosen for its high performance and low power consumption, the Neoverse V1 will help us achieve our goal of 2.5x performance with 60 percent power reduction over alternative HPC and AI accelerator technologies.”
Community support is building for Neoverse.
Oracle said last year that it plans to adopt Ampere Altra CPUs, which use Neoverse N1 cores, for its Oracle Cloud Infrastructure. The Arm-based 160-core bare metal instances are scheduled to debut in the first half of this year.
Marvell revealed its OCTEON DPU family of networking solutions based on Neoverse N2 will begin sampling by the end of 2021. The company projects a 3x generation-over-generation performance boost.
Tencent and Alibaba Cloud also announced support, with the latter reporting a 50 percent performance improvement for its DragonWell JDK on Neoverse N1.
Nvidia recently announced plans to build an Arm chip, called Grace, based on a future Arm core. Given the timeline – Grace systems are expected to debut in 2023 – the V1, N2 or next-generation “Poseidon” platform are all possible candidates.
Arm also touted its partnership with AWS and growing traction for the (Neoverse N1-based) Graviton2 instances within the EC2 footprint. Citing research by Liftr Insights, Arm said Graviton2 instances comprise 14 percent of the total number of instance types within EC2.
Addison Snell, CEO of Intersect360 Research, sees today’s news as another marker of Arm’s rise in HPC.
“x86 is naturally still dominant, but Arm is getting trial from over 25 percent of HPC users, with penetration highest in the commercial sector, according to our most recent survey,” he said. “It’s noteworthy that HPC and hyperscale are two of the four stated key target markets (along with 5G and edge computing), and Neoverse could therefore be well-situated for HPC in the cloud.”
Snell remarked that he was especially interested in the licensing announcement for the Indian exascale computing project at MeitY. He noted that while India has a poor track record at following through on its supercomputing plans, the announcement nevertheless builds on the continued interest in Arm at scale.
Last month, as part of the Neoverse Armv9 launch announcement, Arm CEO Simon Segars said that v9 based chips are on track to arrive early next year, with partners beginning to sample v9 parts later this year.
Arm Holdings is owned by Japan’s Softbank Group, which has agreed to sell the U.K. chip IP firm to U.S. chipmaker Nvidia for $40 billion. The deal is currently under review by regulatory agencies.