Low power processor design has been applied to arm with more than 180 billion chips. It began to explore the edge, cloud and 5g markets two years ago, and released the neoverse e series processors. Arm said at that time that the performance of this series of processors would achieve 30% per year by 2021. In 2019, the first generation of neoverse N1 will be launched, with a performance improvement of 60%, twice the target set in 2018.
This week, arm updated the neoverse product line again, releasing the second-generation N-series neoverse N2 and the new neoverse V1 platform.Compared with neoverse N1, neoverse N2 improves single thread performance by 40% while maintaining the same level of power and area efficiency. Also, compared with neoverse N1, the performance of neoverse V1 can be improved by more than 50%.
Is arm competing with X86 in high performance computing market?
Why does neoverse CPU emphasize single core performance?
Two years after the launch of neoverse series, it has three series, e series, n series and V series.Neoverse n series considers performance, power and area (PPA) at the same time, and is good at scalability; V series aims to provide the best performance and consumes more area and power; E series focuses on efficiency and optimizes power consumption and area reduction.
As a newly released product series, V series is more performance oriented. Therefore, neoverse V1 is designed for 7Nm and 5nm, and is the first to support scalable vector extensions (SVE). Based on the software programming model of unknown width vector unit, SVE can execute SIMD integers, bfload16 and floating-point instructions.
At the same time, V1 also supports the connection of pci5.0, ddr5, hbm2e and ccix 1.0, which can realize bidirectional consistent communication between chips between slots and small chips in package.
Chris Bergey, senior vice president and general manager of Arm Infrastructure Division, interviewed Lei Feng network (official account: Lei Feng net): "SVE is efficient in accelerating the HPC high-performance computing field or machine learning, and it is very friendly to software developers, and does not need the width of the vector. In the future, we will apply SVE technology to a series of cores other than N2 and v1. "
Based on ARM architecture, Japan's fugaku supercomputer uses 512 bit SVE technology. It is not only the fastest supercomputing in the world, but also the fourth supercomputer on the latest green500 list.
Another new product released by arm this week is neoverse N2, which is oriented to 5nm process design and supports PCIe 5.0 and ddr5. It is further expanded by supporting hbm3 for high bandwidth memory and ccix 2.0 and CXL 2.0 for architecture. It can support the design from 8 cores, 20W to 192 cores, 350W, and is a highly scalable platform across the cloud, smart NICs, enterprise network and edge devices with limited power consumption.
More details about the new V1 and N2 releases will be announced later, Chris Bergey said. Now we only know that V1 and N2 have improved the single core performance by more than 50% and 40% compared with the previous generation N1.
Why does neoverse CPU emphasize single core performance so much? Chris Bergey explained: "we think cloud service providers will prefer single threaded processor cores, because proximity, security, multi tenancy and other features can bring them better economy, which is why our products focus on single threaded technology. Of course, neoverse E1 also supports synchronous multithreading (SMT) technology. "
Neoverse's 30% performance improvement per year continues beyond 2021
Arm promises that the performance of the neoverse Series CPU will increase by 30% annually, which will last until 2021. Chris Bergey said: "of course, this is inseparable from the efforts and investment of the arm engineering team. I think it has a greater relationship with the gradual maturity of the software ecology. For example, many cloud native software can now run on the arm architecture more seamlessly."
In other words, there are similarities between cortex and neoverse products in the same period.For example, neoverse N1 and cortex-a76 were developed at the same time. There are many similarities between the two microarchitectures.
Neoverse V1 can be seen as a peer design of cortex-x1, and the two flagship CPUs may share many very large core architectures. Cortex-x1, a new product released in May this year, also pursues maximum performance.
Neoverse N2 is special because it represents the design of the next generation cortex-a, which is the follow-up product of A78 released in the same period with cortex-x1. This may be the key reason why the technical details of neoverse N2 can not be released yet. In order to more intuitively explain the relationship between cortex and neoverse product line, a picture made by foreign media AnandTech can help you understand it better.
In terms of software ecology, the key task of arm is the chip level interface of arm, which provides an opportunity to design system level solutions. Arm has invested in ccix and CXL, providing better interconnection technology, providing scalable switching network and supporting a large number of processor cores.
In addition, arm project Cassini hopes to be friendly to software developers through standards, platform security and reference implementation, so that industry partners can deploy and use the arm based platform.
In terms of infrastructure software including operating system and hypervisor support, Xen, KVM, docker container and more and more kubernetes have announced support for ARM architecture. Many of the early arm driven open source projects are becoming autonomous and commercial ISV applications are evolving.
Arm server CPU "rush" x86 cake, the most lack of time
But arm still has a long way to go to really shake x86's position in the cloud CPU market.Chris Bergey told Lei Feng: "the biggest challenge for arm in the cloud market is time. We see a lot of data center customers have great interest in turning to arm. However, from the IP of arm to the chip design, to the final deployment in the data center, and then to the support of the software ecosystem, it still needs a certain amount of time and the efforts of many industry chain partners. "
Of course, neoverse is for markets from the cloud to the edge.Chris Bergey also pointed out that arm's opportunities actually include the support of cloud native software, as well as helping customers overcome the challenges brought by Moore's law slowdown through customization.
Some people in the industry believe that one of the keys to the success of arm in the cloud market is to meet the customized needs of large Internet companies and cloud service providers.
"In heterogeneous computing, arm has a great opportunity, that is, how we can provide tight coupling or customization capabilities, or even integrate the requirements and technologies from the ecosystem, IP technology and cloud manufacturers through multi-core packaging or multi-chip assembly technology."Chris Bergey said customers who don't have strong design experience are more likely to choose the neoverse core.
Can arm architecture chip cover from terminal to edge to cloud in the future? Now it seems that in addition to arm's efforts, whether or not it will eventually be acquired by NVIDIA has become a key factor.