On March 31, ARM officially released the Armv9 architecture, which focuses on improving "security, AI and performance per watt". The last generation of Armv8 architecture (which we are using) has been released in October 2011, and this decade is the decade when the Armv8 architecture has conquered cities and territories. Arm claims that Armv9 will serve the next generation of 300 billion ARM chips.

CPU performance: 30% increase in two years

First of all, we are most concerned about the most practical performance, which may disappoint everyone. The performance improvement of Armv9 will not be as big as that of V8, which introduced the AArch64 instruction set that year.

The new CPU architecture is expected to be released this year and will not appear on commercial devices such as mobile phones until early 2022. In the roadmap released by Arm for the next two years, it is expected that the performance of the next two generations of mobile CPUs will be improved by more than 30%. That is to say, for the CPU cores of the next two generations: Matterhorn and Makalu, the same frequency IPC will increase by about 14%.

Specifically, it includes reducing the memory latency from 150ns to 90ns (every 5ns increase brings 1% performance improvement), increasing the frequency from 2.6GHz to 3.3GHz, increasing the bandwidth from 20GB/s to 60GB/s, and doubling the L2 and L3 caches.

Safety performance: CCA

Back to the official statement, Armv9 focuses on improving the security architecture. Armv9 introduces a security architecture called CCA, which has the technology of "memory tag extension", making it more difficult for hackers to take advantage of time and space vulnerabilities in hidden code.

Another new function, "Realms", can dynamically create a memory partition "domain", set a completely opaque security container execution environment for the system and virtual machine management program, and isolate other VMs. It claims that even if the mobile device's system is "hacked", the new architecture can continue to protect data security.

AI performance: SVE2 instruction set

The second is to improve AI performance. Armv9 integrates the SVE2 instruction set (full name Scalable Vector Extensions, hard translation is a new extensible vector extension technology), claims to improve the performance of machine learning (ML) and digital signal processing (DSP), and improve the performance of 5G, virtual reality (VR) and augmented reality (AR).

SVE2 was released in April 2019, and the predecessor of the instruction set is the SVE jointly developed by Arm and Fujitsu, which has been used on the A64FX CPU of the world's first supercomputer.

Summary

Both security and AI performance are required by server and mobile device users, but they are both at the bottom level. The user perception is that there is no obvious improvement in CPU performance.

Since the A76 architecture, the CPU increase speed of Arm has slowed down. Unless Apple is squeezing toothpaste in the next two years, it may not be possible to see the future of Android camp catching up with Apple's A-series chips in terms of single core performance

Follow our Weibo @ Love Computer

Follow our WeChat official account: playphone

Of course, we also pay attention to our Bilibili account: love computer