At Arm Vision Day 2021, the company showed off its vision for Arm computing in the next decade. Specifically, the company was out touting its next-gen Armv9 architecture. We did not get in-depth technical briefings since this was a higher-level event (in-depth we will cover later.) Still, Arm showed off a few features it is bringing out for the ecosystem in its next-gen IP.
Arm Vision Day 2021 and Armv9
Arm’s base architectures are designed to last some time. Not only will they ship in products for years, but they will often be deployed in the field where they may be expected to run for a decade or more. Arm said that its Armv8 powered the last decade, but its Armv9 is set to power the next decade.
At the event, the company introduced Armv9 which will bring about key advancements for machine learning, digital signal processing, and security.
One of the key drivers of Arm expecting to see massive shipment growth is the need for specialized compute. Or another way to look at this is that a number of traditional analog devices will convert to some level of “smart” and connected over the next few years. An example was given of a mechanical pump (like a water pump) that could be monitored for failure signs and efficiency versus just pumping water. For each of those applications, there will be different needs in terms of sensor connectivity and processing, general-purpose and accelerated compute (CPU and AI as examples), memory, and communications infrastructure. Arm sees the lower power cost of new chips enabling a wider array of chips and therefore more chips being sold.
As a quick aside here, this is not dissimilar to the Intel Xe HPC Ponte Vecchio Shows Next-Gen Packaging Direction and Pat Gelsinger’s vision as he starts to revamp the Intel Borg. Arm and Intel both see heterogeneous compute with different bundled IP as the way forward.
Another key push will be for Arm SystemReady. This is building on Arm ServerReady which helped Arm servers go from being a science experiment to boot each server to our experience with the Ampere Altra Wiwynn Mt. Jade Server where it worked (mostly) out-of-the-box using a standard image. We had to make one small tweak to get the performance we wanted, but otherwise, it worked. Arm is looking ahead and is attempting to break down maintainability barriers present in the Arm ecosystem by certifying base specifications.
If you have a pile of Arm development boards that you cannot load any OS you want or even just a new OS onto, you have seen this first hand. While Arm has been focused on performance, a lot of the Arm ecosystem is either supported by a sole vendor or is designed for “throwaway” technology. Try putting a current version of Android on a 2010 Arm-based phone versus Ubuntu on a 2010 x86-based PC and you will see the difference. Also, try just loading Ubuntu on an Arm M1 Mac Mini versus a Project TinyMiniMicro node and you will again immediately see the difference. Arm has focused on performance and efficiency, but now it is working on creating a useful legacy to an extent that the company has not focused on previously.
To that next-decade, the idea is that there will be an umber of new capabilities, but perhaps the most pervasive is going to be AI/ ML. When Arm is using “Machine Learning” here, it is mostly focused on inference. Local processing is going to be important as devices proliferate since otherwise network bandwidth scaling will not keep pace. As a result, Arm needs to enable AI inference on its devices that will be deployed at the edge.
Another major focus for its next-gen architecture is confidential computing. AMD powers Google Cloud Confidential Computing and Intel will make a big splash with this with its Ice Lake Xeons as well, but increasing on-chip security is important for the industry.
Arm will be adding this type of feature to its Armv9 in the Arm Confidential Compute Architecture or Arm CCA. Here, the company is using a “Realm” to perhaps harken back to the British Empire’s expansion and colonial era, while others are using terms such as secure enclaves.
The basic idea is that applications and services can be virtualized and prevent memory attacks. We hope to get more details on this since adding virtualization adds overhead which can be power expensive.
This is where the industry is heading, so it is good to see that Arm is helping its partners keep pace with the industry here.
Perhaps the most interesting slide from the presentations was this one:
Arm has a fairly interesting view of how different aspects of CPU design impact overall performance. While this may not be exact in every case, it is at least an interesting way to present its thinking.
Overall, this was a technically light ~3-hour event. The key takeaways are that Armv9 is coming with improved performance and security. We should start to see designs once Arm’s partners adopt the architecture and implement it into products. On the server-side, this announcement is akin to how Intel started talking about Sapphire Rapids some time ago, yet we will not see chips for a little bit longer. Arm does not make chips, but it helps to make the IP that partners can leverage to build chips, so Arm’s releases are usually ahead of product launches.