Ten years ago, Arm introduced a major revision in the architecture of its processor cores. That upgrade propelled Arm down a path of phenomenal success. The company just announced its first new major architectural revision since then, which the company hopes will be good for another decade of expansion.
Arm’s actual announcement was a little light on details, but the company did make some details available, and we happen to know a guy who went through them. This week we examine Arm’s new roadmap for its next 10 years with Tirias analyst Kevin Krewell.
Back in the 1980s, consumers knew the brand names of the computers they bought: IBM and Apple, Commodore and Compaq, Amstrad and Sinclair. The average consumer had never heard of the companies that made the microprocessors that were the engines enclosed in the cabinets of those still-mostly desktop machines – companies like Motorola, Zilog, MOS Technology. That changed in 1991 with Intel’s “Intel Inside” advertising campaign. Pretty soon, average consumers were talking about microprocessors the way their parents might have talked about car engines back in the 1960s. People became aware of what was under the hood.
早在1980年代，消费者就知道他们购买的计算机的品牌名称：IBM和Apple，Commodore和Compaq，Amstrad和Sinclair。普通消费者从来没有听说过制造微处理器的公司，这些公司的产品是被封闭在大多数台式机主机中的引擎，例如摩托罗拉，Zilog和MOS Technology。1991年，随着Intel的“ Intel Inside”广告活动，这种情况发生了变化。很快，普通消费者就开始谈论微处理器，就像他们父母在1960年代谈论汽车引擎的方式一样。人们开始意识到背后的东西。
Arm, today, is a little like Intel before 1991, in that the average consumer hasn’t heard of the company that almost certainly designed the processor in the smartphones that the majority of people on the planet have in their pockets or purses.
About 10 years ago, Arm introduced version 8 of its architecture. It was a fairly big deal. The company announced some key innovations for v8, and also plotted out a roadmap that built out on those innovations. The announcement, combined with the roadmap, built excitement around Arm’s processor cores, which the company rode to overwhelming dominance in the smartphone market, and to solid success in other markets such as automotive and the Internet of Things, among others.
Recently, the company held a virtual conference and introduced version 9 of its architecture. No one can doubt that Arm wants the introduction of v9 to be as big of a deal as the introduction of v8 a decade ago. Arm was a little light on details about v9, but to be fair, it was a little light on details back in 2011 on v8, too.
Kevin Krewell is a principal analyst with Tirias Research. He’s an engineer who in the past has worked at Nvidia and AMD, and he’s been a regular guest on this podcast. We know he knows his stuff. If anyone was going to ferret out whatever details there were about Arm v9, it was going to be our old pal Kevin. I called him up earlier this week to see what he was able to unearth about Arm v9.
Kevin Krewell是Tirias Research的首席分析师。他是一位工程师，过去曾在Nvidia和AMD工作过，并且一直是我们播客的常客。我们知道他懂这些。如果有人能揪出任何有关于ARM V9的细节，那就是我们的老伙计Kevin。我在本周早些时候打电话给他，看看他能够挖出Arm v9的哪些不为人知的秘密。
What details were you able to ferret out that suggest the differences between v8 and v9?
KEVIN KREWELL: Using my technology and analyst skills, I dug deep into their web site.
BRIAN SANTO: You bloodhound, you!
KEVIN KREWELL: Actually, there are lots of details on the Arm developer web site developer.arm.com going into and breaking it down in a little more detail. Arm was forthcoming about the fact that this was a “vision” that they were laying out for the architecture for v9.
V8 was a huge change for Arm. That introduced the 64-bit extensions to Arm. That was a massive change for the Arm architecture from 32-bit now to 64-bit and 32-bit compatible. So it’s really hard for Arm to follow up with that. Arm v9 (and in this case, just specifically the A profile, which is the cortex A, a line used in smartphones and servers and such). There are some big changes in there, but maybe not quite a big change as it was from 7 to 8. And in fact, what I found going through the web site is that a lot of the features that are showing up in Arm 9.0 and 9.1 will also be back-ported to future versions of the Arm 8 architecture.
V8对Arm来说是一个巨大的改变。这为Arm引入了64位扩展，从32位到兼容64位和32位的Arm架构，这是一个巨大的变化。因此，Arm很难跟进此事。Arm v9（在这种情况下，仅是A系列，cortex A，这是智能手机和服务器等中使用的产品线）有一些大的变化，但可能不像从7到8变化那么大。实际上，我浏览该网站，发现Arm 9.0和9.1也将反向移植到Arm 8架构的未来版本。
Some of the stuff is exclusive to Arm 9, but there’s a lot that they mentioned in the vision conference that is also going to be available on Arm v8. So they’re not necessarily making everything exclusive to Arm v9.
有些功能是ARM 9独有的，但还有很多他们在发布会上说到的功能也将会在Arm V8中实现。因此，他们并不一定要使Arm v9每个功能都独有。
The things that are exclusive to Arm v9 include Salable Vector Extension 2 (SVE-2), which is now a variable with vector extension, instead of the SVE-1, which was used in the supercomputer from Japan.
Arm v9专有的功能包括Salable Vector Extension 2（SVE-2），它现在是具有矢量扩展名的变量，而不是日本超级计算机中使用的SVE-1。
But SVE-2 is more dynamic in terms of its variable width. So it will scale from lower end to higher end solutions.
Transactional memory extension, which is used for writing highly parallel code, making it run more efficiently. I think that will show up mostly in servers.
Something called Branch Record Buffered Extension. That’s coming later for the Arm v9 instruction, but it captures branches and allows it better debugging. And then something that they really focused heavily on, which is this realm management platform.
There are extensions to that for confidential computing. And that’s going to come sometime in the future. So that’s actually not going to be available in the Arm 9.0 release. But the realm security management extensions are Arm’s attempt to really go in to enterprise class of computing platforms with an even higher level of security. So I think that’s one of their big separators. And that will be available in 9, but not back-portable to v8.
BRIAN SANTO: These are not insignificant technological advances. Some of them are very significant. But they’re not a clear break, like 32-bit to 64-bit. They’re spread across v8 and v9. But it sounds as if they’re signaling that, moving forward, they’re going to have a set of features and a direction that will support a new set of applications or one that they haven’t been as successful in: that enterprise area that you just mentioned before. Might that be a reasonable way to look at what Arm’s doing?
KEVIN KREWELL: I believe so, yeah. They’re moving the architecture forward, and I think that’s another reason to use the increment, the numbering scheme. There are definitions for Arm v8 up to Arm v8.7 on their web site with details. And some of the stuff like B-Floyd 16, which is used for machine learning applications, and that’s going to be in 8.6.
KEVIN KREWELL：是的，是的。他们正在推动架构的发展，我认为这是使用增量的另一个原因，即编号方案。网站上有Arm v8到Arm v8.7的定义以及详细信息。还有一些像B-Floyd 16，这是用于机器学习应用，这就是将出现在8.6中。
Highprecision timers. There’s a number of features. Generalized matrix multiplication, GMM. That’s in 8.6 as well. But that will also be in 9.1. So there are a number of overlaps, and I think these are high-performance compute functions that could be back-portable to 8. But when you start talking about “enterprise class,” I think Arm wanted to elevate the architecture to another level. And I think that by going from v8 to v9, they’re signaling that the Arm architecture is evolving to another level.
So I think the target markets here are high-performance computing applications and also enterprise class computing applications. And I think the security features, the scalable vector extensions give you more high performance compute on the platform. The transactional memory is another unique feature that can be useful in increasing performance in data center applications. So I think they are elevating the architecture to a new level and to bring it into direct competition and compatibility with enterprise class security, enterprise class computing.
BRIAN SANTO: These are some of the extensions paralleling what some of their competitors are doing in the same area.
KEVIN KREWELL: The goal here in the enterprise class is to compete with Intel Xeon, AMD Epic and even IBM Power. They’re bringing the architecture up to that level of performance and security. Arm has evolved over the years, and I think it wants to be considered a first-class architecture to any other architecture in the data center. And I think this is their play for that. This is their movement into that arena.
KEVIN KREWELL：企业级的目标是与Intel Xeon，AMD Epic甚至IBM Power竞争。他们正在将架构提升到该级别的性能和安全性。多年来，Arm不断发展，我认为它应该被视为数据中心中的一流架构。我认为这是他们的目标。这是他们进入那个舞台的动力。
BRIAN SANTO: With the performance improvements… They’ve always hit on their ambition to get into those areas with the thinking that their expertise in low-power processing seemed to fit hand in glove with what data centers wanted, but perhaps they might not have had the horsepower with their processors before. And perhaps also they might not have had a clear roadmap for where they were going with performance improvements, which data centers love to see. And it looks like they’re providing both with this transition to 9.
KEVIN KREWELL: Yeah. Initially, the focus on power efficiency is a holdover from their smartphone and tablet business. Power efficiency was their dominant theme. But then with the advent of neoverse set of cores, and even the Arm X1 core, they have really moved to step up performance as a lead feature. Still power efficient, but more on the performance side, not just necessarily performance per watt. Just to accelerate that.
KEVIN KREWELL：是的。最初，对功率效率的关注是其智能手机和平板电脑业务的后遗症。节能是他们的主要主题。但是随着新的内核集的出现，甚至是Arm X1内核，它们确实已经开始提高性能，并成为领先的特性。仍然具有高能效，但在性能方面更多，而不仅仅是每瓦性能。只是为了加速。
Part of it is, I think the competition from even Arm licensees like Apple, was done with very performant cores within the Arm architecture, that they would like to offer something from the standard Arm core that would be comparable to what Apple has developed. Qualcomm has bought Nuvia to also step up its performance to compete with Apple.
The cores themselves have now moved from just power efficient to performance, still with power efficiency. Then adding all these other features, security features, that are required for enterprise class. One thing we can’t forget is that Nvidia is trying to buy Arm at this point in time. So Nvidia’s focus and Arm’s focus are now coming into alignment. Nvidia’s focused on enterprise, it’s focused on AI, it’s focused on high-performance compute. So now we’re seeing a further alignment of the corporate goals of Arm and the corporate goals of Nvidia prior to an acquisition taking place. So I think that also could be part of the push for this Arm v9 at this point in time.
BRIAN SANTO: We’ve talked a little bit about performance. We’ve talked a little bit about the AI. Is there anything more about those before we move on to a couple of the other points that Arm CEO Simon Segars was hitting on at the introduction? He was talking about confidential computing and IoT security. Before we move to those, anything else about AI or performance?
BRIAN SANTO：我们已经谈了一些性能。我们已经讨论了一些有关AI的内容。在我们继续介绍Arm首席执行官Simon Segars在介绍中提出的其他几点之前，还有其他内容吗？他正在谈论机密计算和物联网安全性。在我们转向这些之前，还有其他关于AI或性能的问题？
KEVIN KREWELL: One of the things that Arm is doing is rolling out AI-performant features throughout the entire Arm cortex family. That’s important because, actually I’m a believe that AI workloads are going to be prevalent in everything, from microcontrollers all the way up to servers. So therefore, having some level of AI acceleration in your CPUs and your GPUs and your DSPs and other accelerators is going to be one of the megatrends for CPU architecture and architecture going forward.
KEVIN KREWELL：Arm正在做的一件事是在整个Arm cortex家族中推广AI性能功能。这很重要，因为实际上，我相信从微控制器一直到服务器，人工智能工作负载将在所有领域中普遍存在。因此，在您的CPU，GPU，DSP和其他加速器中实现某种程度的AI加速，将成为CPU架构的大趋势之一。
Intel’s doing it with DL Boost, their architecture. You’re going to see other architectures step up. Certainly RISC-V is adding their own vector extensions as well and focused on AI. So Arm is joining this wave of AI everywhere. So I see that as an important part of V9 as well.
英特尔通过其架构DL Boost来实现这一目标。您将看到其他架构也在涉足了。当然，RISC-V也将添加自己的矢量扩展，并将重点放在AI上。因此，Arm加入了AI everywhere的浪潮。因此，我认为这也是V9的重要组成部分。
BRIAN SANTO: So AI everywhere. That is a conceptual leap to the next thing, which is security everywhere. They were talking about confidential computing, and they were also talking about IoT security. And the two things are different. I have an intuitive grasp of why, but I’m wondering if I could get you to give up a sense of where confidential computing fits and where IoT security…. where those two things diverge from each other.
KEVIN KREWELL: In many ways, they’re all similar. You want to keep your data encrypted and safe. In confidential computing, it’s important that the data in a data center (whether it’s in a virtual machine or running on bare metal), the data at rest, the data in storage and the data in compute, in movement, is all protected. In fact, that’s part of the goal of confidential computing. AMD’s involved in it; IBM’s involved in it; Intel’s involved in it. It’s a major data center and enterprise class standard going forward to make sure that all the data is protected no matter where it is and what you’re doing to it.
On the IoT side, you see that, but on-the-fly encryption/decryption or working on encrypted data is a little more complicated, a little harder to do. So there, mood of trust is important. The communication… almost all the IoT or edge devices have to communicate back to a host of some sort. And in transaction, that communication of data, that needs to be protected along the way. So I see it a little bit different: You don’t necessarily have to worry about large data files on an IoT device, but you need the fact that the communication links in and out of the device and the boot are all protected.
But in a data center, the data in flight and transaction on a data path within the data center need to be protected as well. Little more detail, a little more deeper level of security.但是，在数据中心中，也需要保护数据中心内数据路径上正在运行的数据和事务。细节多一点，安全级别更深一点。
BRIAN SANTO: It’s really easy for gearheads to get attracted to the performance numbers, to the number of transistors, to whether this is implemented in chiplets or not. But the security and the confidential computing… do they actually show up as an “architectural” feature? Either way, it seems like neither of these is unimportant. It just that they don’t seem to have a clear physical engineering implementation yet.
KEVIN KREWELL: Going back to the v9, the first silicon implementation of v9 is supposed to come out later this year. And I believe Media Tech is going to be one of the first that’s out. So what I assume is, it’s going to be a performant core with probably SVE-2 on it. Actually, this is one of the big differences from Arm now. The last time when they talked about v8, when they released it, we didn’t see silicon for a while. I don’t remember exactly, but the first actual implementation of 64 bit was the Apple A series, ahead of everybody else. Arm itself didn’t even have a 64-bit core when Apple released their 64-bit core.
KEVIN KREWELL：回到v9，v9的第一个芯片实现应该在今年晚些时候问世。我相信Media Tech将是首批出来的。因此，我认为这将是一个高性能的核心，上面可能装有SVE-2。实际上，这是与目前Arm产品的最大区别之一。当他们上一次谈论v8时，当他们发布它时，我们已经有一段时间没有看到芯片了。我记忆不是很准确，但64位的第一个芯片实现是苹果A系列，领先其他人的。当苹果发布了自己的64位内核芯片的时候，Arm本身还没有一个64位内核。
This is different now. Again, this is a vision conference, so they didn’t release an Arm v9 core at the conference. They didn’t announce any. But I expect there is one, and it’s in the hands of Media Tech right now. So we will find out more detail probably in the next few months as some other announcements are made.
现在不一样了。同样，这是一次虚拟会议，因此他们没有在会议上发布Arm v9核。他们没有宣布任何消息。但是我希望有一个，并且现在就在Media Tech手中。因此，随着其他一些公告的发布，我们可能会在接下来的几个月中找到更多详细信息。
Arm is moving much faster now than it had in the past. So this announcement… we’ll see silicon later this year. Whether it’s broken up into chiplets or not, that doesn’t necessarily affect the architecture per se, but the other thing that Arm talked about was accelerating innovation in areas like Mali, their graphics core. You could see a chiplet structure with a Mali core and an Arm v9 core. You can do it in chiplets, or you can do it in a full SoC. You can do it either way. That may be something we’ll see more of in the future, but it wasn’t specifically tied to this announcement. That’s one of the things we may see going forward, breaking these things up into chunks where parts are going to be implemented in different process nodes or just by different foundries and then glued together in the package.
One interesting thing about the Intel announcement is that their foundry plan includes Arm cores. You could go to Intel for Arm as well.
BRIAN SANTO: That was almost shocking! Right?
KEVIN KREWELL: Well, yeah. Arm and RISC-V are both IP that Intel will be supporting in their foundry program. As well as X86. It would be crazy: You can mix X86, Arm and RISC-V all in one design.
BRIAN SANTO: And you’re just the guy to do it! Just because.
KEVIN KREWELL: Just turn on the blender and blend it all together.
BRIAN SANTO: You should sell that to the EDA companies. “Stop calling it what you’re calling it; call it a blender.”
KEVIN KREWELL: Yes.
BRIAN SANTO: All right! Now that we’ve got trademarks settled, Kevin, thank you so much for your time today. It’s always a pleasure to talk to you.
KEVIN KREWELL: And you, too, Brian. It’s always fun.
BRIAN SANTO: That was Kevin Krewell, principal analyst at Tirias Research.
BRIAN SANTO：那是Tirias Research的首席分析师Kevin Krewell 。
We invite you to keep reading EE Times and listening to this podcast for ongoing coverage of processor technology, data center innovations, and of course we’ll keep tabs on the intended merger of Nvidia and Arm.