Earlier this month at one of the world’s biggest tech tradeshows—Computex 2024 in Taiwan—Arm announced annual updates to its CPU and GPU IP, with a strong focus on AI. With Arm supplying much of the smartphone ecosystem with its chip architecture, this year’s announcements felt different. While much of the IP that Arm announced at Computex will find its way into chips that launch late this year and into 2025, there was also a clear indication that Arm is excited about what’s happening in the PC ecosystem right now. That’s because Computex was the show where Arm got to take a victory lap (with help from Microsoft and Qualcomm) for the role of its instruction set IP in the just-launched Copilot+ PCs. It’s worth noting that, while Copilot+ PCs run on Qualcomm’s Snapdragon X Series processors that use Arm’s instruction set, they do not use Arm’s CPU or GPU design IP.
For context, there has been an arms race of sorts happening in the world of semiconductors. This arms race has predominantly pitted the various chip vendors that produce smartphone and PC chips against Apple, which has led the industry in growth of unit shipments in the U.S. With its A-Series and M-Series of chips, Apple has proven to be a major challenge for PC and smartphone manufacturers to contend with. And even though Apple is both an Arm licensee and one of the founding companies that created Arm, it is also the company that Arm and most of its customers want to beat.
On top of that arms race, there is also a race to AI relevance and having a story that matches what people believe the market is demanding. As a result, Arm is also shifting away from its TCS platform, which delivers tight integration for Arm’s IP, in favor of CCS, an even more tightly integrated platform that delivers more performance and efficiency.
Arm Cortex CPUs Achieve Major AI Improvements
Arm’s Cortex CPUs are the lifeblood of the company’s IP and have driven much of the company’s success. The Arm instruction set architecture is currently on version 9.2 and is the basis of many of the latest Arm processors, including Apple’s M4. Arm’s own CPU designs run the same v. 9.2 ISA; Arm’s partners can take those core designs and configure them within their own SoCs. This formula has delivered some of the most impressive improvements in smartphone performance and, with the advent of Apple’s M-Series, in the PC as well. It is against this backdrop that Arm has announced its three latest CPU cores: the Cortex-X925, Cortex-A725 and Cortex-A520.
The new Cortex CPU cores follow last year’s cores that introduced the Arm v. 9.2 ISA, which brought significant improvements in security and performance. This new generation is optimized for 3nm process nodes, which are expected to be used for most of next year’s flagship SoCs. Last year’s CPU cores landed in MediaTek’s Dimensity 9300 chip, which I had the pleasure of testing inside of the Vivo X100 Pro; it outperformed most Android devices, including the Qualcomm’s Snapdragon-based devices, in multicore performance thanks to its unique all-big core design. I believe that we might see this trend continue into this year with even faster and more efficient big cores.
As a refresher, Arm has a three-tier CPU hierarchy with X series, A700 class and A500 class CPU cores. All of these CPUs are clustered together by Arm’s new DynamIQ Shared Unit, which is almost the same as last year’s DSU-120. The DSU is how Arm helps its partners manage CPU clusters as well as power consumption and bandwidth. Arm is claiming that the new CPU cluster has up to 36% faster peak performance than last year’s premium Android devices thanks to the new X925 core. Arm has also optimized AI performance and claims an improvement of 172% compared to last year’s TCS23 platform. Arm says that the X925 is the largest uplift in the history of the Cortex-X family, which was on its fourth generation last year. These cores will also likely be at the heart of most Arm-based PC designs that seek to compete with the likes of Apple and Qualcomm next year. (There are already a lot of rumors about MediaTek possibly getting into that game early in 2025.)
The A725 is part of the A700 family of processors, which have slowly become Arm’s efficiency CPU cores after once being the performance cores. Because of their efficiency and high performance, they enable the best sustained performance where most applications will run. Arm says that the A725 is 25% more power-efficient than the A720 and has 35% better performance efficiency. Arm also refreshed its A520 low-power cores with an additional 15% power savings over last year’s A520 from the TCS23 platform.
Arm believes that its CPUs are ample for AI workloads and that a lot of the focus on NPUs is somewhat misguided. While I do agree that the data today does support Arm’s arguments for running AI on the CPU, I also believe that most of Arm’s partners will attach an NPU to its CPU and GPU cores because that’s what’s necessary to be a Copilot+ PC, as I’ve detailed here. Arm says that its AI performance on the X925 is 46% faster than the X4, which it seems to believe will be competitive in the world of NPUs and GPUs.
Arm CSS For Client
Arm CSS for Client is the company’s new platform play that further deepens the integration between hardware and software and optimizes for a specific process node with Arm physical implementations. This includes the integration of Arm’s Kleidi libraries for AI and computer vision as well as a reference software stack for Android and more. Ultimately, with CSS, Arm’s partners will get more but will also probably have to pay more; I say “probably” because Arm doesn’t generally discuss licensing terms or pricing with press or analysts.
Arm wants to get into the business of delivering platform-level performance with its combined cores, which is why it is claiming significant improvements for CCS24 versus TCS23. These improvements include 30% better graphics performance, 33% better app launch speed, 42% to 46% faster AI LLM performance and 60% better web browsing performance. CSS for client also includes Arm physical implementations to help improve power performance area on the 3nm node. These physical implementations are tape-out-ready for 3nm and embody Arm’s desire to help its customers get to market faster. Arm has also teased that while it prefers that its customers use its Immortalis GPUs, there is also support for third-party GPUs, including for next-generation AI PCs.
Arm Immortalis And Mali GPUs
As with its CPUs, at Computex 2024 Arm introduced its latest line of Immortalis GPUs—improving on last year’s models—and once again it couldn’t help but talk about AI. With the new Immortalis G925 (note the continued number matching with the X925 CPU), the company is claiming AI performance uplift ranging from 29% to 50% over the G720 from the last generation. Arm is also claiming significant double-digit improvements in gaming performance, with Call of Duty Mobile gaining as much as 72% more performance over the last generation.
One thing to note is that these comparisons are not necessarily apples-to-apples because Arm’s G925 is a 14-core GPU design compared to the 720’s 12-core design. Arm also says that it will enable up to 24-core designs with up to 50% more shader cores. Arm is also claiming improvements to ray tracing with the Immortalis family, which are welcome; these improvements will come only in GPU configurations with at least 10 cores. Meanwhile GPU configurations with six to nine cores fall into the Mali G725 family, while even smaller implementations of five cores or fewer are in the Mali G625 line.
Will Arm Sustain This Momentum?
Arm is having a moment in the spotlight, especially in servers and the PC—both segments that it simply didn’t have much of a presence in before. For Copilot+ PCs, Microsoft and Qualcomm are optimizing Arm on Windows before x86, and it’s quite clear that Arm wants to continue to be part of that discussion. I believe that we will see more of what Arm has in store early next year from partners like MediaTek, if the rumors are true. My experience with Arm’s last TCS23 platform inside of the MediaTek Dimensity 9300 has shown that Arm and its partners will absolutely put up a fight against Apple and Qualcomm.
I personally can’t wait to see how much diversity Arm will bring to the PC ecosystem, especially because Arm seems confident that it will take significant share away from x86 in the PC market. That all said, Intel and AMD are not sitting still, and I will talk about their own Computex announcements in my next few articles. Intel’s Lunar Lake and AMD’s Ryzen AI processors look to be some of the best low-power platforms the companies have ever created, which could put them in direct competition with Arm.
Read the full article here
