ARM has unveiled its next generation CPU and GPU designs and is doing some rebranding at the same time – meet the ARM C1 CPUs and G1 GPUs, which will form the ARM Lumex compute subsystem (CSS). “Lumex” is the new branding that ARM will use for mobile-focused designs (for others, “Niva” will be used for PC, “Zena” for automotive and so on).
This generation of CPU and GPU designs brings massive improvements to buzzwordy workloads (AI, ray tracing), while also delivering solid improvements to less flashy workloads.
Note that ARM has also created a customizable chipset, which can be licensed and fabbed at different 3nm foundries – we cover this in a separate post.
Let’s look at the CPUs first and we’ll start with a cheat sheet to help you keep track of the new branding – Cortex-X and Cortex-A are no more!
- Cortex-X9xx → ARM C1-Ultra
- (New) ARM C1-Premium
- Cortex-A7xx → ARM C1-Pro
- Cortex-A5xx → ARM C1-Nano
The new C1 cores are the first to be built on the ARMv9.3 architecture. All of them support ARM’s Scalable Matrix Extension 2 (SME2), which accelerates AI workloads, but also improves more routine stuff – e.g. decoding HDR video is 10% more efficient with SME2.
The ARM C1-Ultra is focused on increasing Instructions Per Cycle (IPC). In single-threaded tasks, the C1-Ultra is up to 25% faster than the Cortex-X925.
ARM also introduces a new tier of CPU core, the ARM C1-Premium. This one is aimed at sub-flagship designs – it has 35% smaller surface area than the Ultra (meaning that chips with it will be cheaper), which gives it the best-in-class area efficiency (i.e. performance per mm² of silicon).
The ARM C1-Pro core will be used for sustained workloads in high performance chips and as the big core for mid-range designs. Compared to the Cortex-A725, it can deliver up to 16% higher performance in things like gaming. Alternatively, it can be up to 12% more power efficient for video playback, web browsing and social media.
The ARM C1-Nano changes are focused almost entirely on power efficiency – this tiny core is up to 26% more power efficient than the Cortex-A520. It also brings a modest performance improvement and has a 2% smaller core area.
The CPU cores will be orchestrated by the new C1-DynamIQ Shared Unit (DSU). This is responsible for things like sharing the L3 cache among all cores, power management for the cores and more. The new DSU enables power savings of up to 26% compared to the previous DSU-120.
An ARM C1 CPU cluster can be configured with as little as 1 CPU core and as many as 14 cores. Chipset designers can mix and match, combining up to three core types, choosing between Ultra, Premium, Pro and Nano.
ARM claims that for real-world workloads, a C1 CPU cluster brings an average of 30% higher performance on industry-leading benchmarks and an average of 15% speed-up for things like gaming and video streaming. Additionally, it uses an average of 12% less energy for things like video playback, web browsing and social media compared to previous generation CPUs.
As we mentioned above, the new C1 cores enable massive performance improvements for AI through the SME2 extension. We’re talking up to 4.7x lower latency for Whisper Base (an Automatic Speech Recognition model, i.e. speech to text), 4.7x higher AI performance for Google’s Gemma 3 model and 2.8x faster audio generation for Stability AI (a text to audio model that can generate background audio, music and more).
App developers will get the improved performance “for free” on next-gen hardware as SME2 support is integrated into major AI frameworks from ARM itself, Google, Microsoft, Alibaba and Meta.
According to a recent study, a whopping 83% of gamers play on mobile. It’s a lucrative business, both in terms of the games themselves and hardware. ARM says that it has shipped over 12 billion GPUs to date – and these are the most powerful it has designed yet.
The ARM Mali G1-Ultra introduces a second generation Ray Tracing Unit (RTUv2), which doubles the ray tracing performance compared to the one inside the Immortalis-G925 GPU.
Note that there’s more to rendering a game scene than ray tracing, so you can expect to see 40% higher frame rate in games that use hardware ray tracing. Additionally, the RTUv2 is now a separate module and brings a number of power efficiency optimizations, including a simple one – it can go to sleep when the system is idle.
Ray tracing aside, the Mali G1-Ultra can deliver 20% higher raster performance than the G925 across leading benchmarks. It can also be 9% more power efficient. GPUs can be used for AI too and thanks to a brand new GP16 matrix compute path, the G1-Ultra is 20% faster on AI inference.
Here are some more concrete numbers on popular games:
| Arena Breakout | +25% |
| Honkai Star Rail | +19% |
| Genshin Impact | +17% |
| Fortnite | +11% |
| ARM in-house game demo (Mori) | +26% |
ARM Mali-G1 GPUs can be configured with between 1 and 24 shader cores. The G1 GPUs support ARM’s Accuracy Super Resolution (ASR), which is the company’s temporal upscaling tech (think DLSS). This is already supported by Unreal Engine 5 and is integrated into Fortnite.
Of course, you don’t have a smartphone with C1 CPU and G1 GPU yet. ARM says that the next generation hardware will be “in consumer devices in the very near future”.