Arm’s bid for smarter, AI-powered phones and PCs begins with Lumex

With a little luck, the new CPU LUMEX of Arm platform can give us an index on what to expect for the Windows to come on ARM PCs: four CPU power levels, as well as an engine of rays and an increase in improved graphics.

ARM says that her new fleas in the Lumex C1 series will offer 25% more performance than the Cortex X925 processors series that she launched in May 2024. Like the X925, the latest Lumex C1 nuclei are optimized with 3 NM process technologies, with physical implementations and fund collaborations for customers.

(While Arm hinted at the construction of its own physical nuclei for years, however, the representatives of the company would only say that it includes “physical implementations ready for production for partners” and is “not a chip”))))

The LUMEX platform is the ARM brand for smartphones. The PC -specific Arm chips will be marked as “Niva” under the new ARM denomination scheme, although they share certain common features with the Lumex nuclei. Qualcomm, which actually manufactures Arm processors, uses the brand “Snapdragon”, which will almost certainly continue.

Arm chips feed the vast majority of global smartphones, because ARM applies its conceptions to customers who can choose to market ARM cores or take an architectural license and a compatible design but otherwise new conceptions of their own manufacturing. This is the approach that Apple and Qualcomm have adopted – which, in addition to designing smartphones, brought Arm into the Mac OS space as well as Windows on ARM PCs. Although Qualcomm and Arm have had their legal differences on the license in the license – which has since been settled – an arm representative refused to comment when he was asked if ongoing legal problems would prevent Qualcomm from taking a license.

Arm

And you will see the lumex apparently in more than PCs. “Thus, the LUMEX platform will feed flagship smartphones through PCs and tablets,” said James McNiven, vice-president of product management for Arm, in a press point.

Smartphones, however, have their own requests: low power, for which the RISC architecture of Arm has been designed; and maximize local AI. The Cortex C1 series understands which Arm calls the extension of evolving matrix 2 (SME2), using the ARM V9.3 instructions.

Arm does not use dedicated NPU. Instead, it uses a technology called Kleidii which mainly uses software libraries to respond to specific AI functions inside the CPU, regardless of the version of ARM architecture. In the CPU C1 cluster, says Arm, you will see a 5x uprising in AI performance.

Arm says that this will increase the performance of applications that always have a certain form of IA technology in progress, such as audio generation, camera inference or computer vision. More specifically, ARM claims more than 4.7 times the improvement of latency, voice recognition and classic tasks of the large -language language model, and approximately 2.8 times the faster audio generation.

The other thing that the arm has been known is that it was essentially the pioneer of the concept of performance and efficiency cores, a strategy known as “Big-Little”. But with LUMEX, it was brought to a new level with the addition of a new “premium” nucleus. Now there is four Different levels of CPU nuclei: C1-UTRA, C1-Primium, C1-Pro and C1-Nano.

Arm

Nothing has really changed as much: ultra and premium nuclei are simply two levels of “performance” nuclei, while the pro and nano offer different levels of efficiency. Arm leaders said that the premium nucleus could represent ultra nuclei on non -flagship and cheaper devices because it offers ultra performance, but in an area of ​​35% smaller. The C1 Pro “improves” the Cortex A75 in terms of performance and efficiency, said McNiven, while the Nano has the “smaller imprint” and will play a role in flagship and entry-level devices, he said. Overall, ARM leaders have said that AI on the device is three times more economical in power than previous implementations, and that the pro is 12% more economical in power at the same frequency.

Arm also introduces a new GPU, Mali G1-UTRA, which will promise 20% better graphic performance, twice the tracing performance of the shelves and 20% faster for the treatment of AI than the previous immortalis G925. Mali G1-Ultra will also consume less power, because the block is on its own Power island with less leak when it is inactive.

More specifically, Arm claims that the image frequencies on the games on the shelves will be 40% higher than its predecessor, part of the transition to a “unique ray” model to improve efficiency and more realistic lighting, “said Mcniven.

Arm

“One of the examples we have seen recently was some of the new tracing benchmarks like [UL’s 3DMark] Solar Bay Extreme, and I think we see a doubling of performance there, because it is so heavy with rays. It therefore really depends on the quantity of tracing content of the shelves, “said Mcniven.

An idea behind the LUMEX platform, said managers, was to move certain AI functions based on the cloud on the device. More specifically, a large language model in the world of Krafton Inzoi (A spiritual successor to Sims) was directed on devices at the GDC conference, they said, as well as a “coach” who saw you play in Tencent Honor of kings And offered advice. A large online payment supplier also strives to put agentic AI on the device to manage payments treatment during peak hours, instead of engaging in expensive and costly cloud servers, said Chris Bergey, senior vice-president of the Customer section in Arm.

“If your device is able to run a large language model, you have an additional way to interact with the game that increases your experience,” said Mcniven. “But if you don’t have it, the game is still playable.”