ARM’s next-gen CPU and GPU cores are more efficient

ARM is launching a new line of CPU and GPU technologies… as well as a new option for customers to purchase a complete solution that implements all of those features in one package.

The new ARM Compute Subsystems for Client (or CSS for Client) bundles the latest Cortex-X and A-series CPU cores with Immortalis graphics and ARM’s latest Corelink System Interonnect and System Memory Management technologies.

Compared with previous-gen ARM Cortex-X4 CPU cores, the new ARM Cortex-X925 CPU core brings up to a 36% boost in single-threaded performance boost, as measured in Geekbench, up to 50% more TOPS for AI performance, and 41% faster time-to-first-token when running the tiny-Lllama (Q4) AI model, compared with

ARM says these new CPU cores are based on ARM v9.2 architecture and ready for manufacturing on a 3nm process node, and support single-core CPU speeds up to 3.8 GHz.

And ARM says when compared to the previous-gen Immortalis-G720 GPU, the new Immportalis-G925 GPU is up to 37% faster overall, while boosting ray-tracing performance by 52% and offering 34% faster inference for AI tasks.

While the Cortex-X remains the highest-performance CPU architecture in ARM’s lineup, the company is also updating its Cortex-A CPUs, saying the new Cortex-A725 brings a 35% improvement in performance efficiency over the previous-gen Cortex-A720. And while the Cortex-A520 isn’t getting a name change, ARM says this year’s iteration gets a 15% power efficiency improvement compared with the previous-gen version that was part of the company’s TCS23 (total compute solution) version.

One interesting thing is that while ARM, like most companies these days, is highlighting the AI capabilities of its processors, the chip designer still isn’t announcing a first-party NPU. Instead the company says its CSS for Client solutions are designed to compliment custom AI accelerators from the chip makers who license ARM’s designs. But ARM is introducing new software tools called Arm Kleidi which “is a set of compute kernels for developers of AI frameworks” looking to leverage ARM architecture. It includes “KliediAI for AI workloads and KlediCV for computer vision applications.”