ARM is introducing a new range of CPU and GPU technologies and a new way for consumers to buy a comprehensive solution that combines all of those characteristics into one package.
The newest Cortex-X and A-series CPU cores are combined with Immortalis graphics and ARM’s most recent Corelink System Interconnect and System Memory Management technologies in the new ARM Compute Subsystems for Client, also known as CSS for Client.
The new ARM Cortex-X925 CPU core, when compared to the ARM Cortex-X4 CPU cores of the previous generation, boosts single-threaded performance by up to 36%, increases TOPS for AI performance by up to 50%, and results in a 41% faster time-to-first-token when running the tiny-Lllama (Q4) AI model.
According to ARM, these new CPU cores allow single-core CPU speeds of up to 3.8 GHz and are based on the ARM v9.2 architecture. They can be manufactured on a 3nm production node.
Additionally, ARM claims that the new Importalis-G925 GPU is up to 37% quicker overall than the Immortalis-G720 GPU from the previous generation. It also offers 34% faster inference for AI applications and a 52% jump in ray-tracing performance.
ARM is upgrading its Cortex-A CPUs, claiming that the new Cortex-A725 offers a 35% increase in performance efficiency over the Cortex-A720, even though the Cortex-X is still the highest-performing CPU architecture in the company’s range. The Cortex-A520, according to ARM, has a 15% increase in power efficiency over the previous generation, which was a component of the company’s TCS23 (total computing solution) edition, even if the name hasn’t changed.
It’s noteworthy that, in contrast to most firms these days, ARM is not unveiling a first-party NPU despite showcasing its chips’ AI capabilities. Instead, the business claims that its CSS for Client solutions is made to work in tandem with the proprietary AI accelerators that chip manufacturers that license ARM’s designs create.
However, to take advantage of ARM architecture, ARM is releasing new software tools called Arm Kleidi, which “is a set of compute kernels for developers of AI frameworks.” “KliediAI for AI workloads and KlediCV for computer vision applications” are among its features.