Arm has just announced new Armv9 CPUs and Immortalis GPUs for mobile SoCs, as well as the Kleidi AI software optimized for Arm CPUs from Armv7 to Armv9 architectures.
New Armv9.2 CPU cores include the Cortex-X925 “Blackhawk” core with significant CPU and AI performance improvements, the Cortex-A725 with improved performance efficiency, and a refreshed version of the Cortex-A520 providing 15 percent efficiency improvements. Three new GPUs have also been introduced namely the up-to-14-core Immortalis-G925 flagship GPU which delivers up to 37% 3D graphics performance improvements over last year’s 12-core Immortalis-G720, the Mali-G725 with 6 to 9 cores for premium mobile handsets, and the Mali-G625 GPU with one to five cores for smartwatches and entry-level mobile devices.
Arm Cortex-X925
The Arm Cortex-X925 delivers 36 percent single-threaded peak performance improvements in Geekbench 6.2 against a Cortex-X4-based Premium Android smartphone, and about 41 percent better AI performance using the time-to-first token of tiny-LLama (Q4). The Cortex-X925 core was implemented on an FPGA platform with the following configuration: Cortex-X925 @ 3.8 GHz with 2MB L2 cache, 16MB L3, 32MB SLC, DSU @ 2 GHz, and LPDDR5x-8533 memory.
The AI performance was measured to be improved by 46 percent using the time-to-first token for Phi3, and Arm also says X925 SoCs can deliver 33 percent faster application launch times on average across five of the top 10 applications (in Android), and 60 percent faster web browsing measured using the Speedometer 2.1 browser benchmark. The slides shared by Arm mention support for Android, Linux, and Windows operating systems, so it will not only be used in smartphones but also mobile and AI PCs.
The Arm Cortex-X925 core is optimized for 3nm manufacturing processes. You’ll find more technical details about the new core on the developer’s website.
Arm Cortex-A725 and improvements for the Cortex-A520 core.
The Cortex-A725 further improves the performance and efficiency compared to the Cortex-A720 and Cortex-A78 cores. The new core delivers a 35 percent performance efficiency boost over the Cortex-A720, a 25 percent better power efficiency, and 20% L3 traffic improvements. Performance efficiency is defined as the ratio between the improvement in Performance and the improvement in Power for the said performance. The Cortex-A725 peak performance was apparently measured on a 3nm test chip with 64KB K1 and 8MB L3 caches, and compared to a 4nm Cortex-A720 chip. Besides the different process nodes, Arm claims most of the improvements to performance efficiency are due to the microarchitecture of the Cortex-A725.
The Cortex-A520 has been refreshed with updated implementation and a 3nm process delivering up to 15 percent efficiency improvements compared to Cortex-A520 in TCS23.
Immortalis-G925, Mali-G725, and Mali-G625 GPUs
Like the Arm Cortex-X925 CPU, the Immortalis-G925 offers significant performance improvements over the previous generation Immortalis-G720 with 37% better performance in graphics apps, 34% faster AI inference (testing in fp16 mode), and 52% faster ray tracing. Arm further states that the Immortalis-G925 GPU delivers 46 percent performance improvements in mobile, on average, compared to the Immortalis-G720. Some examples include Genshin Impact with a 49 percent boost and Roblox which is 46 percent faster, and the company also tested Call of Duty Mobile, Diablo Immortal, the Day After Tomorrow, Fortnite, and PUBG Mobile with improvements ranging from 29 to 72 percent. We’re also told efficiency has improved by 30 percent on average in leading games.
Immoratlis-G925 AI performance improvements with image classification, image segmentation, object detection, natural language processing, and more…
Arm did not expand on the Mali-G725 and Mali-G625 GPUs. Those look to be smaller variants of the Immortalis-G925 with fewer cores and now ray tracing capabilities optimized for mid-range and entry-level devices.
Kleidi AI software
Arm Kleidi is a suite of software libraries and developer communities designed to accelerate AI development. The Arm Kleidi libraries support popular AI frameworks and are optimized for Arm CPUs from the Armv7 architecture using the Advanced Single Instruction Multiple Data (SIMD) Extension for machine learning (ML) workloads up to the new Armv9 architecture with more advanced features enabling generative AI workloads on the Arm CPU.
Kleidi is comprised of two main projects for now: KleidiAI for neural networks and inference engines and KleidiCV for OpenCV computer vision library.
KleidiAI is a collection of highly optimized AI kernels that work through MediaPipe (via XNNPACK), LLAMA.cpp, PyTorch (via ExecuTorch), and TensorFlow Lite (via XNNPACK). Arm says KleidiAI can accelerate the time-to-first token for Meta’s Llama 3 and Microsoft’s Phi-3 LLMs using llama.cpp by 190 percent on the new Arm Cortex-X925 CPU compared with the reference implementation based on llama.cpp. KlaidiAI is also being integrated into Unity Sentis on-device AI inference engine for game developers
KleidiCV is developed in partnership with OpenCV to optimize over 2500 computer vision algorithms in the popular open-source library. NEON/SVE2-optimized implementations deliver a 75 percent performance uplift on average. Android builds are also being submitted to the Maven Central repository of open-source software components and libraries for Java development
You’ll find a few more details about Kleidi in the announcement.
All those new IP blocks and software make the Arm Compute Subsystems for Client (Arm CSS for Client) that will be integrated into SoCs for AI PCs, smartphones, consumer devices, and more. We’ll likely have to wait until 2025 at least before the first devices with Arm Cortex-X925 or Cortex-A725 come to market.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress
