Particle.news

Download on the App Store

Arm Launches Lumex, a CPU-First Platform to Run AI On Device

Arm pitches a CPU-first path for on-device AI via SME2 with KleidiAI support to curb fragmentation, speed adoption.

Overview

  • The Lumex compute subsystem bundles new C1 CPUs in four tiers (Ultra, Premium, Pro, Nano) with Scalable Matrix Extension 2 plus a Mali G1-Ultra GPU.
  • Arm claims about 25% higher CPU performance versus Cortex X925 and up to a 5x uplift in on-device AI, while G1-Ultra targets roughly 20% faster graphics and double the ray-tracing performance.
  • System-level IP adds a channelized interconnect and an updated System MMU that Arm says can reduce memory latency by up to 75%.
  • KleidiAI libraries integrate with PyTorch, Llama, LiteRT and ONNX to enable SME2 acceleration on the CPU rather than relying on a dedicated NPU.
  • Optimized for 3nm manufacturing, Lumex targets wearables, smartphones and PCs, with partners expected to ship Lumex-based chips later this year or early next year.