Why Ethos-U55: Bringing cloud-proven AI tasks to the MCU class — predictable, efficient, and scalable.

One toolchain. One chip. Deterministic results instead of guesswork.

Arm® Cortex™-M55 + Arm® Ethos™-U55:
Engineered for the Edge

  • Arm® Cortex™-M55 with Helium (M-Profile Vector Extension) Vector DSP/ML performance for audio and sensor preprocessing — spectrograms, filters, downsampling. Eliminates the need for a separate DSP → lower BOM and power.
  • Arm® Ethos™-U55 microNPU Accelerates convolutions, depthwise layers, fully-connected and matrix multiplications — the core operators of ASR, TTS, and compact language models. In tandem with M55 it delivers performance uplift into the hundreds of GOPS, with Arm citing up to 480× ML speed-up compared to previous Arm® Cortex™-M systems without an NPU.
  • One SoC, one memory space No shuttling of tensors between external chips. Predictable latency, minimal jitter — especially vital for multimodal pipelines where audio and vision must align.

From Cloud Workloads to MCU Execution

  • Classes of workloads we bring down:
    • Speech recognition — enabling voice as a natural input channel.
    • Speech synthesis — devices that respond in human voice.
    • Language models — compact dialog and logic engines.
    • Acoustic event detection — identifying critical sounds for safety.
    • Multimodal fusion — combining audio, vision, and sensor streams.
  • How it works:
    • Models compiled with Arm Vela into Arm® Ethos™-U55-optimized binaries.
    • Optimizations include INT8 quantization, pruning, operator fusion, static memory allocation.
  • What OEMs gain:
    • Predictable latency and power profiles.
    • Fit into tight MCU memory budgets (RAM and Flash).
    • Offline-by-default operation with reproducible results.

Multimodal Advantage: Seeing and Hearing Together

  • Division of labor: M55 handles DSP filters and channel sync; U55 crunches CNNs and matrix math.
  • Unified memory: no wasted cycles copying features across interfaces — fusion happens with minimal latency.
  • Stable response: deterministic timing across audio and vision streams, critical for access and safety.
  • Event-driven power: M55 runs ultra-low-power always-on filters, U55 wakes only on trigger — multimodal on battery for weeks.
  • Scalable complexity: from rule-based fusion to full neural fusion, all within the same architecture.

Proven in Silicon

    • Himax HX6538 (WE2) Arm® Cortex™-M55 + Arm® Ethos™-U55 built for battery-driven edge scenarios. Multi-level power management makes it ideal for “always-on” PoCs running for weeks without recharge.
    • Alif Ensemble (E3/E5 series) Arm® Cortex™-M55 + Arm® Ethos™-U55 with multi-core MCU options, built-in security subsystems, and extended connectivity. Suited where certification, robustness, and long lifecycle are mandatory. Features aiPM (Autonomous Intelligent Power Management) — multiple power domains, dynamic clock gating, domain shutoff and sleep modes, so only the necessary parts of the chip remain powered.

Together these two SoC families prove that U55 scales both to ultra-efficient consumer devices and to heavy-duty OEM systems.

Why It Matters for OEMs

  • Determinism: Vela reports memory and performance budgets → no surprises at integration.
  • Lower TCO: one SoC vs. “CPU + external DSP/NPU” → fewer components, simpler integration, reduced jitter.
  • Privacy & SLA: offline-by-default, no dependency on connectivity.
  • Future-proof toolchain: the very same flow extends seamlessly to Arm® Ethos™-U85.

Next Step: Arm® Ethos™-U85 — Same Path, More Headroom

  • More compute: scaling from ~256 GOPS up to 4 TOPS @1 GHz, with improved efficiency per watt.
  • New model classes: native transformer support — longer ASR context, more fluent TTS, stronger small-LM dialog, robust multimodal fusion.
  • Continuity, not disruption: still compiled with Vela, still portable across frameworks (ExecuTorch backend for U55/U85). Engineers extend tasks, not re-architect systems.
  • FOEM reality: everything you validate today on U55 runs on U85 with breathing room — more parameters, longer sequences, richer fusion. One step closer to cloud-class experiences, still fully offline.
Request PoC