A Cortex-A320 That Listens While It Sleeps

Amlogic has detailed two TSMC 6 nanometre system-on-chips built on Armv9.2-A Cortex-A320 cores with SVE2 vector extensions: the quad-core A123X and the dual-core C305X2. Both offer a choice of a 4 TOPS ADLA2 NPU for convolutional object detection and tracking or an 8 TOPS ADLA3 NPU with hardware-accelerated Transformer operations supporting vision transformers and language models, and both include a neural-network hardware sound event detection engine for low-power audio identification. Video is H.264 and H.265 encoding at 4K 60 fps and decoding at 4K 30. The A123X carries a low-light HDR image signal processor handling 16 megapixels at 30 fps SDR and 13 at 30 fps HDR with AI-ISP to 4K 30, 32-bit LPDDR4, 4X or 5 to 6400 Mbps, gigabit Ethernet, USB 3.0 OTG, dual four-lane MIPI CSI, eight-lane Sub-LVDS, MIPI DSI to 1080p60 and two CAN FD. The C305X2 runs memory to 4266 Mbps, drops Sub-LVDS and adds always-on video and pre-roll low-power modes. Amlogic cites a 34 percent SPECint2006 gain for Cortex-A320 over Cortex-A35 and a tenfold machine-learning uplift from SVE2. Both are sampling now with mass production scheduled for the fourth quarter of 2026.

The sound event detection engine is the part worth pausing on, because it is a hardware block for a job usually done in software. Recognising a specific acoustic signature — glass breaking, a smoke alarm, a bearing squeal, a compressed air leak — at a few milliwatts means the main processor can stay asleep until something is worth waking for. In a battery or solar-powered device that is the difference between weeks and years of operation, and the same logic transfers directly to industrial condition monitoring: a sensor that listens continuously and wakes only on an anomaly is a fundamentally better power profile than one that samples on a timer and hopes to catch the event.

Pre-roll is the other quietly significant feature and it solves a familiar frustration. A camera woken by a trigger has already missed the beginning of what triggered it. Pre-roll keeps a rolling buffer encoded before the wake event so the recording starts before the trigger rather than after — which for a security camera means seeing the approach, and for an industrial fault camera means seeing what the machine was doing before it stopped.

For industrial designers the honest positioning is that these are consumer-derived parts with industrial applicability rather than industrial parts. The named targets — dashcams, sweeping robots, solar doorbells, hunting cameras — set the volumes and therefore the price, and the CAN FD interfaces and the temperature question are what determine whether they can be used elsewhere. Cortex-A320 itself is the interesting development: an Armv9 core designed for exactly this point in the range, where a device needs more than a microcontroller and far less than an applications processor, and where SVE2 makes small-model inference practical without a separate accelerator. Whether these particular parts fit an industrial design will come down to the temperature grade and longevity commitment, neither of which the announcement states.