Amlogic A123X Is the First Cortex-A320 Edge AI SoC

Amlogic Ships the First Silicon Built on Arm's Cortex-A320

Amlogic has announced two edge AI system-on-chips built around Arm's Cortex-A320, the low-power Armv9.2-A core Arm introduced for always-on IoT work. The quad-core A123X and dual-core C305X2 are, as far as public announcements go, the first shipping parts to use it. Both are fabricated on TSMC 6nm, both are sampling now, and both are scheduled for mass production in the fourth quarter of 2026.

• A123X: quad-core Cortex-A320 with a 4 TOPS ADLA2 NPU plus an 8 TOPS ADLA3 NPU, LPDDR4/4X/5 to 6400 Mbps
• C305X2: dual-core Cortex-A320 with ADLA2 plus the same 8 TOPS ADLA3 NPU, LPDDR to 4266 Mbps
• Both on TSMC 6nm, both encoding H.264/H.265 at 4K60 and decoding at 4K30
• Sampling now; mass production scheduled for Q4 2026; pricing not disclosed

What Cortex-A320 Brings to a Battery-Powered Camera

The Cortex-A320 exists to replace the Cortex-A35 in devices that need to stay awake for months. Amlogic cites a 34% SPECint2006 improvement over Cortex-A35, while Arm's own product page claims up to a tenfold improvement in machine learning performance against the same predecessor, credited to the SVE2 vector extensions in Armv9. Both are vendor figures and no independent benchmarks exist yet, so read them as claims rather than measurements.

The architectural point still stands on its own. SVE2 on the CPU means the pre- and post-processing around a neural network — resizing frames, normalising tensors, running non-maximum suppression — no longer bottlenecks the NPU. On a small Armv8 core that glue work often costs more wall-clock time than the inference itself.

Why Two NPUs Instead of One Bigger One?

This is the design decision worth dwelling on. The A123X pairs a 4 TOPS ADLA2 block aimed at conventional CNN object detection with an 8 TOPS ADLA3 block that has hardware acceleration for Transformer operations — vision transformers and compact language models. The C305X2 uses its ADLA2 for AI-ISP image processing up to 4K30 and carries the same 8 TOPS ADLA3 alongside it.

Splitting the silicon this way lets one engine run continuously on the cheap, boring work while the other stays powered down until a Transformer workload actually arrives. For a solar doorbell that has to survive on a small panel, duty-cycling a big accelerator is the whole ballgame. It is a different philosophy from the single large NPU we saw on the Amlogic A311Y3, which put 8 TOPS behind an octa-core application processor.

Where Each Chip Is Meant to Land

Amlogic is explicit about targets. The A123X goes after machine vision in devices with a power budget: dashcams, robot vacuums, access control panels, and video conferencing hardware. The C305X2 is tuned for the battery tier — solar doorbells, portable trail and hunting cameras, and compact 2K or 4K battery IP cameras — and adds always-on video and pre-roll low-power modes so a camera can capture the seconds before a trigger without running the full pipeline continuously.

That pre-roll capability is quietly the most consumer-visible feature here. Anyone who has reviewed doorbell footage and found the event began just before the clip did knows exactly what problem it solves.

What This Means for Board Makers

Cortex-A320 parts landing in volume changes what a cheap Linux-capable board can be expected to do. The broader shift toward Armv9 in low-power silicon is the same current that produced Arm's C2-Ultra mobile platform at the high end, and our mini computer coverage has been tracking it from both directions all year. Expect the first A123X development boards and system-on-modules to appear shortly after mass production begins, most likely from the same module houses that built around Amlogic's previous generation.

No pricing has been disclosed for either chip, which is normal at the sampling stage. The figure to watch is what a finished A123X module costs next to an 8 TOPS A311Y3 module — if the newer, lower-power part lands at a similar price, it will be an easy recommendation for anything running on a budget of watts rather than tens of watts.

Sources: CNX Software — September 11, 2026; Arm Cortex-A320 product page — 2026.