Boardcon CM311Y3 Packs 8 TOPS Into a 52x45mm Module

An 8 TOPS Edge AI Module on an Eight-Layer Board

Boardcon has built a system-on-module around Amlogic's A311Y3, and the CM311Y3 is a compact one: 52 x 45 mm, eight-layer PCB, 210 castellated pads along the edges. CNX Software covered it on September 3, 2026. We looked at the A311Y3 silicon itself back in June; this is the first module that turns it into something you can design onto a carrier board.

• CPU: two Arm Cortex-A78 cores at up to 2.4GHz plus six Cortex-A55 cores at up to 2.0GHz, with two RISC-V cores handling system control, power management and sensor hub duty
• NPU: 8 TOPS with INT4, INT8, INT16, FP8, FP16 and BF16 support
• Memory and storage: 4GB, 8GB or 16GB LPDDR5/LPDDR5X, paired with 32GB to 256GB of eMMC flash
• Software: Android 16 with Linux 6.12.69 and U-Boot 2025.01

Why the Castellated Format Matters for Edge AI

A castellated module is the pragmatic middle ground between an off-the-shelf single board computer and a full custom design. You get the hard part — the SoC, memory, power sequencing and high-speed routing — as a validated unit, and you design the cheap, low-speed carrier yourself. For a product that needs specific connectors, a specific enclosure and a ten-year supply commitment, that is usually the right trade.

Boardcon has been consistent about this format. Its Tiny1126B squeezed AI vision onto a 34 x 30mm module earlier this year, and the CM311Y3 is the same idea scaled up to a considerably heavier compute budget.

The dual RISC-V cores are a small detail with real consequences. Offloading power management and sensor polling to always-on RISC-V cores means the Cortex-A78 cluster can stay asleep in idle states without losing responsiveness to a wake event — which is exactly what a battery-adjacent or fanless design needs.

What Can You Build With 8 TOPS?

Eight TOPS is not a data-centre number, and it is not trying to be. It is a useful number for the workloads this module targets: 4K video surveillance with on-device analytics, interactive kiosks and terminals, enterprise thin clients, and autonomous mobile robots. Object detection, segmentation, keyword spotting and pose estimation all run comfortably in that envelope.

The INT4 and FP8 support in the NPU is the part worth planning around. Quantised inference at 4 and 8 bits is where most practical edge deployments end up, and hardware that natively handles both saves you from software emulation penalties that quietly eat the headline TOPS figure.

The media block backs up the surveillance use case: decode for H.265, AV1, VP9 and AVS3 up to 4Kp120, H.264 decode to 4Kp60, encode for H.264 and H.265 to 4Kp60, and a 16MP ISP. Dual MIPI CSI inputs let you run two camera streams into the same module.

Connectivity and Carrier Options

I/O runs to USB 3.0, PCIe 3.0 x2, dual MIPI CSI, and a display path that can be configured as HDMI, MIPI DSI, LVDS or eDP. Networking arrives through a Realtek RTL8211H Gigabit Ethernet PHY on the module. Graphics are handled by an Arm Mali-G625 MC1 with OpenGL ES 3.2, Vulkan 1.4 and OpenCL 3.0 — the OpenCL support being the one that matters if you want to push work at the GPU rather than the NPU.

Boardcon has not published pricing, which is normal for this class of module: volume quotes come through the sales channel, and the evaluation kit usually lands first. The A311Y3 is fabricated on TSMC's 6nm process and includes hardware support for post-quantum cryptography, which is an unusual thing to find at this price tier and a reasonable box to tick on a product with a long service life.

Where It Sits in the 2026 SoM Landscape

Against the Rockchip modules that dominate our mini computer coverage, the CM311Y3 competes on NPU throughput and media decode breadth rather than on ecosystem maturity — Rockchip's mainline Linux story is further along. Against Nvidia's Jetson line it is not a serious competitor on raw compute, and it is not priced to be.

What it offers is a fairly modern Arm cluster, a genuinely flexible NPU and a full 4K media pipeline in a form factor you can commit to a product design. For a lot of industrial builds, that is the whole shopping list.

Sources: CNX Software — September 3, 2026; LinuxGizmos — September 2, 2026.