Advantech AOM-6741 Packs 100 TOPS Into a SMARC Module

The interesting number on the Advantech AOM-6741 is not the 100 TOPS. It is the 16. As in sixteen concurrent camera streams, into a module small enough to drop onto a carrier board. Announced through CNX Software on August 25, 2026, the AOM-6741 is a full-size SMARC 2.2 module built on Qualcomm's Dragonwing IQ-9075, and it is aimed squarely at factory vision work that has outgrown a single sensor.

• The Dragonwing IQ-9075 pairs an octa-core Kryo Gen 6 CPU at 3.2 GHz with quad-core Cortex-R52 real-time cores, an Adreno 663 GPU and a Hexagon Tensor Processor rated at up to 100 TOPS INT8
• Configurations reach 36GB of LPDDR5 memory and 128GB of UFS storage
• All I/O exits through a standard 314-pin MXM 3.0 edge connector: four DisplayPort outputs, dual 2.5GbE, two USB 3.2 Gen2 ports and PCIe Gen3 x2 and x4
• The module runs Ubuntu 24.04 with the Advantech Robotic Suite, including AI frameworks, multimedia pipelines and ROS2 tooling

Why a SMARC Module Instead of a Complete Board

Computer-on-module formats exist to separate the part that changes fast from the part that does not. The compute module carries the SoC, memory and storage; the carrier board carries the connectors, power circuitry and whatever industrial interfaces a given deployment needs. When the SoC generation turns over, you replace a module rather than requalifying an entire product.

For industrial vision that argument is unusually strong, because the carrier design tends to encode years of accumulated field knowledge about connectors, isolation and enclosure fit. SMARC 2.2 keeps that investment intact. We saw the same logic in the AAEON uCOM-Q6490 SMARC module last week at 12 TOPS and in the Jetway SMC-ARK1 with an RK3588 at 6 TOPS. The AOM-6741 sits at the top of that stack.

What Do You Do With 100 TOPS and 36GB?

Advantech's framing is a move from traditional machine vision to what it calls real-time vision reasoning — and the memory configuration is what makes that phrase more than marketing. Classical machine vision runs a fixed pipeline: threshold, find edges, match a template, pass or fail. That fits comfortably in a few gigabytes.

Running a vision-language model that can be asked an open question about a scene does not. Thirty-six gigabytes of LPDDR5 is enough to hold a genuinely capable multimodal model resident alongside the video buffers for many streams, which is the actual unlock here. The quad Cortex-R52 real-time cores handle the deterministic side — the motion control and interlocks that cannot wait for a Linux scheduler — while the Kryo cores and Hexagon NPU do the thinking.

Availability and What Is Still Unknown

Advantech says the AOM-6741 is sampling now and has not published pricing, which is standard for this class of part and means real-world cost will come down to volume and carrier design. The module is part of a wider Dragonwing IQ9 push the company began earlier in 2026 across edge AI systems, not a one-off.

For anyone building on the mini PC and single board computer side of edge AI, this is a useful data point on where module-level performance now sits. A hundred TOPS on a socketed module is a specification that would have described a full workstation not many years ago.

Sources: CNX Software — August 25, 2026; Advantech Newsroom — 2026; eeNews Europe — 2026.