Chess launches low-SWaP CHARM50 edge processor

Chess Dynamics has expanded its Vision4ce video-processing portfolio with CHARM50, an approximately 50mm by 35mm edge-processing module intended to bring AI-enabled detection, classification, and tracking into defence sensors with tight size, weight, and power constraints.

The Horsham-based Cohort Group company is targeting airborne surveillance systems, uncrewed platforms, compact electro-optical gimbals, and hand-held sensors. Processing takes place close to the camera, reducing the need to transmit every raw video stream to a larger central computer before useful information can be extracted.

CHARM50 can process two HD-SDI video inputs simultaneously, allowing sensor packages combining, for example, daylight and thermal cameras to run through the same module. With a single input, the processor can handle higher-bandwidth 3G-SDI video.

The board measures roughly half the size of the company’s CHARM100NX module. Chess has not published a complete power-consumption figure in the launch announcement, so the low-SWaP description cannot sensibly be reduced to a single electrical specification; the disclosed advantage at launch is primarily the smaller physical processing footprint.

Video analysis moves towards the sensor

Distributed surveillance systems can create large communications loads before an operator sees anything useful. High-definition electro-optical and thermal cameras generate continuous streams of data, and moving several of those streams from uncrewed vehicles or remote sensors to a central processing node can consume limited radio bandwidth regardless of whether the images contain targets of interest.

Edge processing changes where the first stage of that workload takes place. Detection, classification, and tracking can run beside the sensor, allowing a platform to transmit tracks, metadata, alerts, selected imagery, or processed video rather than treating the communications network as a pipe for every unfiltered frame.

That can also shorten the time between sensing and response. Sending imagery to a remote processor introduces transmission and processing delay before the result returns to the platform or operator, while local analysis can keep parts of the processing loop within the sensor system.

The trade-off is a harsher environment for the electronics. A processor embedded inside a compact gimbal, vehicle sight, airborne sensor, or uncrewed platform has to work within restricted cooling, electrical supply, mechanical volume, vibration limits, and environmental conditions that differ substantially from those experienced by a rack-mounted computer.

Reducing board dimensions is therefore useful only if the processing architecture can retain sufficient capability within those constraints. Chess is positioning CHARM50 as the smallest end of an existing family, extending Vision4ce algorithms into installations where the company’s larger CHARM processors may be impractical.

The portfolio approach gives system integrators several processing scales rather than forcing one module across every application. CHARM100NX occupies a larger embedded format, while other CHARM products address progressively heavier computing requirements, allowing processing hardware to be matched more closely to sensor count, available volume, power, and algorithm workload.

CHARM50 also uses a modular interface intended to support integration with different camera systems. That matters over the life of a defence sensor because electro-optical technology changes more quickly than many of the platforms carrying it, with new thermal imagers, visible cameras, resolutions, frame rates, and digital interfaces introduced during upgrades.

Keeping image processing separate from the individual camera can make those changes easier, provided interfaces and software are sufficiently controlled. It also gives integrators the option to update analytics without replacing every sensor component, although revised algorithms still require verification before operational deployment.

Software configuration becomes increasingly important as AI-enabled functions move into the sensor itself. Detection and classification results depend on algorithm versions, thresholds, training data, sensor calibration, image conditions, and the environment in which a system is operating.

A fielded fleet therefore needs to know which software is running on which processor, how updates are authorised, and whether revised models change detection performance in ways that affect the operator. Compact hardware does not remove that lifecycle burden; it moves more of the decision-making capability to the edge of the network.

Chess has not published benchmark figures covering CHARM50 inference throughput, detection accuracy, full electrical consumption, environmental qualification, or a named customer programme. Those omissions make direct performance comparisons with other embedded AI processors premature.

The disclosed specifications still establish a clear engineering direction. A 50mm by 35mm processor that can handle two HD-SDI feeds brings Vision4ce’s existing video-analysis functions into a much smaller installation envelope, allowing sensor manufacturers to evaluate edge processing without accommodating a larger standalone computer.

The communications argument grows stronger as military platforms carry more sensors. An individual high-definition feed may be manageable, but several cameras across multiple uncrewed platforms can quickly compete for radio capacity, especially when networks must also carry command traffic, telemetry, targeting information, and other mission data.

Local processing cannot eliminate the requirement for full-motion video. Operators may still need raw or lightly processed imagery for identification, intelligence exploitation, post-mission review, or circumstances in which automated classifications are uncertain.

It can, however, change when that bandwidth is consumed. Systems can transmit higher-value information continuously and reserve full video for occasions when an operator or downstream process actually requires it, making communications capacity less directly dependent on the number and resolution of deployed cameras.

For Chess Dynamics, the commercial measure will be integration rather than board size. CHARM50 needs to move from a product launch into qualified sensor and platform programmes where its smaller footprint produces a measurable reduction in communications load, latency, or installation complexity.

The defence sensor market is steadily adding cameras and algorithms faster than deployed bandwidth is increasing. CHARM50 addresses that imbalance from the processing end, putting more computation beside the sensor so that the network carries information rather than every pixel the camera produces.