IN Brief:
- Logiicdev has joined the Siemens Cre8Ventures Factory-One programme with its proposed Mantrum AI accelerator architecture.
- The initial technical work is intended to use FPGA hardware and representative LiDAR, vision and multisensor datasets.
- No measured performance improvement, dedicated semiconductor production commitment or completed industrial deployment has been reported.
Siemens is examining the potential of deterministic AI processing for industrial robotics through the addition of semiconductor developer Logiicdev to its Factory-One programme. The collaboration, announced by Siemens Cre8Ventures on 19 October, will explore whether specialised hardware can support predictable, low-power processing of information from multiple sensors in autonomous and industrial robotic systems.
Logiicdev is developing Mantrum, an AI accelerator architecture intended for real-time sensor fusion and physical AI applications. Its proposal focuses on reducing processing latency and managing power consumption while producing repeatable execution timing. The companies have not disclosed benchmark measurements demonstrating these advantages in a complete industrial robot.
Factory-One is part of the Siemens Cre8Ventures mission-led programme, which brings technology developers together around industrial engineering problems. Rather than announcing a finished robotics product, the latest collaboration provides a possible route for evaluating Logiicdev’s architecture against a defined application and deciding whether further development is justified.
The proposed starting point is an FPGA implementation of Mantrum using representative LiDAR, camera and multisensor datasets. Field-programmable gate arrays allow digital processing structures to be implemented and modified in configurable hardware, providing a way to explore acceleration, memory arrangements and data movement before committing to an application-specific semiconductor design.
Robotic perception systems often collect information from several devices operating at different rates. Cameras provide image frames, LiDAR sensors produce spatial measurements and motion sensors can report acceleration or rotation at substantially different sampling intervals. Those measurements must be associated with the appropriate time and position before a controller can interpret the environment.
Sensor fusion combines information from these different sources to form estimates of object location, movement and the robot’s relationship with its surroundings. A camera may provide visual features that help identify an object, while LiDAR contributes distance measurements. Combining the information can help resolve uncertainties present in either sensor alone, provided the measurements have been calibrated and aligned correctly.
Timing errors can undermine that process. When a moving robot receives measurements captured at different moments, treating them as simultaneous can produce incorrect estimates. Timestamping, buffering and synchronisation are therefore part of the sensing architecture, alongside the algorithms used to interpret the data.
Artificial intelligence introduces further computational requirements. Neural-network inference may involve large numbers of repeated arithmetic operations and significant movement of information between memory and processing elements. Dedicated acceleration can execute some of those operations in parallel, although overall performance depends on the model, numerical precision, available memory bandwidth and supporting software.
In control-related systems, predictable execution timing can be as important as average processing throughput. An algorithm that completes rapidly under typical conditions may still cause difficulties if occasional delays exceed the interval available before a control decision must be made. Scheduling becomes more demanding when several sensing and processing tasks share computing resources.
Deterministic execution aims to reduce uncertainty about when a computation will complete under specified conditions. Achieving that requires examining processing paths, memory access, communication delays and the behaviour of concurrent workloads. It cannot be demonstrated merely by quoting an accelerator’s peak arithmetic throughput.
Logiicdev proposes that Mantrum can address these requirements through a processing architecture designed for predictable operation and low power consumption. The FPGA evaluation is intended to investigate those claims using representative workloads. Siemens has not identified a particular robotics benchmark, published worst-case execution times or quantified the proposed energy savings.
FPGA implementation allows hardware designers to investigate architectural trade-offs before manufacturing dedicated silicon. Processing pipelines can be rearranged, buffering strategies modified and interface arrangements tested. Measurements of resource utilisation, clock frequency and data throughput can then help establish whether an alternative implementation offers advantages for the selected application.
Results from an FPGA are not automatically equivalent to those of a future system-on-chip. The two technologies have different physical structures, power characteristics and implementation constraints. A dedicated semiconductor design may improve some measures of efficiency, but its performance must be established through its own design, verification and testing.
Power consumption affects both mobile robots and fixed industrial equipment. Computing hardware generates heat according to its electrical activity and efficiency, while cameras, LiDAR, communications and control electronics introduce additional loads. Reducing energy consumption within one accelerator may be useful, but the change must be assessed within the complete sensing and processing subsystem.
Thermal constraints can also influence sustained performance. A processor may operate at one rate for a brief workload but require a lower operating point when installed in an enclosure with limited cooling. Any useful comparison should therefore consider electrical power, temperature and processing performance under representative operating conditions.
The companies identify possible longer-term integration with motion-control systems and industrial communications technologies. Such work could connect perception processing more closely with equipment responsible for physical movement, but the October announcement does not establish that those interfaces have been implemented or validated.
A further distinction concerns functional safety. Predictable processing time does not guarantee that a perception algorithm will identify an object correctly, and accurate perception does not establish that an actuator will respond safely. Industrial systems require risk assessment and verification appropriate to the machinery, operating environment and consequences of a failure.
Siemens’ Factory-One programme may provide simulation and engineering resources for examining those interactions, but programme participation does not confer certification on Logiicdev’s technology. Any safety-related use would require separate evidence covering the selected hardware, software, interfaces and system-level behaviour.
The companies have not announced an industrial customer deployment, semiconductor tape-out, manufacturing agreement or commercial order associated with Mantrum. The next development will be the selection of a representative robotics problem and measurement of the proposed FPGA implementation against defined timing, accuracy, power and integration requirements.
