It describes NM Card as a PCIe product with 16 neural cores, five ARM cores, and 5 GB of memory, while noting that its published 512 GFLOPS FP32 metric is not directly comparable with the INT8 TOPS figures commonly used for Jetson Orin Nano or Huawei Ascend 310.
The operational signal is an RZD computer-vision deployment. According to the review, a replacement of Jetson Nano initially encountered roughly fivefold lower speed, unsupported YOLOv4 layers, and the absence of the TensorRT/FP16 path used by the earlier stack. The project moved to YOLOv5 and used four independent compute units to run four model instances in parallel, reportedly reaching 30 fps for YOLOv5n at 320×320.
This is evidence of deployment-oriented adaptation: model choice, graph support, precision, and workload partitioning can determine whether a local accelerator is viable in an existing industrial vision system.