The MP10304 PCIe card features four M1076 Mythic Analog Matrix Processors (AMPs), delivering up to 100 TOPs of AI performance and supporting up to 320 million weights for complex AI workloads at less than 25 W of power. The combination of high-performance and power-efficiency in a compact form factor makes the MP10304
PCIe card ideal for edge AI applications in video security, commercial drones, and product inspection in manufacturing.
The MP10304 PCIe card can be configured to run large DNN models utilizing the combined AI compute of all four AMPs for high-definition camera applications that need to detect smaller objects with minimal or no downscaling, for example in commercial drones and physical security. The MP10304 can also run a set of smaller DNN models for applications that process and analyze independent video streams concurrently from multiple cameras, such as network video recorders (NVRs).
“We continue to see increased compute requirements for edge AI applications, including the ability to run multiple models simultaneously. The MP10304 is ideal for this type of use case, making it easy to run object detection followed by further classification of the detected object, for example,” said Tim Vehling, senior vice president, product & business development at Mythic. “In addition, the MP10304 can also leverage its 100 TOPS of AI compute for manufacturing applications where a complex model will ingest data from a high-definition camera to accurately detect assembly defects.”
Integrating four M1076 AMPs, the MP10304 PCIe card offers 4X the performance of each M1076, packaged in a compact half-height, half-length PCIe card, offering an unparalleled combination of performance and power-efficiency for embedded edge AI applications. With the MP10304, complex AI networks can be deployed in edge appliance and network video recorders (NVRs) to ingest video data from many cameras in the field, and complex networks can be deployed at high resolution for
applications requiring high accuracy. Mythic currently supports object detection, classification, and low-latency human body pose estimation. Other popular models such as depth estimation and image segmentation will be available soon.