MP10304 Quad-AMP PCIe Card
Overview
The Mythic MP10304 Quad-AMP PCIe card enables high performance, power efficient AI inference for edge devices and servers. The half-height, half-length (HHHL) PCIe card simplifies the integration effort into platforms where space is a constraint. The MP10304 features four M1076 Mythic Analog Matrix Processors (Mythic AMP™), delivering up to 100 TOPs of AI performance and supporting up to 320 million weights for complex AI workloads at less than 25W of power. Large DNN models can be deployed on the MP10304 PCIe card using the combined AI compute of the four M1076 AMPs. The MP10304 can also run a variety of smaller DNN models for video analytics applications processing images from multiple cameras.
Features
Four M1076 Mythic AMPs
AI compute performance of up to 100 TOPs
No external DRAM required
Support for industry standard AI frameworks
Pre-qualified networks including object detection, classification, pose estimation, depth estimation, and image segmentation networks
Support for up to 320 million weights on-chip
On-chip storage of model parameters
4-lane PCIe 3.0 for up to 3.9GB/s bandwidth
OS Support: Ubuntu, NVIDIA L4T, and Windows (future release)
Workflow
DNN models developed in standard frameworks such as Pytorch, Caffe, and TensorFlow are implemented and deployed on the Mythic Analog Matrix Processor (Mythic AMPTM) using Mythic’s AI software workflow. Models are optimized, quantized from FP32 to INT8, and then retrained for the Mythic Analog Compute Engine (Mythic ACETM) prior to being processed through Mythic’s powerful graph compiler. Resultant binaries and model weights are then programmed into the Mythic AMP for inference. Pre-qualified models are also available for developers to quickly evaluate the Mythic AMP solution.
Models
Mythic provides powerful pre-qualified models for the most popular AI use cases. Models have been optimized to take advantage of the high-performance and low-power features of Mythic Analog Matrix Processors (Mythic AMPTM). Developers can focus on model performance and end-application integration instead of the time-consuming model development and training process. Available pre-qualified models in development:
