草榴社区

MulticoreWare

MulticoreWare, a software product and engineering services company, is headquartered in San Jose, CA, USA and has offices across the world. MulticoreWare primarily operates in the areas of media artificial intelligence (AI) and autonomous vehicle technologies, with key areas of expertise in human behavior AI, computer vision, deep learning, video engineering, image/data processing and compilers. MulticoreWare has developed key IP around all areas of human behavior AI including face recognition, pose estimation, action intelligence and other areas of human intelligence. MulticoreWare serves a wide group of customers with its AI/ML/DL algorithms and ensuing analytic solutions optimized for cloud and edge computing implementations in various programming languages and frameworks.

Products

MulticoreWare's foundation lies on homogeneous and heterogeneous computing (CPU/GPU/DSP/NN) platforms. The company's experience in CUDA, OpenCL, intrinsics, and Halide enables microarchitecture-aware optimizations for seamless performance and energy-efficient solutions. MulticoreWare has developed and deployed many integrated circuit microarchitecture tools, including compilers, PPA, MxPA, and has ported libraries for ease of platform solution deployment to meet our customers' time to market requirements.

  • Face AnalytiQ - Highly scalable, feature-rich AI engine that delivers high performance and high accuracy to the edge devices and enterprise market

    Features include face detection & recognition, landmark detection, eyeball tracking, facial attributes, and spoof detection. The library is very lightweight, portable and has a multitude of integration options.

  • Pose AI - An AI-based, lightweight, cross-platform tool to detect and digitize human poses in a video stream

    Features include multi person pose estimation, multi person detection & tracking, real-time live PoseAI, etc. The library is computationally inexpensive, and can precisely localize key points by accurately detecting the person.

  • DMS - Ultra lightweight, real-time AI based driver monitoring system optimized for edge platforms

    Features include driver ID, distraction detection, drowsiness detection, emotion recognition, and driver specific activity. The solution keeps the driver attentive by giving alerts and can be optimized for customer specific platforms.

  • LipSync - A format and content agnostic “watch and listen” AI tool that decides and flags media content with lip-sync errors

    Features include auto detection of audio-visual (AV) sync errors, intelligent hotspot detection, AV watermarking agnostic, and scene and speaker detection. LipSync automates the process of detecting audio and video sync errors.

  • x265 - The leading Open-Source encoder for HEVC

    Features include ABR Ladder, Dolby Vision support, AVX512 optimizations, application (with use of ABR Ladder) can be used to create multiple instances. x265 is available as an encoder library and application and there is no difference between a general purpose license and a Commercial license.

  • UHDkit - A commercial software encoder library for AVC, HEVC & AV1

    Features include running x264, x265, SVT-AV1 simultaneously, FFmpeg integration, segment parallel encoding, and capped VBR. UHDkit makes it easier to build optimized video processing pipelines by effectively utilizing many core servers.

MulticoreWare has standard software products (licensable), builds custom software products (application/hardware specific) and provides software engineering services to accelerate the time to market for customers/partners.

ARC-specific Support Details

MulticoreWare has worked with 草榴社区 on model optimization techniques like pruning, compression and quantization for compressing and downsizing CNN models with very minimal or no loss in accuracy in Caffe and TensorFlow frameworks.

Key achievements include:

  • Designed a converter tool for converting ‘TensorFlow’ CNN models to ‘Caffe’, supporting up to 10+ models
  • Achieved significant compression on few selected CNN models with no loss in accuracy on Imagenet dataset in Caffe Framework
  • Developed an auto pruning tool to automate the manual analysis to compress Caffe based model
  • Provided behavior and OPs support analysis for the Tensorflow CNN model for pruning and compression
  • Support for 草榴社区 EV Quantization logic to perform quantization for classification & detection models in TensorFlow framework

Learn more about how MulticoreWare and 草榴社区 work together.