Llama2, YOLO Family and MediaPipe family of networks now available

October 31, 2023

Quadric today delivered a Halloween treat for all registered users of the Quadric Developers’ Studio. (Link).   This latest update of the industry’s flagship GPNPU software tools showcases a significant new batch of machine learning benchmark networks and several big enhancements to the underlying ChimeraTM software development toolkit that powers Quadric’s Chimera general purpose neural processor.

Llama2 Code Available

Quadric’s first to market implementation of the Llama2 large language model is now available for customer viewing and inspection in the benchmarks section of the Studio.   Users can explore all aspects of the implementation of the 15M parameter “baby” Llama2 implementation, including all source codes.

MediaPipe Network Family, Yolo Networks

A suite of the Google Mediapipe family of networks is now available: Hand Landmark (full, lite); Palm Detection (Full, Lite); Face Landmark; Face Detection short range; and SSDLite Object Detection all run on the Chimera GPNPU and can be benchmarked and viewed by prospective licensees.

Additionally, full implementations of the YOLO V3 and YOLO V4 networks are also released to DevStudio users.  These YOLO model ports are of the complete YOLO networks – all layers, all operators – running entirely on the Chimera GPNPU processor.  As with all networks that have been ported to the Chimera family of GPNPUs (spanning a range from 1 TOPs to 16 TOPs) the entire network always runs 100% on the GPNPU and there is no companion CPU or DSP required. 

Significant Compiler and Tool Advancements in Chimera SDK 23.10 Release

The quality and performance of code generated by the Chimera Graph Compiler and Chimera C++ compiler continues to make substantial strides with each release.  This latest SDK release adds a further 33% performance improvement to Quadric’s already market-leading performance levels on the ViT_B vision transformer, and a 13% performance improvement on the older RESNET18 benchmark network.  A full performance improvement chart showing the speedups for more than a dozen networks is available inside DevStudio.

A host of new and enhanced code development and debug tools are also now available to further speed the porting and debug process including a new tool for numerical validation/correlation of accuracy of results between Chimera code and ONNX runtime results, and enhanced performance visualization tools at both the operator and subgraph levels.

Immediate Availability

Current and prospective users can access these new features and benchmarks at https://studio.quadric.io/login

© Copyright 2024  Quadric    All Rights Reserved     Privacy Policy

linkedin facebook pinterest youtube rss twitter instagram facebook-blank rss-blank linkedin-blank pinterest youtube twitter instagram