An artificial intelligence/machine learning (AI/ML) development kit has been announced by Mentor, a Siemens business, with added AI/ML enhancements to two tools to help its customers deliver smarter, AI/ML-powered ICs to market faster.
The new Catapult software High-Level Synthesis (HLS) AI Toolkit and HLS ecosystem are designed to help customers jumpstart the development of complex machine learning IC architectures.
Meanwhile, Mentor has also announced it is adding AI/ML infrastructure throughout the Calibre platform, and is launching the first two of these AI/ML-powered technologies: Calibre Machine Learning OPC (mlOPC) and Calibre LFD with Machine Learning, both of which leverage machine learning software for faster, more accurate results.
These new offerings further expand Mentor’s fast-growing portfolio of AI/ML-powered solutions. Last year, Mentor acquired Solido, a pioneer in AI/ML-enhanced EDA tools. Solido’s customers include 15 of the world’s 20 largest global chip design firms.
More than 3,000 design engineers use Solido’s AI/ML-powered tools to drastically speed development for many of the world’s most popular and sophisticated semiconductor designs in production today.
“It’s becoming evident that a vast majority of products being developed for the foreseeable future will incorporate AI/ML in some capacity – what’s smart today will become smarter with AI/ML,” said Joe Sawicki, Executive Vice President of IC EDA for Mentor, a Siemens business. “Mentor is committed to developing solutions with functionality that will help our customers more easily integrate AI/ML into their products.
“In addition, Mentor is incorporating adaptive machine learning into our own tools, and as a result customers are seeing vast improvements in runtimes and accuracy, which in turn further enable our customers to deliver their smarter, AI/ML-powered technologies to market faster.”
JY Choi, Vice President of Foundry Design Technology Team at Samsung Electronics, added: “Samsung has been collaborating with Mentor on Calibre LFD with Machine Learning to improve full-chip accuracy and performance, in addition to other ML-powered solutions from Mentor such as Tessent YieldInsight software, the Solido Variation Designer and the Solido ML Characterization Suite.
“With the new Calibre LFD with Machine Learning, we realized a dramatic boost in performance and a 25 percent improvement in accuracy over earlier Calibre LFD solutions.”
The New Catapult HLS AI Toolkit
Mentor’s new Catapult HLS AI Toolkit is designed to help customers developing AI/ML-based accelerators for edge applications get to market faster. Based in easy-to-use HLS C++, the toolkit provides an object detection reference design and IP to help designers quickly find optimal power, performance and area implementations for neural network accelerator engines, a task not possible with hand-coded register-transfer level (RTL) designs.
The solution also includes a complete setup to build an AI/ML demonstrator platform, with live HDMI feed on an FPGA prototyping board. The toolkit is a key component of an expanding Catapult HLS ecosystem for AI/ML applications, which also includes open-source HLS IP, TensorFlow integration, easy system integration with interconnect that implements the Arm AMBA 4 AXI interface, and HLS On-Demand training and consulting.
Mickey Jeon, Chief Technical Officer for Chips&Media, said: “As a provider of high-performance, high-quality video and computer vision IP, incorporating deep neural networks (DNNs) enables us to deliver differentiated quality and functionality to our customers.
“For these DNN-based IP cores, it is key to have a hardware architecture that is highly optimized for power, performance and area (PPA). Mentor’s HLS technology enabled us to do this extremely efficiently. Our project was an outstanding success, and we plan on deploying an HLS flow using Catapult on our next project.”
Enhancing Calibre Platform with AI/ML
In addition to the Catapult AI Toolkit introduction, Mentor’s Calibre group is diligently integrating machine learning on a massively scalable architecture targeting full-chip manufacturing databases. These capabilities are being leveraged across the existing Calibre tool base to increase performance, accuracy and capacity while also enabling advanced customers to craft custom applications.
The first of these AI/ML-powered Calibre tools commercially available and in use today are Calibre mlOPC for optimized optical proximity correction and Calibre LFD with Machine Learning for advanced lithography simulation.
The new Calibre mlOPC product provides 3x faster OPC run time compared to the prior Mentor optical proximity correction technology.
mlOPC utilizes fast, intelligent feature extraction and machine learning algorithms to predict the OPC output to within a single nanometer of accuracy, and in the process eliminates up to 75 percent of OPC run time.
Calibre LFD with Machine Learning
Mentor has also added a machine learning option for its lithography simulation tool. The new feature is engineered for high accuracy and improved performance on large blocks as well as full-chip analysis.
The feature’s predictive capability focuses on high-risk layout patterns for detailed lithography simulation, removing low-risk patterns from this compute-intensive step. The result is a 10 to 20x performance improvement over full chip-model based simulation while maintaining optimal accuracy.
Building on a foundation of AI/ML leadership in EDA
These new solutions add to an already deep portfolio of AI/ML-powered tools from Mentor, including:
- Solido Variation Designer, which incorporates adaptive machine learning for variation-aware custom IC design. Used by analog/RF, memory and standard cell designers, Solido Variation Designer can improve design performance, power, area and yield, giving accurate, full corner and statistical variation coverage in vastly fewer SPICE simulations. Variation Designer helps customers speed SPICE simulation cycles between 10 to 1,000,000x faster than traditional brute force approaches.
- Solido ML Characterization Suite, which uses machine learning to perform comprehensive validation of characterized Liberty files, and accelerates the generation of Liberty models and statistical data for standard cell, custom cell and memory cell libraries. The solution’s analytics tool uses machine learning to identify outliers and critical issues in characterized data, and visualizes library information while enabling the debugging of issues using an interactive environment. The suite’s Generator tool uses machine learning to speed up characterization time by 2-4x.
- Tessent YieldInsight software, which uses machine learning to identify and understand the root causes of yield loss from scan test data. This helps operations teams identify systematic yield limiters, prevent wasted physical failure analysis (PFA) on known issues and eliminate costly physical localization in the PFA process.