Using Machine Learning for Characterization of NoC Components
Synopsys Users Group (SNUG), March 2019, SNUG Silicon Valley
Download this conference presentation titled, "Using machine learning for characterization of NoC components”, presented by Benny Winefeld, Arteris IP's Solutions Architect , at the Synopsys Users Group (SNUG), March 2019, SNUG Silicon Valley.
- Conference presentation, 33 slides
- Abstract: Modern NoC (Network-on-Chip) is built of complex functional blocks, such as packet switches and protocol converters. PPA (performance/power/area) estimates for these components are highly desirable during early design phases – long before NoC gate level netlist is synthesized. At this stage a NoC component is a soft module, described by a set of architectural parameters, like the bit width of ingress and egress ports, number of virtual channels, etc. The proposed approach attempts to predict the PPA behavior of NoC components based on machine learning non-linear regression algorithms. The system consists of several layers. Atthe bottom Synopsys Physical Compiler is used to synthesize a NoC component with one combination of input parameters (features) and capture its characteristics. This result becomes a data point in a training set. When it gets sufficiently large, this set is being used for trainingfast models predicting PPA for components with parameter values not exercised during the training. These models can be plugged into a NoC design tool assisting the user with feasibility and what-if analysis.
- Author: Benny Winefeld, Arteris IP