CEO interview: Machine learning boost for analog simulation
A spinout of Oxford University is moving AI technology from fusion research into analog semiconductor design. Bijan Kiani, the US based CEO of Machine Discovery talks to Nick Flaherty
Machine learning has just raised $4.5m to develop its machine learning technology for analog simulation and prediction.
“The aim is to simplify and automate the verification flow and deliver real time prediction and in the as a companion to existing simulators and in the future add optimisation,” Kiani tells eeNews Europe.
“We believe the next step change will come form applying AI technology, We are focusing on cutting the product development cycle and the technology works throughout the design flow using AI powered models rather than spending months and months developing Verilog A models.”
- Machine Discovery raises $6m for semiconductor AI design
- Oxford startup promises fusion gain by 2024
- $29m for US AI research on fusion energy
The technology was originally developed at the University of Oxford for the physics of fusion reactors, and Machine Learning is part of the £12m project developing a small fusion reactor with and other Oxford spinout, First Light Fusion.
“We are addressing two markets,” he said. “Given we are coming from a physics background clean energy and fusion is our comfort zone and both the optimisation and real time prediction is being used in this at pace.
“I worked at Synopsys for 20 years and joined in 2020 because the cofounders backgrounders is stunning. When I joined with a semiconductor background we expanded our scope to semiconductors. Our critical mass for R&D is based in Oxford and our business and customer support is based in the US.”
Analog AI for prototyping
The analog virtual prototyping technology is trained in the cloud on signals from a SPICE simulator, giving the ability to then explore the design space.
“We use spice simulators and collect small sample data for a given design in a few hours or collect from silicon then we use that to train the emulator – this varies depending on the design, typically from a couple of hours to a day for a very complex design. Some run several hundred simulations in parallel on the cloud for the training,” he said.
“During the training we show how everything progresses and this emulator has a user interface and shows the design parameters with a slider bar to see the output. It’s a neural network technology developed internally to make sure it is credible,” he said.
“First you get the simulation waveform and you can identify areas to focus on and we can deliver accuracy of over 99% in these areas, and we go through training and show the progress and the accuracy so you can stop the training once you have the accuracy you want. Using the slider you can move the temperature range, the voltage, current. This is in contrast to a simulator where you have to run a separate simulation for every setting.”
“This gives us the ability to see the circuit performance at the click of a mouse with 90% accuracy on prediction vs the golden spice simulator. The beauty we have as we don’t need the netlist, the model is very secure so you are not exposing your IP.”
This also provides the ability to do data driven decisions on selecting IP, he says. “As we engage with customers they are finding new applications.”
One of these is as a companion to instrumentation for testing the silicon. This can identify the areas that need more testing rather than ad hoc testing.
- Synopsys adds generative AI data analytics across its EDA
- Thalia raises $2.7m for analog IP reuse tool
The AI-powered models can be used to create dynamic data sheets. Customers can train the emulator on simulation data and create a model for the end user that represents the entire design that the user can explore in detail,” he said.
“So far we have been able to demonstrate we can handle different design styles from teaching partners. We can handle typical designs available as standalone IP and power management.”
Optimisation technology is being used in the fusion project but not semiconductors.
“We haven’t rolled the optimisation technology into semiconductors but what we develop in the fusion space we use in the semiconductor space,” said Kiani.
The analog AI platform is available on subscription and can be accessed on the cloud, and developers can use the emulator for prediction. This is hosted in UK and US on Google Cloud and uses Nvidia GPUs for training while the emulator can be on a local CPU.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
