Xilinx adds ML, incremental compile to FPGA design tool
Xilinx has added incremental and nested compilation as well as machine learning algorithms to the latest version of its FPGA design software.
The incremental compilation in the Vivado ML design tool allows parts of the FGPA fabric to be compiled independently at different times. This allows blocks of IP, for example from a partner or a customer, to be added entirely separately into a more complex design. Customers such as board maker Abaco are already using the tool for this, while test system developer Keysight is using it to allows different internal teams to develop separate blocks completely independently.
The tool both uses AI and is aimed at developers working on AI designs that would be implemented in its FPGAs in data centres or for edge computing.
“This is the next generation Vivado,” said Nick Ni, director of marketing, software and AI at Xilinx. “AI requires orders of magnitude more compute performance so there’s a huge need with designs with billions of transistors.”
“EDA is mainly heuristic algorithm designers and for them to adopt ML there is a talent gap so this is about how to apply ML for EDA tasks,” he said. “Xilinx has been working on the research for a while and finding a way to productise it. The design iteration aspect is something we are pretty excited about as ML has the intelligence to help pick the right strategy early on so it has big promise.”
The tool breaks the design down into smaller pieces from the beginning to the end into small reusable pieces. These are then compiled separately, speeding up the compile time of the final design by a factor of five, says Ni and boosting the quality of results (QoR).
“We are deploying a number of ML techniques, with an intelligent timing closure tool,” he said. As the chip gets bigger timing closure is more challenging but this is a single button push. It automatically applies machine learning algorithms for congestion estimation and delay estimation to optimise the design based on past experience. This will get closer to timing closure after 4 or 5 iterations then apply strategy prediction for the ‘last mile‘ closure, and try the top strategies from 60+ designs to find the appropriate approach to hit the target frequency.
Next: Machine learning strategies
The Vivado ML tool uses a number of different machine learning approaches, from a traditional neural network for congestion estimation to the strategy prediction.
“Most of the training data comes from our designs as we have been running huge amounts of regression testing, Sometime we use customer designs with permission to make sure the tools do what they want. Many are designs are half the size of our largest devices,” said Ni.
“For deployment we have an approach we call dynamic function exchange (DFX) to change blocks without taking the system down. This uses reconfigurable modules (RM) with a feature called secure handoff. Some customers have to handoff some region of the FGPA for further customisation and this hides everything other than the RM. For example with FPGA as a service, this leaves a big RM for developers to use. For other use cases, for genomics for example, its pretty evenly split with different RMs to quickly reload different algorithms on the fly. The smaller the RM the bigger the benefit on compilation, and we see an average 5x faster compile,” he said.
“It’s a feature we had before but we have enhanced it with a new feature called Abstract Shell that allows the RM to be compiled truly independent,” he said. “We are also extending the abstract shell to the back end and the front end. Block design container uses this as a module inside an IP graphical flow and these can directly cascade through the systems.”
This idea of replaceable, compilable blocks throughout the design is being further explored with nested DFX where a DFX block is compiled within another DFX. “That’s something we have deployed as a feature but is still in a development programme,” he said.
Related Xilinx articles
- Xilinx targets Versal FPGAs at edge AI
- Xilinx extends FPGA dynamic reconfiguration capabilities
- Debug tool steps up support level for Xilinx FPGAs
- Custom FPGA configurations to offload and accelerate image processing
Other articles on eeNews Europe
- Intel eyes Bavaria for wafer fab
- Micron, Mouser sued in Germany
- SiBrain module takes on microcontrollers
- €8m project for Europe’s first RISC-V supercomputer chip
- Delivering power from the back of a chip
- Memory module halts Hubble