Enabling secure open source ML products with Open Se Cura

Today at the RISC-V Summit in Santa Clara, we’re pleased to participate in Google’s announcement of the open source release of project Open Se Cura. The announcement crowns a many-year collaboration towards developing a secure, heterogeneous multi-core ML solution based on RISC-V that serves as a template for subsequent commercial adaptations that can build on a solid open source foundation. The release includes both open source software, hardware, simulation-based co-development workflow using Renode that we helped bring to life, as well as open source RTL, closing the loop between a great concept to its practical implementation.

The goal of Open Se Cura, launched by Google today, is to provide an open source framework to accelerate the development of secure, scalable, transparent and efficient AI systems. The project was previously known as Project Sparrow, and in the previous years we have given presentations about it on several occasions, demonstrating subsequent developments in the methodology and workflow that powers Open Se Cura in anticipation of today’s full open source release.

Open source AI, from the hardware up

Google’s ambition in the project, convergent with Antmicro’s mission to enable complete vertical integration using open source, is to evolve a set of open source design tools and IP libraries that will accelerate the development of full-stack systems with ML workloads through co-design and development. This would enable system designs to center around security, efficiency, and scalability, empowering the next generation of AI experiences.

Open source has been central to the recent AI advancements and is a crucial base for safe and trustworthy AI applications. However, developments in the hardware space - and especially the open source flavors thereof - lag behind machine learning models and software development, which hinders collaboration and practical deployment of secure and efficient full-stack systems. The demand for secure, local, low-power AI still outpaces the supply which typically relies on a fragmented ecosystem of very vendor-specific solutions.

Antmicro has been working to remedy this with its open source developments including the Kenning edge AI optimization deployment library, but Open Se Cura, including an open source reference silicon design that can be integrated into products following a Renode-powered co-design methodology, takes this to a new level. We had already enabled the support for the Open Se Cura ML accelerator in Kenning via Renode earlier this year, and are excited for the upcoming capability to deploy ML models on real, open source ML hardware, and help implement the RTL and the accompanying methodology in real chips and products.

Collaboration and Antmicro’s involvement in providing Renode for full-system ML co-development

Open Se Cura is a collaborative project which, besides Antmicro, involves lowRISC, VeriSilicon and multiple academic partners like Cambridge University and University of Michigan that are coming on board to help bring Open Se Cura to new use cases and explore its incredible customization potential.

Antmicro’s role in the project was primarily focused on enabling the simulation-driven co-development methodology underlying Open Se Cura development with the Renode simulation framework.

As part of the collaboration, Renode was expanded with broad support for OpenTitan open source silicon Root of Trust, covering the most important parts of the SoC, including such subsystems as OpenTitan’s Big Number Accelerator, cryptographic infrastructure with RISC-V Vector extension and a range of accelerators, lifecycle management, entropy distribution features, and communication interfaces. Thoroughly tested against the OpenTitan extensive smoke-tests suite, this support became the basis of the Open Se Cura Continuous Integration and testing system, providing developers with more confidence and reduced iteration time of their work.

With the great introspection capabilities of Renode, the Open Se Cura team was able to benefit from improved execution tracing features for in-depth analysis of their executed software. Thanks to the detailed reporting features, they could fine-tune their Machine Learning software and toolchains to fully use the potential of the platform. Recently, the quest to provide comprehensive information to the hardware and software teams developing ML solutions based on Open Se Cura led us to developing an integration with Google’s Trace Based Model which can used to measure performance on a more detailed level, based on real microarchitecture implementation details.

Renode’s ability to run various software payloads as part of heterogeneous multi-core and multi-node systems allows users - both current and future - to develop their secure ML software end-to-end, including a complete boot flow and inter-core communication, as well as data ingestion and output via flexible simulation of sensors and actuators, etc. which will facilitate practical implementation in real AI scenarios.

Build complete ML systems with Open Se Cura

A big part of our excitement with Open Se Cura lies in the core competence of our company: on a day-to-day basis, Antmicro helps build AI-focused products. With the open source release of the complete flow and RTL, Antmicro can help companies embed Open Se Cura and its derivatives into ML focused products, just like it’s doing today with proprietary ML silicon - but with more transparency and control.

Antmicro offers a wide range of services including integration of Open Se Cura into custom SoCs, simulation-first co-development and testing, BSP & Software development for chips including Open Se Cura, developing edge ML boards and products with custom hardware designs based on our open source flow and hardware component database.

On top of those hardware-centric scenarios, Antmicro also supports customers with open source tooling for fast-paced RISC-V-based ML ASIC development and integration, AI application development including model training and optimization with Kenning and integrating these solutions into secure, real-world devices by providing custom OTA and fleet management systems based on our open source frameworks.

The Open Se Cura project will be presented today, November 7, at the RISC-V Summit at 1:55 PM. An extended technical deep dive will follow at the CHIPS Alliance Technology Update November 9th, at 9:10 AM at Google, Sunnyvale.

If you want to learn more about how Antmicro can help accelerate your AI system design, reach out to us at contact@antmicro.com.