info@embecosm.com

I spent last weekend, with my colleague, William Jones at FOSDEM in Brussels. This is an annual pilgrimage for me, to mix with working open source engineers and here all that is new. I was also a speaker, sharing some of my own work on compiler development.

For me the most significant devroom was Low level AI Engineering and Hacking. This is only the second time FOSDEM has offered an AI devroom, and once again William Jones of Embecosm was one of the co-organizers. The two of us presented our work on Accelerating AI with open source hardware and software, one of the big growth areas for our company. AI is now pervasive, and so I spoke with Frédéric Desbiens of the Eclipse Foundation about the key role the Open Hardware Foundation in our talk on How to quickly build an AI startup on open source RISC-V Cores as part of the RISC-V devroom. Meanwhile William Jones was speaking with Ian Potter, a highly experienced shool leader, on Open Source in Education of Neurodiverse Students. Once again, development of AI systems was central to engaging students and the success of this initiative.

But not everything was AI. My entire career has been focused on computer performance, primarily through the role of compilers. I investigated this in my talk, How good is RISC-V: Comparing benchmark results, in the RISC-V devroom, where I looked at how well the compilers for RISC-V are performing. Not just in how fast the code runs, but in how small the code can be (which really matters for the embedded engineering community). The good news is that performance is improving, both over time, and through the various ISA extensions being added to the RISC-V architecture. And for embedded processors, the code size performance gap between RV32 and Arm Cortex M is closing, albeit very slowly.

But it is not all good news for RISC-V: for server class machines, AArch64 and x86_64 architectures still score much better on the SPECMark™ benchmark suite than RV64. And the new RISC-V Vector extension does not appear to be helping. Preliminary analysis using instruction counts, suggests that, while it improves one of two benchmarks significantly, for most it has not effect, and for a number it makes things worse!

I have for a long time been interested in how you measure the success of open source communities. Almost all established communities have goals they wish to achieve, but how well do they achieve them. I touched on this in my talk, noting that 90% of the work on RISC-V GCC is down to just 15 engineers worldwide. And, despite RISC-V International having more than 4,000 member organizations, just 12 of those organizations contributed to the development of GCC for RISC-V. This is a subject I shall revisit later this year.

Finally, open source depends on engineers willing to give back to the community. My last talk of FOSDEM 2025 was in the GCC devroom, where I gave a tutorial on How to add a builtin function to the GCC backend. For many years we have had an in-house training course for new graduate employees on compiler development. That course lasts 3 months full-time, but I try regularly to take small parts of the course and present it at meetings like FOSDEM.

Once again, I have returned from FOSDEM refreshed and full of new ideas to try in the coming year.