At the RISC-V conference late last year, one of the keynotes was by Linton Salmon titled A Perspective on the Role of Open-Source IP in Government Electronic Systems . It was not specifically about RISC-V, although the RISC-V ISA and many of the implementations to date (but not all) are open source. Linton started giving a little of his background. He is a Program Manager at DARPA/MTO. These program managers are people pulled from academia or industry to work for the government for a few years. His background is in both. He was a professor in EE and physics at BYU and Case-Western and then had 15 years in different executive positions directing R&D at GF, TI, and AMD. He was even a member of the steering committee on the first ITRS. He calls himself "middle-aged" meaning that he has been in his government position for three years. The institutional memory is all in the contractors, the program managers are all "temporary". The challenge that the government faces is that the Department of Defence (DoD) needs custom SoCs since the commercial world doesn't make everything that they need. There are often factors of 1000 or more improvement in power for a custom SoC versus just programming a processor with software (say 4W on an ASIC, 1200W on a CPU). Since their volumes are low (thousands of parts), over 90% of the cost of a program is in design and only a little in the manufacture. As he said, making it concrete: In DoD, I've yet to see a total production that is more than the samples that we would send out in the commercial world. So putting all that together leads to the obvious conclusion that DoD should be doing something to reduce design costs. How's that going? Design costs are skyrocketing. Which, in turn, means DoD program costs are skyrocketing. In the commercial world, people use IP to address complexity. For example, a Qualcomm Snapdragon phone chip is a compilation of IP from many sources (foundation IP, Qualcomm internal IP, licensed IP, and third-party IP). A 2016 average SoC has 175 IP blocks, of which 80% are reused. Is Open-Source IP the Answer There is good news, at least potential good news: Open-source IP can sharply reduce the resources, time, and complexity required for DoD custom SoC design Open-source IP can enabled SoC specialization to unique DoD applications Detailed information available with open-source IP can enable greater trust Open-source IP permits increased use of unique DoD security approaches Security is actually a big one. The government often wants what commercial already has but with the secure-o-meter turned all the way clockwise. In particular, at the IP level, they need to ensure that the block does what it is supposed to do and only what it is supposed to do. Commercial IP doesn't really give you the ability to check, but open-source lets you do whatever due diligence you want. Further, hardware security often requires the ability to modify the third-party IP (security against tampering, security against counterfeiting and cloning, security against cyber-attack). As he put it: it is easier if we can modify it without having to convince a large corporation to move their whole product line Another challenge is that design teams in DoD are small, and they only do one design every few years. That means that they never get good. So it is great when methodology experience is embedded in the tool chain, and design experience is embedded in IP, especially standards-based IP enabling everything to plug-and-play. DARPA is interested in RISC-V, given all those constraints, since they often just need a "C engine". Lots of DARPA programs have died because they couldn't do the software: the reality is that it is a lot harder to change software than hardware…we should swap the names A little aside: when I was doing my PhD, we came up with the term "stiffware", which was software for which the source code had been lost, or the right version of the compiler no longer existed, such that it could no longer be rebuilt. I'll bet the DoD has lots of stiffware. However, in addition to the good news, there is bad news: The open-source community needs to develop a complete infrastructure Open-source IP needs to be more robust than is currently the case The community needs to develop a model to fund infrastructure The support model must assure long-term support and continued development of open-source IP He looked at each of these in turn. Full Ecosystem Infrastructure It takes a village...or in this case an ecosystem. Linton apologized for showing an Arm slide at a RISC-V workshop but that's what it takes. Common infrastructure could be a great strength for RISC-V but it has to work: ISA, common processor implementations, interconnect and associated IP, EDA tools. Robustness and Dependability DoD requires robustness and dependability of the open-source infrastructure: complete verification, clear documentation, silicon implementation, and test. One big issue is that this is neither the role nor the strength of universities. That's before worrying about the DoD's different concerns such as operating conditions (extended temperature, etc). Model to Fund Long-Term Infrastructure First, successful open-source infrastructure requires a lot of work: not just work to design the correct IP modules, but also work to maintain, upgrade, and facilitate use of the infrastructure. It is a lot work. Not just that, it is a grind, difficult but not exciting, and certainly not the source of PhD dissertations. But DARPA funds projects, not longer-lasting infrastructure. As Linton put it: please don’t ask me why DARPA doesn’t do that. We don’t do infrastructure. If we did then we would still be stuck on some old infrastructure 30 years ago Improvements over Time But somehow, if it is going to work for the DoD, there need to be improvements over time. DoD timescales are 10-25 years. Moore's Law has driven the expectation of continual improvement. and a static technology can’t be used. So it needs to update regularly. DARPA Programs Driving Open-Source IP Power Efficiency Revolution for Embedded Computing (PERFECT) Circuit Realization at Faster Timescales (CRAFT) System Security Integrated through Hardware and Firmware (SSITH) Posh Open Source Hardware (POSH) Intelligent Design of Electronic Assets (IDEA) Q&A In the Q&A, Dave Patterson (RISC legend, these days of Google) said that a lot of Linton's concerns about open-source hardware don't seem to apply to software. Linton said that open source has worked marvelously in software. "I can find out if software will work but we don't have that in hardware yet." It is not just money, but recognition. People establish their reputations through open-source software and become employable. That is starting to occur in hardware. "A lot of people in government believe we can afford to buy it all, drive it all, but we can't. Semiconductor is a $300B business, we are so tiny." Robert O'Neill (Western Digital) pointed out that some of the goals for government and industry are aligning with autonomous vehicles and wondered if that is something that DoD can take advantage of. Linton said that it is work in progress. Most of the contracts for CRAFT and SSIHF will not be with traditional government contracts. Most commercial companies won't do it since it costs too much to set up the paperwork. DoD understands that on the industry side it also has to meet their commercial needs. But lots of opportunities for win-win. Ed Humenberger (Symbiotic EDA) said that Linton had only talked about ASICs, but wondered what happened for people who couldnt afford a $100K "cheap" tapout. Linton loves FPGAs, even though he is often accused of being an "ASIC bigot". But "we lose a factor of 50-200X in higher power for same performance, which is a real problem in many applications. It seems like in a jet fighter why worry about a few watts given we are throwing flames out the back. And we also care for the guy schlepping the batteries. When power is the limiting factor, I want an ASIC." Sign up for Sunday Brunch, the weekly Breakfast Bytes email.![]()
↧