When the University of Michigan partnered with Arm Research on a project for the USA’s Defense Advanced Research Projects Agency (DARPA), there was a remaining technical gap the team needed to bridge. Arm knew the ideal partner from a longstanding collaboration: The School of Informatics at the University of Edinburgh’s software expertise would be the perfect complement to Michigan’s knowledge of hardware. Here, the key players from the two universities explore what it means to their institution – and their students – to work with Arm, DARPA – and each other.
While Moore’s Law has proven itself right up to the early 21st century, traditional transistor scaling is now unable to give the extra performance that cutting-edge compute requires. The hunt is on for heterogeneous systems that push that compute even further. The theory: you need to augment traditional cores with accelerators or different interconnects and computing styles, to eke out that sought-after performance within strict power and area constraints.
One organization driving these experiments is the US Defense Advanced Research Projects Agency (DARPA). DARPA is famous for taking a bold approach to achieving real-world results, through cutting-edge collaborations with innovative university and industry research teams. In 2017, DARPA launched the Electronic Resurgence Initiative (ERI) – which supports fundamental research into shaping this post-Moore’s Law future.
A key ERI project is the Software Defined Hardware (SDH) program. The SDH team worked to develop the transmuter, a reconfigurable computer that can radically adapt its internal structures to suit different applications. Previously, programmability has come at the cost of performance, but the transmuter should still allow data-intensive workloads to run at near- ASIC efficiency. Target workloads included graph and stream processing with application in big data analytics and neural network processing for accelerated AI algorithms.
When the University of Michigan’s Division of Computer Science and Engineering found itself invited to tender to run the SDH program, it brought in Arm Research, a long-time collaborator, which in turn looped in the School of Informatics at the University of Edinburgh, providing the latter with an introduction to the funding-rich US research circuit.
We asked the key players to walk us through that collaboration…
Ron Dreslinski, Michigan, principle investigator on DARPA’s SDH program: “The end of Moore's Law is the key driver for our work here at Michigan. We are focused on researching ways to improve on the scaling of traditional compute, so the DARPA program was a natural fit. We also have a strong history of taping out, of having chips manufactured, tested, and working, and there is only a small number of universities that have that capability.
These DARPA contracts are extremely competitive, but the combined academic and industrial team were able to make a persuasive case that any technology resulting from the program would enjoy a fast path to commercialization. We have a long history with Arm, so it made sense to bring Arm in, to leverage its commercial IP and its ecosystem of compilers, toolchains, and all the off-the-shelf elements that are critical in the development of a vast range of devices.
But while we had the hardware expertise, we were not addressing the software side. We asked our contacts at Arm if they knew anybody. They pointed us to these excellent researchers up in Edinburgh, doing great software stuff with Arm’s GPU IP, who would fit the hardware design we were looking at.
Mike O’Boyle, personal chair in computer science, Edinburgh: “Our school is like a super-sized computer science department. It is one of the largest in Europe, with 100 academics, 250 research staff and 400-500 PhD students. We do everything from computational neuroscience, building mathematical models of the brain, right down to designing processors. Our aim is to make a difference to industry. The real world matters to us.
The DARPA program offered a once-in-a-lifetime opportunity for us to do something really big, working with a world-class team to build a software stack for a new piece of hardware that would be better than anything currently out there. And to do so in a three-year timeframe. This would really bring that real-world focus to our research and make us see things in a different way.
DARPA also offers the kind of funding we would never normally get access to. It can be challenging to secure funding for a project of such scale, where you are developing test chips and getting everything going.
RD: “The connection with Edinburgh probably would not have happened without Arm Research. Arm knows all the academic institutions and what they are good at. And they can find the right ones to team with for a particular project. This was the perfect example.”
MO: “Arm saw us as two sides of a coin: Michigan is world-leading in hardware and design and architecture; we have a strong background in architecture and compiler software.
In this program we act as the glue between what the programmer wants to do, and what the hardware is able to do. If we did not understand what the hardware could do, there is no way we could reconfigure it. Meanwhile, Michigan needed to understand the applications and what kind of things we could do, so it could design the hardware to do it. There is no point having the fastest hardware available if the software cannot use it.
To this end, Arm provided us with the Revere Accelerator Management Unit, which allowed the two sides, hardware and software, to talk to each other seamlessly. Without such a clean interface to the IP, the project would not have happened. On the software side we are writing a software stack that is open source, so we do not have to worry too much about the IP. But the relationship there between Arm and Michigan has been critical.”
RD: “Arm provided us with IP and its software tools, including compilers, versions of OS that already run on Arm-based chips and drivers for devices. And, because its IP has that rich ecosystem around it, it is much more compelling than just using a custom core you designed, which is very common in academia because it can be hard to get an NDA in place to get IP from elsewhere. That gives us potential to be picked up by industry in the end, which again makes us more competitive.”
Trevor Mudge, the Bredt family Professor of Engineering, Michigan: “Arm is a unique company in that it is selling the IP, the blueprints, rather than the actual chip. And that fits very well with what we want: we want to get hold of those blueprints and Arm has let us do that. We have an arrangement that allows us to download copies of various components, especially the CPU designs it sells, which we can use to build our prototypes.”
MO: “We would not be able to do this with any other company in the world. No one else would be so open and helpful. Arm has provided simulators and emulators so we can all constantly check what we are doing. We were able to get to know the details of the GPU design, and we could work with Arm’s engineers to build a very fast, hyper-accurate simulator.
We spent a lot of time getting the engineering right in the first year, but now all the students are interacting, and even during Covid-19, we have been developing two or three paper ideas a month. It is been incredibly fruitful. We have developed the first prototype of the software stack, which is working on top of the emulator. And we are waiting for the chip to come back in around March 2021. The next stage is showing that these ideas we had in abstract really do work in practice, coding the software and hardware together, to see what performance options we get. And we would not have been able to do any of this if we did not have real hardware.
RD: “The other good thing about Arm is, it does not just provide the IP and the connections and general support; in our case it is a performer as well, actively working on the research too. Alongside Edinburgh, we are collaborating with someone from the Arm product team who is working on one of its interconnect IPs. He joins in our weekly meetings, providing support and updates on what is happening at Arm’s end. And he gathers information from us to make adjustments to the IP Arm is sending us, so it can continue to meet our requirements.
MO: “These close collaborations have become a key part of moving the tech forward. AI, and the explosion of embedded chips around it, are huge areas now – and that is the direct result of collaboration between academics and industry. If I look back ten years, industry and academia would meet each other at conferences and have a coffee together and that was it. Arm recognizes that academics can look at things in a way it would not. The academics are the ones firing thousands of ideas off, so it is really important for industry to have us around.
But it goes the other way too. We simply cannot compete with industry’s engineering efficiency and short-term roadmaps. In academia, our timescales are always ten years away, so if we are wrong, we are never going to be found out. These companies have market time pressures and have to make hard decisions in the now. They will have a six-month timescale with deadlines set by their customers, so there is no room to have three or four failed designs. I was quite amazed to see how productive and efficient industry can be, as well as the depth of thinking, the trade-offs, and the planning they do in the time available. If an academic is doing systems research with no industrial interaction, they will fail.”
RD: “The Arm team provides us very useful feedback: if we suggest an idea, they may come back and say that partners are not going to want to pick it up because the risk is too high, or it is not clear that it is going to help outperform something that is being developed elsewhere. They can steer us in a direction to create useful research, rather than just developing something that appears in a paper and never makes it anywhere else.”
Working with a group like Arm Research has been a tremendous opportunity for us. It has connected us up with collaborators, and brought a rich set of IP with a comprehensive software ecosystem to make our proposals stronger. And Arm is constantly working to reduce the friction of legals and NDAs to bring that IP on board and get it up and running quickly, so we can focus on doing the research we want to, rather than wasting engineering effort building the system around our research. By cutting down those overheads, we can focus our students' time and effort on the problems we find truly interesting.”
TM: “Students who have worked on programs like these become highly sought-after as graduates, because they have experience of things they would not normally get in a university. They are work on cutting-edge technology thanks both to Arm and the funding agencies we work with. The likes of DARPA tend to be very forward-looking, to the point where the projects we are working on seem almost impossible. That is a great thing for students to be involved with.”
Ronald Dreslinksi is Morris Wellman Faculty Development Assistant Professor of Computer Science and Engineering at the University of Michigan.
Trevor Mudge is Bredt Family Professor of Computer Science and Engineering at the University of Michigan.
Michael O’Boyle is Personal Chair Professor of the School of Informatics at the University of Edinburgh.
Arm Research is committed to supporting our academic partners. Find out how we have aided their research, and how their contributions have influenced us. University of Cambridge: Partners in Parallel How can Arm support your research?
Arm Research is committed to supporting our academic partners. Find out how we have aided their research, and how their contributions have influenced us.
University of Cambridge: Partners in Parallel
How can Arm support your research?