Chip and system design environments will undergo a radical restructuring during the next several years, if Richard Newton, dean of engineering at the University of California at Berkeley, is correct. Newton's attitude toward EDA these days is let's figure out what we're designing first and the "how" of the tools will follow.
Newton believes that new silicon methodologies, such as platform-based design, will cause a complete overhaul of the RTL-based ASIC design flow that dominates the EDA industry today.
As former head of Berkeley's Electrical Engineering and Computer Science Department, Newton has helped shape the EDA industry during the past two decades. During that time, a well-worn design flow has developed-one that encompasses RTL synthesis, simulation and physical layout.
Now, Newton believes, we have reached the end of a "methodology" cycle and are no longer sure just what we'll be designing several years from now. We can't predict what tools will look like until we figure out the methodology, Newton believes.
Still, he said, some general conclusions can be drawn. Newton predicts that hardware and software design will move much closer together, silicon sign-off will occur at a very high level and designers will work in the "language of the problem," not Verilog or VHDL.
"At a certain point, Moore's Law creates enough change that the fundamental problem is methodology," Newton said. "We're just now sorting out what the methodology will be. Change will be driven by methodology and the tools will follow."
One promising new methodology, Newton believes, is platform-based system-on-chip design. Here, hardware architectures are largely predefined and designers create differentiation through programmable logic and embedded software. "I think large markets will be addressed by programmable substrates," Newton said. "They'll be domain-specific platforms."
Another candidate is what Newton calls the 24-hour chip, an idea now being developed by Bob Broderson, professor of engineering at Berkeley and chair of the Berkeley Wireless Research Center. The goal, said Newton, is to start with a Matlab description, go through automatic synthesis and get a power- and performance-optimized layout in 24 hours.
Either concept could change the world, Newton said. "Both will probably happen, for different market segments and different problems."
|At Berkeley, Richard Newton, dean of engineering, is taking a new look at what it means to be an engineer: 'Everyone is working across boundaries.' |
Whatever methodologies prevail, Newton believes there are several key challenges facing designers in the 21st century. One such challenge is managing concurrency, which is essential in both hardware and software to overcome latency in high-performance systems.
"Humans can't visualize concurrent systems well and come up with reliable solutions," said Newton. "That is a key research problem we have to solve and it will apply to both hardware and software."
Another big challenge, Newton said, is "building hardware with a deeper understanding of the software context and building software with a deeper understanding of the hardware context."
Looking toward software
Newton believes, in fact, that the next big market for the EDA industry is embedded software development. Here, he sees a need for a new generation of verification tools. "The challenge is how to verify the correctness of systems implemented as software with a lot of concurrency," he said.
One might conclude that Newton is an advocate of hardware/software codesign, but he's never liked that term. "It casts the problem as hardware vs. software, as opposed to a functional problem that gets decomposed into hardware and software as a process of implementation and evaluation," he said.
Newton also said that "system-level design" is a term that means too many different things to too many people. The real issue, Newton said, is the ability of designers to accurately predict power, performance, throughput and application efficiency for a very complex SoC, at a much higher level of abstraction than is possible today.
"Speaking the language of the problem, you need to reliably predict the performance of the silicon," he said. "That's a major research problem we have to solve."
The days of RTL coding are numbered, Newton believes. Instead, he said, designers will write "in a language more specific to the domain at hand." This language could be a set of C language packages and routines, or something else. But it needs to support concurrency and manage memory automatically, he said.
As for silicon signoff, Newton believes it must move well beyond the gate level. Reiterating the views of the Gigascale Silicon Research Consortium (GSRC), a multiuniversity research project that Newton heads, Newton said that signoff should move to "the boundary between architecture and microarchitecture."
In GSRC's lexicon, an architecture is an abstraction that's independent of implementation, while a microarchitecture has an implementation and an instruction set. Or, as Newton puts it, a microarchitecture is the point at which "you have a block diagram that looks like the chip."
Supporting this level of signoff requires incredibly accurate prediction and estimation tools. "We're talking about being able to make final design decisions at the architecture-microarchitecture boundary and never stopping to think twice," Newton said.
What about the skill set that will be needed for next-generation electronic design? Newton is thinking about that a lot in his new position as dean of engineering. He is, in fact, working on what he describes as the first major overhaul of how universities teach engineering since the Second World War.
Newton is overseeing top-rated departments in electrical, mechanical and civil engineering. One big problem faced by the School of Engineering, he said, is that two-thirds of the most qualified students want to go into the Department of Electrical Engineering and Computer Science.
Mechanical and civil engineering departments aren't getting the students they need, even though some very interesting work is going on in these areas. For example, Newton noted, much of Berkeley's microelectromechanical systems research takes place in mechanical engineering, and civil engineers are trying to wire every bridge in California to help predict earthquakes.
"We're starting to take a new look at what it means to be an engineer," Newton said. "It's not about civil or mechanical. Everyone is working across boundaries."
At Berkeley, Newton hopes to launch an ambitious program that would erase some of those boundaries and put engineering students into multidisciplinary research groups. This might entail a common engineering program for the first year or two, with specialization afterwards.
"We've lost the binding between the names of the departments and the types of problems people solve," Newton said. "We need to morph the structure into one where kids will identify more with what they're doing. The 'how' will follow."