At 45 nm, is there still a logic design hand-off? Design managers chime in.
Ron Wilson - September 10, 2008
A panel at the CDNLive conference in San Jose yesterday took on a question that has vexed managers of leading-edge design teams, and will within a few years puzzle many others: what is a logic hand-off like at 45 nm and beyond? Is there still a point in the design process where the RTL design team passes a discrete set of files to the physical design team? Or has what was once a clean hand-off become a group hug, or a free-for-all? Panelists Rajiv Sharma, technical lead at Cisco’s Core Router ASIC development group, Fred Jen, director of engineering for Qualcomm CDMA Technology division, Sandeep Mirchandani, associate technical director at Broadcom’s mobile and multimedia group, and Herve Menager, NXP Semiconductor corporate design technology architect addressed the issue.
The question arises because increasingly as we use finer geometries, side-channels are developing around the conventional design hand-off. These side-channels handle all sorts of vital information that is not included in the formal hand-off, according to the panelists. This is causing a breakdown in once-clean interface between RTL design and physical design.
It is also causing some rearrangement of tasks. One panelist pointed out that floor-planning, which used to be a physical-design task, has now moved into front-end design, almost as a transition between architecture and implementation. Conversely timing closure, which used to be the responsibility of the RTL team, has now passed to the physical designers, with the RTL group only using rough estimates to try to ensure that closure is possible.
Panelists felt that to some degree better tools would preserve the concept, at least, of a logic handoff. Better front-end estimation tools could, in principle, give architects, floor-planners, and RTL designers the ability to pass an RTL file or a placed netlist to physical design, knowing that it could be routed successfully, and that it could meet timing, and the power budget. Such estimators could potentially replace much of the information that physical designers must now provide to the RTL team early in the flow. Jen emphasized that such tools could head off problems particularly in intensively power-managed chips, with their proliferation of corner-cases, growing number of domains, and increasing sensitivity to variations.
Tools could help with information flow in the other direction as well. Design-intent-capturing tools could convey vital information about what the architects and RTL designers wanted to do through the formal handoff, Menager pointed out that, for example, all information about the intent of power management design used to be passed around the wall informally. But with the evolution of more tools, all this information can now be passed through the handoff in CPF or UPF files.
As to what the hand-off of the future would look like, exactly, opinions differed. One panelist stated that except on very straightforward, small logic designs, any attempt to do a traditional RTL hand-off would result in numerous iterations from physical design back to RTL or even floorplanning. Clearly that’s not an acceptable outcome.
But what to do? Menager suggested that instead of a hand-off, what we really have today is a handshake—a temporary merger of the front-end and physical design teams as they come to a mutual understanding of the data passing between them. Mirchandani argued that this is a good thing, not a bad one. "The hand-off is an important check point in the design—an important chance for everyone to stop and understand the state of the design," he said.
Sharma, whose ASIC teams as Cisco sometimes use outside physical design contractors, pointed out that organizational issues can influence the shape of the hand-off. Some systems OEMs regard the RTL as very sensitive data, as it could easily reveal proprietary algorithms. And so an RTL handoff to an outside contractor would be very problematic. Yet if the RTL and physical teams are in different companies, perhaps on different continents, informal communications become even more problematic.
If there was a single conclusion, it was that the front-end team must do whatever it can to be sure it is passing something to the physical design team that can be made to work. Predictive tools are one answer, but one slow to come and one that always raises issues of tool and file compatibility. An increasingly viable alternative is for the front-end team to have enough physical-design expertise to actually do trial placement, routing, and closure before formally passing the design over the wall. No guarantees, as there can always be surprises once the final closure process starts, but such local iterations could in principle prevent the much more expensive large iterations when the physical design team has to request major changes in the design.