3D ICs are an attractive technology, but what will it take
to make them successful? Presenters at the recent Electronic Design Processes (EDP) workshop
didn't have all the answers, but they had a lot of interesting insights into
how EDA tools and flows will need to change to support stacked die with
through-silicon vias (TSVs).
Salvation for Moore's Law?
In the workshop's opening keynote address, entitled
"An(other) Inconvenient Truth," Tom Williams (a well-known design for test
expert, now retired) talked about some of the inherent problems with the
"nanometer rush" fueled by Moore's
Law. As we move to lower process nodes, he noted, there are increasing problems
with Ioff current, power density, leakage, and variability. Gate
utilization is declining at each new technology node, and increased costs mean
that revenue will shrink if volumes don't double at each new node.
"More of Moore means more
than just Moore,"
Williams said. "It needs some creative thinking in another dimension." That
dimension, he suggested, is 3D. He said that 3D ICs will offer much easier
analog/digital integration, smaller footprints, higher bandwidth, shorter
global interconnect, better timing, and lower power.
But there's much to be done, he noted. One challenge is
test, and the solution Williams proposed is built-in self test (BIST). 3D power
analysis, thermal analysis, and floorplanning will become important as well.
The bottom line: "The third dimension will keep Moore's Law alive and well."
Making 3D ICs
Rahul Deokar, product manager for the Encounter Digital
Implementation System at Cadence, offered a presentation entitled "3D ICs, are
we there yet?" He noted that "cost is the new driver in today's world," and
that moving an entire system to a new process node is risky and costly. With a
3D IC, he noted, you can still build your CPU on a low process node and keep
analog and memory on a mature process node.
Much will need to change, however, in the EDA world,
including new layout rules, a new backside layout layer, new floorplanning
rules, and thermal and mechanical constraints. Customers are especially
concerned about thermal issues when die are stacked on top of each other. "It's
a heat pocket, and you've got to figure out how to release that heat," Deokar said.
There's more. Deokar outlined 3D IC design flow challenges
including system-level exploration (how do you stack the die?), 3D
floorplanning that optimizes TSV locations, placement and routing with TSVs,
extraction and analysis across multiple die, and new approaches to design for
test (DFT). He also stressed the need for a 3D ecosystem including IP providers
and foundries. "Indeed it takes a village," he said.
Some Tough Challenges
One thing that hadn't occurred to me is how big TSVs really
are. Sung Kyu Lim of the GTCAD Laboratory
at Georgia Tech illustrated that very dramatically in his presentation. The
slide below shows that a TSV with landing pad and keep-out area is several
times larger than gates and memory cells:
Lim went on to identify a number of issues with TSVs:
they're so large, TSV count is critical. As you use more TSVs the wire
length goes up.
location is critical. There's often a tradeoff between timing and
cause coupling. To reduce it, you can add more space to the keep-out zone,
but then you add to area.
require design for manufacturability (DFM). They cause stress and CMP
issues, but no litho problems, fortunately.
require DFT. Many questions remain here, such as how to test the TSVs in
each die and how to provide power to each die.
It's still worth the effort, though. Lim spoke of a recent
3D IC with 64 processors that GTCAD recently taped out - possibly the first
many-core 3D processor from academia. If you want to know more about TSVs, Lim
published a recent article
on the DAC web site that summarizes TSV-aware tool needs.
There have been some recent blogs on 3D ICs as well as EDP.
Here's a quick summary:
Rahul Deokar posted
a blog on his impressions of 3D IC discussions at EDP. Samta Bansal of
Cadence recently wrote
about 3D IC activity at the DATE conference in Europe.
Clive Maxfield wrote about Cadence's approach to 3D ICs in a blog
Harry the ASIC Guy (Harry Gries) posted an interesting
overview of EDP. I wrote an Industry Insights blog
about an EDP debate on parallel programming for EDA tools. Main problem
with EDP: Too much good stuff, too little time.