Ed’s Threads 080304Musings by Ed Korczynski on March 4, 2008
DFM matures along with industry
Hundreds of technologists over-packed the room in the San Jose Convention Center at 8am on the fourth day of SPIE
to hear keynotes from IBM, Intel, and TSMC on the real reality of design for manufacturability (DFM) in the IC fab industry. As the two leading integrated device manufacturers (IDM) in DFM, IBM and Intel provided thorough overviews of technologies and methods used at 65 to 45nm nodes. In contrast, TSMC gave what seemed like a sales pitch which was not well received by the audience of peer technologists. A raucous panel discussion that evening raised the need for a modeling environment to test new DFM approaches in virtual space and time.
“It’s always possible to increase yield by throwing money at the problem,” declared IBM’s Lars Liebmann. “We need to keep costs under control because we’re chasing incremental yield.” One of the most costly aspects of implementing DFM today is quantification experiments to prove the value of a considered new technology. “You have to convince management that yield will increase, and that value is unique to the product, the time in the current manufacturing node, and the business goals,” said Liebmann. “…there is no universal DFM.”
New DFM tools build upon the proven models used in the past. Critical area-analysis (CAA) is one of the oldest predictors of mature device yield
, since the area of the wafer subject to failure due to random physical defects (such as particles and scratches) can still be accurately extracted from any new design. “While there’s a lot of hype about systematic problems taking over from random problems, CAA is still an excellent indicator of yield,” said Liebmann.
Intensive number-crunching will be needed for design-technology co-optimization, and one example that has already been demonstrated is electrically-driven optical proximity correction (OPC). Instead of tuning a mask to produce optimized shapes, the mask is tuned to produce shapes with optimal electrical performance.
IBM showed that using highly restrictive design-rule (RDR)—specifically pdBrix layouts created by software now owned by PDF Solutions
—at 65nm created dice with the same area and performance with fewer hotspots and less variability in a PowerPC405 core.
“DFM has sort of just barely scratched the surface,” said Liebmann. “We’ve just reached our teenage years and the best years are still ahead.”
In his keynote address, Intel Fellow Clair Webb
explained that his group does a lot of simulation and modeling of design rules, such that the first test-chip is expected to confirm and calibrate the rules which are not supposed to change. Intel ramps processes to very high volumes very quickly, so the process must be very robust, and a very fast yield learning rate is essential.
What is really meant by co-optimization? For Intel, the many factors to be considered include the characteristics of litho tools, resists and illumination sources, tape-out technology, mask processing, device performance targets and architectures (incl. variability requirements), and ultimately even the product targets (incl. power/density, cell layout, time to market, CAD tools available, etc.).
Design rules for pitch start with a 1D target set by first-order density goals. “All the fun comes with the 2D targets,” quipped Webb. Starting with learning from the previous process, Intel then extrapolates 2D models for OPC, illumination techniques, reticle enhancement techniques (RET), and photoresist for critical parameters (e.g., DOF, MEEF, etc.). All of this leads to an OPC/litho test-chip to quantify the models for things like new off-axis illumination techniques. “The test-chip is the outcome, it is not part of the modeling process,” explained Webb. By the time the first design hits the fab, 80% of the design-rules should already be set. “We may take learning and feed it forward into the next process, but we’re not going to change the rules at ramp,” explained Webb. “If I have to do design-rule changes at ramp then I’ve made a mistake.”
For example, modeling variations in line-length on MEEF and CD with different illumination sources showed that line-lengths between 0.2 and 0.3 µm created problems. The real-world DFM trade-off involves checking back with the designers to determine whether they really need lines of this length, and searching for another interdependent parameter which can be constrained with a rule to eliminate the MEEF problem.
Intel’s Webb explained that 90 nm was the first time that the design rules started to change dramatically with each node, where there was a 47% increase in the number of rules for poly. At 65nm there was a 65% increase in poly rules, primarily to enable PSM and to handle proximity effects, though there was still variable poly pitch and width, and two directional routing. By 45nm, Intel has gone to the extreme constraint of gridded layout rules (GLR), and the total number of design rules went down 37% compared to 65nm. “It’s hard to measure the results of any one particular rule, since it would take thousands of wafers,” explained Webb.
In regards to the trade-offs between design and process, “As a foundry, we say the customer wins a lot…which means the design wins,” said Dr. Fu-Chieh Hsu, vice president of Design Technology Platform, TSMC. “We’re always challenged by designs pushing the limits of design rules.” Since the foundry supports legacy processes, as well as half- and quart-nodes for customers, TSMC sees a continuum of process technologies instead of discrete jumps between nodes. TSMC therefore sees broad general trends in process-design trade-offs. TSMC DFM solutions start with a DFM-design-kit, and include certified DFM-compliant EDA tools and 3rd-party IP, all of which has been used on over 113 tape-outs based on ~1000 IP validated blocks.
In a lively evening panel discussion moderated by Mark Mason (TI) and Juan Antonio Carballo (Argon Venture Partners), the prevailing sentiment seemed to be that of hope over hype. Joe Sawicki (Mentor Graphics) and Srinivas Raghvendra (Synopsys) provided perspective on the business constraints of commercial EDA vendors, while the playfully soft-spoken Riko Radojcic (Qualcomm CDMA)
expressed the perspective of the designer.
Regarding the challenges faced in attempting to model manufacturing variability and then feed that information back to designers in some way, Radojcic opined, “The two communities speak entirely different languages. If a manufacturing guy says this is the variability that you have, what does the designer do with it?” Radojcic advocated for a simulation environment which could be used to explore DFM options in virtual space and time, instead of waiting for expensive “spins” in silicon.
Radojcic said that the main limitation to the use of new DFM tools is quantifying benefits, “Trade-offs in area, variability, yield, and cost at the whole chip level are nightmares, so we all just shrug our shoulders and keep doing what we did before.” Raghvendra replied that, “We’re making progress towards solutions that are holistic, where you can look at the whole picture, and it’s not like you look at timing and lose power.”
Extensive DFM will continue to be needed for ICs made using less than quarter-wavelength lithography: 45nm and below for 193nm litho tools. The trouble really started at half-wavelength (~90nm), and unless EUV (~13.5nm) becomes an option to get patterning back to super-wavelength litho, the world will need more and more DFM going forward. IBM’s Liebmann says that high-index 193nm immersion won’t be ready for the 22nm node, and so litho will be addressed 100% computationally using ultra-regular layouts, extreme RET ( incl. source mask optimization, ‘SMO’), and virtual fab-ing using predictive modeling.
To that end, the recently formed DFM Consortium (DFMC)
—founding members including Cadence, Freescale, IBM, Samsung, ST, and TI, and—announced new members including Infineon, Intel, Mentor Graphics, and UMC
. Let us hope that the DFMC now has sufficient leverage to develop new standard models and metrics to allow for innovation to be quantified and rapidly implemented into design flows.
Labels: CAA, DFM, gridded design rules, modeling, restricted design rules, virtual fab
posted by [email protected]
080304: DFM matures along with industry
Ed’s Threads 070413Musings by Ed Korczynski on April 13, 2007MRS meeting specs the future
The Materials Research Society (MRS) spring meeting was in San Francisco April 9-13
, and the near- and far-term possibilities for process technology in our industry were presented to a record numbers of attendees. Researchers showed results from the world’s leading labs for electronic materials development: CMOS high-k gate dielectrics, nano-imprint lithography, organic semiconductors, quantum dots, and nano-tubes. It’s like sipping from a firehose, unless you’re interested in just one of the 36 parallel sessions.
Sachin Joshi of UT-Austin showed that hybrid-orientation technology (HOT) silicon wafers
based on the MEMC direct silicon bonding (DSB) approach contain inherent defect-rich junctions between orientations. Shallow-trench isolation (STI) regions 60-140nm wide may be used to eliminate these defects, though this seriously limits circuit density, he pointed out. Non-silicon channels will probably also limit density, so their use will probably be limited to RF and mixed-signal applications in small portions of chips.
Arief Budiman from Stanford analyzed the grain orientation in submicron damascene copper lines using the synchrotron x-ray beam from the ALS Berkeley Lab
. This very bright x-ray source and submicron spot-size (0.8 x 0.8µm) allows for resolution of crystal bending/stress as well as the dislocation density. Starting with large single grains spanning across the width of the line (“bamboo structure”), Budiman’s group observed clear directionality of EM-induced plasticity and thus the orientation of slip-planes. They found that <112> orientations were most susceptible to plastic deformation, so any grains with such orientations that line up with the induced EM-stress will deform. Grain orientation controls plasticity, which in turn influences EM degradation mechanisms and circuit reliability.
An analysis of the influence of microstructure on void formation in failed copper interconnects
, from Intel's Sadasivan Shankar, revealed that voids first nucleate at triple-boundaries caused by stress-induced de-cohesion at copper interfaces. These voids can be easily pinned by a grain boundary, which provides a fast diffusion path for the void to grow across the width of a line or via. “It almost unzips the grain-boundary,” he commented. A 2D model developed with Brown and UT-Austin accounts for current flow and stress, diffusion along surfaces and interfaces, void migration, and the interaction of voids and grain boundaries.Duane Boning
, the MIT professor who created one of the first useful pattern-density step-height CMP models in the 1990s, showed progress on new physically based models. By explicitly including pad properties—elastic response (including lateral coupling across the pad), slurry transport, and average asperities—he showed how chip-scale uniformity can now be predicted.
Roland Rzehak of Qimonda in Dresden, Germany, provided both an overview and details of inexplicable CMP removal-rate variations using ceria-slurries. A counter-intuitive “slow start phenomena
” slows the removal rate for the first minute of pattern planarization to be ~2.5X lower than that for blanket films. Ceria particles may initially adsorb in trenches to take some of the pressure load. However, Qimonda observes additional non-uniformities implying influences of pattern pitch, the pad material, and possibly effects from chemical additives to the slurry.
MRS meetings also cover wilder technologies like superconductors, neuro-prosthetic interfaces, and “the nature of design using nature’s portfolio” like the self-assembly of sea-shells
or the nano-hairs of gecko feet
. Materials scientists and engineers continue to explore the structure-property relationships of the physical world, and confirm that there is indeed “still plenty of room at the bottom
Labels: electronics, materials research, modeling, nano, self-assembly, silicon
posted by [email protected]
070413: MRS meeting specs the future