28 nm IC Design: The Devil Is In The Details

Nora — Mon, 14 Mar 2011 23:00:00 GMT

Smaller process technologies are enticing chip makers with bigger rewards from their end products. The shorter gate lengths at 28nm promise faster transistor speeds and less leakage power, and can double the amount of the logic that can be put into the same die area. Most importantly, however, more die on a wafer means lower per unit cost.

Meanwhile, design complexity is compounded as chip content grows significantly larger. Chip sizes are significantly smaller, but are running faster into the gigahertz range, and they must consume less power. When designers break barriers to pursue new innovations at advanced nodes, design flaws are common and more costly.

A recent highly publicized glitch in the Intel's chipset captured attention from the semiconductor industry, EDA community, and across businesses and consumers. This is not only because Intel's circuitry powers electronic notebooks and laptops which are ubiquitous in our daily life, but it is also because the cost of the flaw is notable. This news serves as another reminder that similar incidents have happened before, and can happen again.

Details are Challenging

How can we prevent a costly design flaw from happening again? The common wisdom tells us that the devil is in the details. Oftentimes, the last 10% of the task is taking care of few neglected design details. Such tasks can take many iterations to resolve, drain more engineer resources, and extend project schedules. From a digital implementation context, design detail refinements usually relate to the tradeoff between meeting performance and power targets while trying to be compliant with the complex physical, electrical, and process rules and guidelines. For example:

Large and complex design rules have a heavy impact on cell architecture as well as interconnect routing topology and via configuration. Cross coupling and resistance in interconnects and vias are a lot higher in 28nm node than previous generations.

Complex via rules

Complex interconnect rules

If stress and litho effect around cell boundary is not accounted for, timing slack and leakage power problems can be more severe. This is due to intra-cell dependence.

Current density per unit area and electromigration are big concerns.
Metal thickness variation has a bigger impact on parasitic resistance and capacitance. Metal thickness modeling needs to consider the chemical mechanical polishing (CMP) effect precisely for density analysis.

Thickness variation

Smaller wires are extremely sensitive to random defects, lithography distortion, and other types of variation in manufacturing process. Such litho effects may create open and short nets after GDSII is created and cause catastrophic problems. If such problems are not prevented and detected early, the design may pass the traditional LVS/DRC verification, but still fail to operate properly in the end.

The designs at 28nm usually have more aggressive objectives for timing, power and higher density. It is very difficult and sometimes impossible to make corrections at a later stage without a redesign. If the above details are not addressed properly and timely during implementation, they can lead to a functional flaw in silicon, or become the designer's nightmare to fix before tape out. Some predict that more than half of the sub-28nm designs are likely to face re-spins one way or another due to these oversights.

Taking Care of Design Details Through a Unified Design Flow

At the 28nm and below process nodes, having a set of best in class point tools provides a good foundation for design implementation success, but not enough to prevent error oversight. It also requires a consistent method to enforce a tighter design discipline from the start. Encounter Digital Implementation System (EDI System) provides a unified constraint driven methodology with an integrated concurrent signoff analysis flow. It systematically evaluates design constraints, and executes design intents and guidelines down to the smallest implementation details across partition boundaries and design hierarchies.

For example:

For placement, its context-aware placement automatically takes intra cell interdependence into account to ensure timing and leakage power objectives are met.
For routing, NanoRoute recognizes lithography unfriendly patterns, takes a "correct-by-construction" approach to route, preventing hotspots in wire configuration on the fly. The last few hotspots left in the circuit are detected by the in-design Lithography Physical Analyzer for timely repair.

Litho unfriendly pattern

For metal density, its smart metal fill is guided by foundry certified CMP model. Thus the thickness variation is factored into the equation when deciding the metal fill shapes and locations.

Many designers from the semiconductor industry, IP providers, and design service companies have already found benefits from this disciplined and unified design flow, and reached design silicon success faster without costly incidents.

To continue to satisfy the design community's insatiable appetite for new innovation, Cadence recently charted a unified path to 28nm silicon realization with the new EDI System 10.1 release. It provides a Silicon Realization flow based on unified design intent, abstraction, and convergence with physical and manufacturing data. A previous blog posting shows how it tackles chip design problems with a unified digital flow.

"Great things are not done by impulse, but by a series of small things brought together" - Vincent Van Gogh

Nora Chu

EDA Industry Stays Ahead of Technology Curve

Nora — Tue, 05 May 2009 12:00:00 GMT

The EDA Industry is the unsung hero behind for modern era electronic revolution since early 80s and gets the spotlight it deserves in the recent DAC newsletter.

I would like to applaud the author Geoffrey James, for crediting the EDA industry in rising to the challenges associated with each and every technology process node, in particular advancing the use of multi-core architectures. The EDA industry continues to help the semiconductor sector by solving some of their core business issues, such as chip performance, reliability, power consumption, manufacturability and productivity.

The demand for higher clock frequency and less power/heat dissipation is the main driving force for multi-core configuration. As an enabler, the EDA industry must stay ahead of the curve. We must invigorate parallelism, multi-tasking capability within applications while delivering an array of automation tools for design implementation, analysis and validation. We must also continuously focus on ease-of-use to enable better and faster design of multi-core systems, and enable time-to-market saving for our users. “The EDA industry has always depended upon the most-recent generation of chips to provide the power needed to crunch the data required to build the next generation” noted by Tom Spyrou, Cadence Distinguished Engineer and Encounter Platform RnD Manager, who is featured in this interview.

This challenge is immersive, and the Encounter Digital Implementation System is a great example. One of the obvious challenges is transforming enormous number of lines of code, that were not written to be parallel, to be multi-threaded, distributed or both. The solution must also be optimized for scalable performance and adhere to our customers’ unique computing environments and needs. Another challenge, which is equally important, is to speed up any serial computation steps left in the flow, else the performance gains from multi-threading in a RTL to Signoff flow could be significantly limited, according to Amdahl’s law.

With its unified client-server multi-cpu backplane, the Encounter platform has supported multi-processing since 1999, and has been continuously optimized for performance over time. Encounter’s focus is on overall flow performance from end-to-end, and we often achieve an average 2X- 4X runtime/memory footprint improvement per each major release. The latest release of the Encounter platform, called the Encounter Digital Implementation System is multi-cpu enabled by default.

In addition, to enable efficiency throughout the complete design flow, adjacent products such as sign-off analysis, DFM validation, the Virtuoso platform, functional verification and simulation are all multi-cpu enabled. It is Cadence’s goal to stay at the forefront in this, offering fully integrated and comprehensive parallel processing throughout back-end implementation. No matter what new challenges may come down the road, the EDA industry must continue to help the semiconductor industry overcome obstacles and deliver new, innovative products despite the economic slow down.

To learn more about the state of the art in multi-process Encounter applications, join us at the Parallelism tutorial at the 46th Design Automation Conference – DAC this summer, featuring the Distinguished Engineer Tom Syprou of Cadence, together with Professor Kurt Keutzer of UC Berkeley, Tim Matsson and Michael Wrinn of Intel.

Nora Chu

Nora Chu Blog

28 nm IC Design: The Devil Is In The Details

EDA Industry Stays Ahead of Technology Curve