SMT webinar banner3

A Brief History of Semiconductors: the Foundry Transition

A Brief History of Semiconductors: the Foundry Transition
by Paul McLellan on 09-04-2012 at 11:30 am

A modern fab can cost as much as $10B dollars. That’s billion with a B. Since it has a lifetime of perhaps 5 years, owning a fab costs around $50 per second and that’s before you buy any silicon or chemicals or design any chips. Obviously anyone owning a fab had better be planning on making and selling a lot of chips if they are going to make any money. A modern fab manufactures over 50,000 dinner-plate sized wafers every month.

In the past, fabs were cheaper. As a result most semiconductor companies owned their own fabs. In fact around 1980 there were no semiconductor companies that didn’t own their own fabs since there would be no way for them to manufacture their designs.

The first thing that happened was that some companies found they had excess capacity in their own fabs because an economically large fab might turn out to be larger than their own needs for their own product lines. Correspondingly, other companies may have the opposite problem: they didn’t build a big enough fab or they were late constructing it, and they could sell more product than they could manufacture.

So semiconductor companies would buy and sell wafers from each other to even out their capacity needs. This was known as foundry business, analogous to a steel foundry. In a similar way, semiconductor companies with shortages would take their designs to other semiconductor companies with surplus capacity (often even competitors) and have them manufactured for them.

The next step in the evolution of the ecosystem was that in the mid-1980s some companies realized that they didn’t need to own a fab to have chips manufactured. These companies would purchase foundry wafers just like any other semiconductor company. These companies came to be called, for obvious reasons, fabless semiconductor companies. Two of the earliest were Chips and Technologies, who made graphics chips for the PCs of the day and Xilinx who made what are now known as field-programmable gate-arrays (FPGAs). They purchased wafers from other semiconductor companies and sold them just as if they’d manufactured them themselves. Chips and Technologies eventually was acquired after falling on hard times, but Xilinx is still the leader in FPGAs today. And despite being number one, it still doesn’t have its own fab, it outsources all manufacturing.

Semiconductor companies with fabs became known as integrated device manufacturers, or IDMs, to distinguish them from the fabless companies. In 1987 the first of another new breed of semiconductor companies was created with the founding of Taiwan Semiconductor Manufacturing Company (TSMC). TSMC was the first foundry, created only to do foundry business for other companies who needed to purchase wafers either because they were fabless or because they were capacity limited. It was also known, when the distinction was important, as a pure-play foundry to distinguish it from IDMs selling excess capacity who would often be competing at the component level with their foundry customers.

Until TSMC and its competitors came into existence, getting a semiconductor company off the ground was difficult and expensive. To build an IDM required an expensive fab. To build a fabless semiconductor company required a complicated negotiation for excess foundry capacity at a friendly IDM which might go away if the IDM switched from surplus to shortage as its business changed. Once TSMC existed, buying wafers was no longer a strategic partnership, you just gave TSMC an order.

This lowered the cost and the risk of creating a semiconductor company and during the 1990s, many fabless semiconductor companies were funded by Silicon Valley venture capitalists. Historically, a semiconductor company had to be large since it had to have enough business to fill its fab. Now a semiconductor company could have just a single product, buy wafers or finished parts from TSMC and sell them.

Over time, another change happened. Many system companies also switched from using the ASIC companies to doing their designs independently and then buying wafers from the foundries. The specialized knowledge about how to design integrated circuits that was lacking in the system companies in the 1980s was gradually acquired and by the 1990s many system companies had very large integrated circuit design teams. The ASIC companies gradually started selling more and more of their own products until they became, in effect, IDMs.

As fabs got more expensive, another change happened. IDMs such as Texas Instruments and AMD that had always had their own fabs found they could no longer afford them. Instead some switched to being completely fabless. For example, AMD sold its fabs to an investment consortium that turned it into a foundry called Global Foundries. Alternatively they kept their own fabs for some of their capacity and purchased additional capacity, typically in the most advanced processes, externally. This was known as fab-lite.

This is the landscape today. There are a few IDMs such as Intel who build almost all of their own chips in their own fabs. There are foundries such as TSMC and Global Foundries who build none of their own chips, they just build wafers for other companies. Then there are fabless semiconductor companies such as Xilinx and Qualcomm along with their fab-lite brethren such as Texas Instruments, who do their own design, sell their own products, but use foundries for all or part of their manufacturing.

A Brief History of Semiconductors
A Brief History of ASICs
A Brief History of Programmable Devices
A Brief History of the Fabless Semiconductor Industry
A Brief History of TSMC
A Brief History of EDA
A Brief History of Semiconductor IP
A Brief History of SoCs


3D Memories

3D Memories
by Paul McLellan on 09-02-2012 at 4:42 pm

At DesignCon earlier this year, Tim Hollis of Micron gave an interesting presentation on 3D memories. For sure the first applications of true 3D chips are going to be stacks of memory die and memory on logic. The gains from high bandwidth access to the memory and the physically closer distance from memory to processor are huge.

Micron have a Hybrid Memory Cube (HMC). Bandwidth to the memory is 128GB/s (1 terabit per second). Each die in the stack is divided up into partitions, each of which is a 2-bank autonomous memory element. In turn, these are grouped in the vertical direction into “vaults” that consist of the same partition on each die all linked up together using through-silicon-vias (TSVs). In turn, the memory die are stacked on top of a single logic die to form the whole cube, with fully integrated DRAM and logic.

This way of dividing up the cube is also used to divide up the simulation. Initially partition level simulations are used to capture general noise to timing sensitivity. Multi-partition simulations (same die) capture inter-partition noise coupling, and vault level (vertical slice) capture die to die noise coupling. Finally, cube level simulations capture the impact of system power delivery network (PDN), resonance etc. The fact that there is so much repeated in the architecture make this very hierarchical approach viable.

At the partition level, simulation focused on power-grid impedance analysis, voltage drop analysis, full spice-level simulation with the power grid and decoupling all modeled. Apache’s Totem static analysis was used to correct layout and to validate that the silicon and the model matched.


At the vault level (multi die) Apache’s chip-power-model (CPM) was used to create dummy layers and the simulation focused on ensuring that layers were adequately isolated from noise. Then at the level of the system PDN the focus was on noise and resonance and adequacy of decoupling.

Takeaways from modeling the hybrid memory cube:

  • complexity is a function of how tightly coupled different parts of the system are and how much concurrent activity is allowed
  • physically partitioning the device simplifies design, modeling and simulation
  • 3D challenges must be overcome in parallel.

The Micron presentation is here in Apache’s CPS microsite.


A Brief History of Cadence Design Systems

A Brief History of Cadence Design Systems
by Daniel Nenni on 09-01-2012 at 8:10 pm

EDA software for IC and system design became a commercial business in the early 1980s. In those days, 3 companies – Daisy Systems, Mentor Graphics, and Valid Logic Systems – dominated the emerging EDA industry. However, two small startups that emerged in the early 1980s grew rapidly and merged to form Cadence Design Systems in 1988.


One of those startups was ECAD, which was founded by Glen Antle and Paul Hwang in 1982. Unlike the leading EDA vendors, who also sold workstations, ECAD provided only software. The company developed and sold Dracula, an IC layout verification product that came into widespread industry use.

The other startup, SDA Systems, was founded in 1983 by Jim Solomon, a former product manager at National Semiconductor. Also a software-only company, SDA offered IC physical design tools within an integrated “framework” that could also accommodate third-party tools. Joe Costello, a former R&D manager at National Semiconductor, joined SDA Systems in 1984 and became its president in 1987.

In Feb. 1988 ECAD bought SDA for $72 million in a stock swap. A new company, Cadence Design Systems, was incorporated June 1, with Costello as president and CEO. Within a year Cadence had become the leading provider of IC design automation tools. Under Costello’s charismatic leadership over the next 9 years, Cadence completed a number of strategic acquisitions, built widely-used product lines, moved into system-level and PCB design, and made design services a key part of its strategy.

1989 was a formative year in several respects. First, Cadence bought Gateway Design Automation, developer of the Verilog hardware description language. (A year later Cadence put Verilog into the public domain, and it became the most widely used hardware description language). Secondly, Jim Solomon started the Cadence Analog Division, launching an effort in which Cadence became the undisputed leader in custom/analog design automation tools. Third, the purchase of Tangent Systems boosted Cadence’s leadership in IC layout automation.

Here are some key Cadence milestones since 1989:
· 1991 – Acquired Valid Logic Systems and became the EDA revenue leader.
· 1993 – Bought Comdisco Systems, a pioneer of system-level design.
· 1994 – Launched Spectrum Services consulting group.
· 1997 – Acquired HLD Systems (design planning) and Cooper & Chyan Technologies (automatic routing). Jack Harding, Cooper & Chyan CEO, became Cadence CEO.
· 1999 – Harding succeeded by Ray Bingham as CEO. Cadence acquired OrCAD (PCB design) and Quickturn (emulation).
· 2001 – Avanti found guilty of stealing Cadence trade secrets, ordered to pay $194 million restitution.
· 2002 – Cadence Genesis design database offered to EDA industry as “OpenAccess.” Cadence bought Simplex Solutions (IC physical verification).
· 2004 – Mike Fister became president and CEO.
· 2005 – Cadence acquired Verisity, pioneer of coverage-driven verification.
· 2008 – Lip-Bu Tan became president and CEO.
· 2010 – Cadence EDA360 vision proposed expanded role for EDA. Purchase of Denali Software strengthened Cadence position in design and verification IP.

Fast forward to 2012, and Cadence has enjoyed three years of steady growth in revenues and earnings. 30 years after the launch of ECAD, Cadence employs approximately 4,900 people worldwide, reported 2011 revenues of $1.1 billion, and provides a wide range of IC, PCB and system design products that are used by nearly every semiconductor and electronic systems provider today.

A special thanks to Richard Goering, Senior Manager, Technical Communications, at Cadence, for compiling this data. Richard has covered EDA since 1985 most notably as EE Times’ EDA editor for 17 years, where he wrote hundreds of articles for both the print and on-line editions of the electronics industry’s premier weekly newspaper.

A Brief History of Semiconductors


The Need for OASIS in Post-layout IC Databases

The Need for OASIS in Post-layout IC Databases
by Daniel Payne on 08-31-2012 at 7:20 pm

OASIS is a hierarchical IC file format used for IC designs that is gradually replacing GDS II throughout the mask data stages. The compelling reason for using OASIS has always been the reduction of file size, and speed up of processing times through the use of hierarchy and fewer translation steps.

At the 45nm node an actual M1 layer from design had a file size of 0.066 Gb, however after going through pre-OPC, post-OPC and Fracture stages that same M1 layer has ballooned in file size to 59 Gb, according to a White Paper authored by four Mentor specialists (Deployment of OASIS in the Semiconductor Industry – Status, Dependencies and Outlook). The following figure shows in blue how this file size expansion happens at each stage:


File size and hierarchical content across the 4 different stages of mask data processing for a real 45nm M1 example. The “hierarchy” data is computed as the number of flat geometries divided by the number of hierarchical geometries. These results show that the post-OPC step has the highest return from an investment in hierarchical compression technology such as the OASIS format.

EDA companies rapidly adopted OASIS as a file format however it took a few years longer for the Fracture data stage to use OASIS based on survey results:


Survey results showing OASIS adoption by technology node, broken down by the data-prep handoffs. These results indicate that OASIS has been adopted by a significant fraction of companies doingOPC as early as 65nm, but that hand-off of fractured data in OASIS has lagged by as much as 2 processnodes.

Just how much compression should you expect when changing from GDS II to OASIS? The average compression is 18.57 at the design stage, and 14.24 at the post-OPC stage:


Data size savings of OASIS vs GDSII, as measured by compression ratio of the file sizes, fordesign and post-OPC layout files. Although the median compression ratios are very similar, the entiredistribution of compression rations for post-OPC files consistently shows smaller values (shifted to theleft) than the distribution for design files.

One factor that slowed adoption of OASIS from GDS II has been the low costs combined with expanding capacity of disk storage.

Summary
Work to define OASIS started back in 2001, became a SEMI standard in 2005, and was quickly adopted by EDA vendors. The post tape-out vendors are using OASIS in Pre-OPC, Post-OPC and Fracture stages to reduce runtimes by 3X and reduce file storage sizes by 3X. Old file standards like GDS II have given way to OASIS gradually over the past decade.


Toshiba Corp. has cancelled plans to sell (ARM-based processor) computers …bad news for Toshiba, TI and the end user

Toshiba Corp. has cancelled plans to sell (ARM-based processor) computers …bad news for Toshiba, TI and the end user
by Eric Esteve on 08-31-2012 at 11:55 am

This news announced two weeks ago is not really good news for those expecting to see this new generation of computers, running Windows RT OS and no more based on x86 processor but on ARM CPU core, coming on the market. The reason invoked by Toshiba was “delays in getting adequate supplies of components” and, even if Toshiba did not specifically mention the guilty supplier, it appears that the computer was based on OMAP platform from Texas Instruments, the chip being manufactured on TSMC 28nm technology. Let’s make it clear from the beginning, this blog is not aiming at putting shame on a specific and respected chip maker like TI, neither on the largest and well respected Silicon foundry (TSMC), but rather to bring my point of view about the probably needed evolution in the Foundry/Fabless/Fab-lite equation.

At one extremity of the SC landscape, we find Intel, still undisputed leader with $54 billion revenue in 2011, being the perfect Integrated Device Manufacturer (IDM) example. That means that Intel manufacture in the company wafer fab the IC designed by Intel engineers and, by the way, using the CPU core 100% owned by the company, the x86 family. At the other extremity, we find the many “fabless” companies, which can be small start-up –in this case the fabless business model is the only way to start a SC company, due to the huge cost ($5 billion) and long delays (3 years minimum) you need to build a fab to manufacture in advanced technology nodes like 28 or 22nm- but also companies reaching revenue level of several $ billion, a well known example being Qualcomm (the company has never built any fab), but also AMD (CEO and founder Jerry sanders use to say in the 70’s that “real men have fabs”) since the company has sold the fabs to a consortium creating GlobalFoundries. In this case, the picture is crystal clear, you own a wafer fab to manufacture your chips or not.

But, since the late 1990’s, the “fab-lite” concept has appeared, mainly due to the already high cost associated with the development of new technology and wafer fab equipment, very capital hungry; mainly, but not only, as the other reason is an economic theory which consider that a company should focus on core business. This theory works well for companies who have so much diversified their products that they can’t invest in every product line enough cash to be part of the leaders, this lack of leadership leading to poor capital ROI or even bankrupt. But that we are seeing these days, the Qualcomm, TI or Nvidia being unable to supply IC in 28nm at the level demanded by their customers, is showing the limits of this “MBA type” of thinking.

What can be worse than not being able to supply your flagship product (this one developed in the most aggressive technology node), offering you the best gross Profit Margin (GPM) and to your customers the best supported feature list at the lowest price and lower power consumption?

Because we are working in a fast moving industry, the first reaction have raised quickly. One option is proposed by Mike Noonen, Executive Vice President, Worldwide Marketing and Sales, for GLOBALFOUNDRIES:

“There are warning signs, both technical and economic, emerging in the foundry business that warrant our attention, and in fact require a re-thinking of how best to apply our resources and energy. The solution, ironically, may be a move toward a more IDM-like model. Strategic collaboration that creates a ‘virtual IDM-like interface’ to chip design companies will help further close the gap between process teams at the manufacturing companies and design teams at the fabless companies. Collaboration – early, often and deep – is really the only practical approach given the cost and complexities involved.”

Mike Noonen is Executive Vice President, Worldwide Marketing and Sales, for GLOBALFOUNDRIES. In this role, he is responsible for global customer relationships as well as all marketing, sales, customer engineering and quality functions.You can find his blog on GFI or ARMweb site.

This is not the only option, and I consider we should also think to another solution, which is for a fab-lite company to carefully select which product to fab outside. This remind me a meeting I had in 2005 in Shanghai with one of the SMIC Director, who share with me how this pretty young foundry did to grow so fast. At that time, SMIC business model was to build fabs, using their customer money, and manufacture the mature products for the TI and Motorola, so probably on 250 or 180nm at that time. We could imagine that the IDM going Fab-lite decide that their core business should include wafer fab, as far as these will be used to manufacture their flagship products, most often in the most advanced technology nodes. This should not prevent them to work with foundries, either for mature products, either to add extra capacity on these flagship products when the market demand is passing a certain level, or even to get better cost from foundries…

The risk is to stay in current situation, and not taking any of these decisions. Risk for these pure fabless or fab-lite IDM to miss the market window on high runner SoC for smartphone or media tablet, and get weaker, and risk also for us, the consumer, the be locked to Intel processor based only solution, when we see the opportunity to buy ARM based media tablet running on Windows RT operating system vanishing, because of wafer fab capacity issue.

Eric Esteve from IPNEST


Arteris joins Inc. 500 List of America’s Fastest-Growing Private Companies… thanks to Arteris customers!

Arteris joins Inc. 500 List of America’s Fastest-Growing Private Companies… thanks to Arteris customers!
by Eric Esteve on 08-31-2012 at 4:35 am

Arteris, founded in 2003, is the inventor and leading supplier of network-on-chip (NoC) interconnect IP solutions. Can we say that the company is still a start-up? I would say yes, as their flagship product, FlexNoc (Network on Chip IP function) was a completely new concept when it was introduced. As for every disruptive technology, it takes time for the new concept to be, at first understood by the potential customers, then tested and evaluated, and finally sold. I would say yes as well when looking at the 2011 to 2010 year to year growth rate of 115%. Sales of Arteris’ unique FlexNoC IP product have driven the growth of the company as Arteris’ patented NoC technology displaces older interconnect technologies such as hybrid busses, tiered switches and crossbars.

So, the Arteris start-up has joined the Inc. 500 list of America’s fastest growing private companies, and this is good news for the SC industry: it means that it is still possible to start an IP company, based on real innovation, and finally be successful. As a side remark for our venture capital friends, please note that it took some time (9 years), and that the company is still in pre IPO stage… it takes some time to cook a good recipe, but the result is far better than fast food (at least according with my French taste).

We have talked in Semiwiki about Arteris product port-folio, including FlexNoc, MIPI LLI (Low Latency Interface) and C2C (Chip-To-Chip Link), so it could be a good idea to look at what has made the success of the company, I mean Arteris’ customers. When you say NoC, you think SoC, and you almost immediately link SoC to Application Processor for smartphone and media tablet, which is true, but we will see that Arteris’ customers develop SoC targeting many other market segment: Video Processing, Consumer Electronics, Automotive, Networking, Multimedia (digital TV), even if a majority develop Application Processor SoC. Let’s start with these:

As you can see, most of the leaders in Application Processor SoC have selected Arteris, and they are located in the USA, Korea and China. OK, Intel is missing on this list, but the success of Intel as a low power chip maker for Application Processor is still to be demonstrated…

Another market segment where Arteris FlexNoC seems to be very popular is Video and Imaging, and the company enjoys design-in with many companies, once again located in various places like USA, Japan, Korea or India:

Because the SoC developed for Consumer Electronics are probably the most complexes, integrating an IP count which can be above 100 (SoC for the Set-Top-Box, for example), developed on one technology nodes just behind the Application Processor, integrating a Network-on-Chip is an obvious solution to speed up the Time-To-Market, so we should not be surprised to see a long customer list in this segment:

This should not be a surprise to see some of the customers listed in the Application Processor segment to appear also in this list. This is not a way to count the same chip twice and increase the customer list, but reflects the fact that companies who have made the huge investment needed to develop an Application Processor are very happy to increase the ROI by targeting other segments, close in term of application (Video, Imaging…) so they can minimize the new investment by designing derivatives chips from the initial Application Processor, leading to a different implementation of the NoC IP. If you look for example at OMAP platform from TI, you will notice various chips named OMAP4XX, each of them targeting a specific segment from Automotive to Wireless and Media tablet (or will be positioned from the high range to the most economical solution in the same segment).

Clearly, this long list of (prestigious) customers has made the success of Arteris. But this success has been made possible because Arteris’ unique FlexNoC NoC interconnect IP product was filling a need in the SC industry. According to the March 23, 2012 Gartner Research report, “Market Share: Semiconductor Design Intellectual Property, Worldwide, 2011”, by Gartner analysts Ganesh Ramamoorthy and Masatsune Yamaji, Arteris charted a 115% increase in IP license revenue in from 2010 to 2011. “Infrastructure IP was the fastest growing semiconductor design IP market after microprocessor IP in the past year. Its growth is correlated with the growing demand for interconnect functionality in advanced system-on-chip designs,” said Ganesh Ramamoorthy, research director at Gartner. “The increasing interconnect requirements of SoC vendors is driving the demand for network on chip technology provided by IP companies.”

Let’s give the last word to Charles Janac, President and CEO of Arteris:
“We are very excited to be named to the Inc. 500 list,” said K. Charles Janac, President and CEO of Arteris. “Since our founding in 2003, Arteris has charted our own path by continually delivering unique network-on-chip interconnect IP technology to the market. The fact that our fast growth and rapid adoption by leading companies such as Samsung, Qualcomm and Texas Instruments has occurred during challenging economic times proves that our FlexNoC IP saves our customers money, and helps them earn more revenue from their SoCs. Our internal data, based on customer experience, shows a 10x return on investment (ROI) when using Arteris IP for advanced SoC designs. This ROI results from cost savings, schedule savings, and increased SoC revenue due to better SoC performance.”

Eric Esteve from IPNEST


Did Qualcomm and Apple Attempt to Bribe TSMC?

Did Qualcomm and Apple Attempt to Bribe TSMC?
by Daniel Nenni on 08-30-2012 at 7:45 pm

It’s amazing how these rumors start and go viral. People are literally laughing here in Taiwan. I remember I wrote something that I thought was obviously satire and less than a week later someone repeated it back to me as fact since they “read it on the internet somewhere.”

According to Bloomberg: Apple Inc. (AAPL) and Qualcomm Inc. (QCOM) were rebuffed in separate attempts to invest cash with Taiwan Semiconductor Manufacturing Co. (2330) in a bid to secure exclusive access to smartphone chips, people with knowledge of the matter said.

Both proposals included investments, each of more than $1 billion, for the world’s largest custom maker of chips to set aside production dedicated to making chips exclusively for them, said the people, who declined to be identified because the details are not public.

Sounds like a bribe to me. No wonder why they don’t want to be identified. Dozens of variations have spawned from this rumor so now it is an “internet fact”:

TSMC Spurns Apple, Qualcomm Bids for Guaranteed Fab Capacity
PC Magazine

Apple, Qualcomm failed to buy TSMC chip exclusive rights [report]
ZDNet
Apple, Qualcomm won’t get to hog TSMC chip fab capacity
Ars Technica

TSMC Doesn’t Want Apple, Qualcomm’s Money for VIP Chip Access
DailyTech

Money can’t buy everything: TSMC denied Apple exclusive mobile chips acces
GigaOM

Apple, Qualcomm tried to purchase exclusive access to TSMC chip production
tuaw.com

TSMC rejected billions from Apple and Qualcomm for exclusive access
Inquirer

Semiconductor Producer TSMC Turns Down Bids From Apple, Qualcomm
the Mac Observer

Even IEEE Spectrum got in on it: TSMC’s Morris Chang Says No to Apple, Qualcomm

Would TSMC build fabs for Qualcomm and Apple? Of course, if it was a responsible thing to do for the TSMC stakeholders. Samsung already does that for Apple as an example. Would TSMC accept money from a customer for preferential treatment? Of course not, and anybody who thinks otherwise does not know TSMC or how the fabless semiconductor ecosystem works. And I can assure you that Qualcomm and Apple know.

From the TSMC website:

Our mission is to be the trusted technology and capacity provider of the global logic IC industry for years to come.

Now lets look at the authors of the rumor: Tim Culpan, Ian King, Adam Santariano. Sorry, never heard of them and not an ounce of actual semiconductor experience between them. Just my opinion of course.



Smart mobile SoCs: Made in China

Smart mobile SoCs: Made in China
by Don Dingee on 08-29-2012 at 2:00 pm

One of the comments to previous installments of this series was that there isn’t much left for the merchant suppliers of smart mobile SoCs, considering Apple and Samsung have majority share and design their own parts. The theory is this makes it hard for many suppliers to continue investing at the resource levels needed to bring a complex SoC to market.

Unless, the market we’re talking about is China. Continue reading “Smart mobile SoCs: Made in China”


Mixed-Signal Methodology Guide

Mixed-Signal Methodology Guide
by Daniel Payne on 08-29-2012 at 11:14 am

Last week I reviewed Chapter 1 in the new book: Mixed-Signal Methodology Guide, and today I finish up my review of Chapters 2 through 11. You can read the entire book chapter by chapter, or just jump directly to the chapters most related to your design role or project needs. With multiple authors I was impressed with the wide range of AMS topics they were able to cover from theory to practice. The last sentence in this blog contains a free offer. Continue reading “Mixed-Signal Methodology Guide”


Mixed-Signal Methodology Guide

Mixed-Signal Methodology Guide
by Daniel Payne on 08-29-2012 at 11:14 am

Last week I reviewed Chapter 1 in the new book: Mixed-Signal Methodology Guide, and today I finish up my review of Chapters 2 through 11. You can read the entire book chapter by chapter, or just jump directly to the chapters most related to your design role or project needs. With multiple authors I was impressed with the wide range of AMS topics they were able to cover from theory to practice. The last sentence in this blog contains a free offer, and you can enter to win a free book here.

Chapter 2: Overview of Mixed-signal Design Methodologies
The AMS methodology you choose really depends on what you are designing, so there is no single flow that fits all design styles. Approaches can be top-down, bottom-up, or meet in the middle. There’s a continuum of AMS designs from Analog-centric to Digital-centric.

Historically there were two sets of EDA tools and databases for AMS design: Analog, and Digital. Today however there are common databases and tools for IC design that allow concurrent AMS design.

Based on the amount of Analog or Digital in your design you can choose the appropriate design methodology:

Chapter 3: AMS Behavioral Modeling
SPICE circuit simulation can be used on any transistor-level netlist to predict the analog behavior of blocks of your design however at the expense of long run times. If you were to write a behavioral model of your analog block then it could simulate orders of magnitude faster than SPICE. This chapter has plenty of examples on how to start writing AMS behavioral models:

  • Programmable Gain Amplifier (Verilog-AMS)
  • Analog PGA (Verilog-A)
  • Real PGA Model
  • Digital PGA Model
  • Operational Amplifier (Verilog-A)
  • Digital to Analog Converter (Verilog-AMS)
  • Low-pass filter (Verilog-AMS)

The concept of Real Number Modeling (RNM) is introduced as a method for analog voltages to be represented as a time-varying sequence of real values.

Chapter 4: Mixed-Signal Verification Methodology
At 72 pages this is the longest chapter in the book and reflects that verification consumes more development time and engineering effort than does AMS design, just like in Digital designs. Simulation is the primary tool used to verify an AMS design, unlike in digital where you also have formal methods and Static Timing Analysis.

Assertions are being used in AMS tool flows to capture design intent and report violations. The UVM-MS (Universal Verification Methodology – Mixed Signal) approach offers a direction for a metric-driven verification. Examples of using PSL (Property Specification Language) with Verilog-AMS show how analog events for assertion clocking can be coded.

SystemVerilog Assertions (SVA) are also demonstrated by a mixed-signal Sigma Delta ADC example. Mixed-signal coverage is defined along with mixed-signal metric-driven verification, so engineers can start to measure the effectiveness of their AMS verification.

The Common Power Format (CPF) syntax is used to show how power domains are modeled and the challenges of designing for low power in mixed-signal chips.

Chapter 5: A Practical Methodology for Verifying RF Chips
Jess Chen from Qualcomm wrote a math-filled chapter showing how to verify a direct conversion wireless OFDM link. His SystemVerilog code implements a low noise passband amplifier, IQ modulator, IQ demodulator and baseband amplifier.

Simulation waveforms called Pretzels show the behavior of the RF models:

Chapter 6: Event-Driven Time-Domain Behavioral Modeling of Phase-Locked Loops
Verilog-A and Verilog-AMS was used to model a PLL circuit by decomposing the design into several blocks:

  • Phase Detector
  • Charge Pump
  • Loop Filter
  • Voltage-Controlled Oscillator

Refinements are shown that add jitter and frequency slewing for a PLL. The benefit of the behavioral model is that you can simulate PLL lock time in seconds instead of using SPICE and having to wait up to a week for results.

Chapter 7: Verifying Digitally-Assisted Analog Designs
The concept of using digital circuits to tune an analog block is gaining use, and Art Schaldenbrand from Cadence provides several examples with calibration to show a new verification methodology:

 

  • VCO with calibration
  • Multi-Bit Delta-Sigma ADS with dynamic element matching
  • Active-RC filter

Chapter 8: Mixed Signal Physical Implementation Methodology
Two IC layout flows are highlighted: Custom Design, and Constraint-driven.

The constraint-driven flow uses Pcells (or Pycells if you prefer non-Cadence tools) to introduce more automation and allow for process migration of analog blocks. The idea is to quickly get from schematic to layout, then use the parasitics from layout back in simulation where performance can be measured and then transistors either re-sized or layout changed.

Physical verification is also touched on:

  • LVS (Layout versus schematic)
  • DRC (Design Rule Checking)
  • ERC (Electrical Rule Checking)
  • DFM (Design For Manufacturing)

Chapter 9: Electrically-Aware Design Methodologies for Advanced Process Nodes
Layout Dependent Effects (LDE) can dominate the performance of an AMS design at 28nm and lower process nodes, so using an electrically-aware design methodology can reduce iterations and minimize design time.

During IC layout the EAD flow will take into account several LDE effects that impact variability and performance:

  • Shallow Trench Isolation (STI)
  • Well Proximity Effects (WPE)
  • Length of Diffusion (LOD)

Reliability concerns like Electromigration (EM) can be analyzed while doing IC layout, instead of waiting for final block assembly.

Chapter 10: IC Package Co-Design for Mixed Signal System
Taranjit Kukal of Cadence wrote about System in Package (SiP) as a way to integrate multiple dies with discrete components into a single package. Our smart phones are probably the most successful example of SiP technology in use today. Multiple ways of implementing SiP were shown:

  • Single die in package
  • Multi Chip Module (MCM)
  • RF module
  • 2.5D IC
  • 3D IC
  • 3D Package

Co-design between package and SiP allows for trade-off analysis early in the system design process, optimized I/O locations, and Power Delivery Network (PDN) analysis.

Chapter 11: Data Management for Mixed-Signal Designs
Michael Henrie and Srinath Anantharaman from ClioSoft described how Design Management (DM) of ICs is different than Software Configuration Management (SCM) because of the broad spectrum of design data:

  • Specifications
  • HDL design files
  • Verification test benches
  • Timing and power analysis
  • Synthesis constraints
  • Place & Route
  • Parasitic Extraction
  • Standard Cells
  • Analog Design
  • PDK (Process Design Kits)
  • Custom Layout
  • GDS II
  • Packaging
  • Scripts & Customizations

An AMS design management system should enable:

  • Collaboration across team members
  • Version control
  • Release management
  • Variant development
  • Security and access control
  • Audit trail of changes
  • Integration with bug tracking
  • Design flow integration
  • Checkin / Checkout to control access to cells, blocks and modules
  • Quick response, low disk space use
  • Composite design objects
  • Shared workspaces
  • Visual change analysis
  • Hierarchical design
  • Reuse of PDKs

ClioSoft offers one of a few DM tools that work within the Cadence environment and operates on the concept of a repository:

A DM methodology can enable an AMS team to collaborate efficiently, avoid mistakes of data loss, automate version control, analysis changes visually, plus use IP and PDKs across the organization.

Summary
Even if you are a specialist in AMS design and verification you will benefit from the big picture presented in this new book: Mixed-Signal Methodology Guide. The eleven chapters cover a wide range of relevant topics, plus there are ample references to allow you to further explore a topic. I enjoyed the numerous examples provided and code snippets as a way to learn.

As a reward for those who read this blog to the very end I am offering a free copy of the book to the first two people who post a comment and request their copy.

Also Read

Book Review: Mixed-Signal Methodology guide

Interview with Brien Anderson, CAD Engineer

Managing Differences with Schematic-based IC design