A major acceleration in growth of the FPGA market should occur in 2010 based on several factors listed below. The two companies that control approximately 85% of the FPGA (Field Programmable Gate Array) market are positioned to exploit this inflection point in growth.
1. When 65nm devices go mainstream
a.
FPGA's
as SoC (for embedded systems), especially with ARM and MIPS soft-core
processors
b.
The
continued acceleration of FPGA's stealing market share from ASICs and ASSPs
2. FPGA's in new markets
a.
x86
servers and PCs
b.
Industrial
& Automotive
c.
Consumer
devices
3. Beginning of construction for the next generation communication network driven by
a.
The
communication equipment is nearly half of the FPGA end market
b.
The
glut of dark fiber from the .com bubble will be in use
c.
The
explosion of true mobility due to 802.11n, WiMax, LTE, etc
d.
The
explosion of interactive TV (through high definition video over IP) enabled by
fiber to the home
e.
The
virtualization of the data center, including desktops
4. The end of the global slowdown, re-invigorating world-wide growth of capital goods products, which are increasingly designed with digital devices like FPGAs
Below we will explain how the structure of the semiconductor market is going vertical due to the dramatic increase in semiconductor complexity requiring semiconductor design integration and how this favors the two leading FPGA companies. There is an acceleration in designers using Xilinx and Altera’s platforms for next generation semiconductor products.
You can also view this document as a PDF. If you have a question about a semiconductor term used in this paper please checkout our Semiconductor Overview and Glossary.
1.
The EDA (Electronic Design
Automation) Market
2.
The ASIC (Application Specific Integrated Circuit) Market
3.
The Programmable
Logic Market
5.
Risks for Xilinx and Altera
6.
Convergence of
Semiconductor Solutions, Opportunity for FPGAs
7.
Conclusion
The Electronic Design Automation (EDA) is a four billion dollar market that enables the huge semiconductor industry. EDA companies offer software tools for designing and producing electronic systems such as integrated circuits. As integrated circuit (IC) complexity grows exponentially due to sub-90 nm manufacturing along with the move to system on a chip (SoC) and mixed-signal chips, software will become more and more the breeding ground for IC designers. However, the EDA vendors will not be the recipients of this sharpened focus on software.
According to former EDA industry analyst Gary Smith "The EDA has been holding the cost of IC design relatively constant for many years, in the $10 to $20 million range, despite steep rises in complexity... but the cost of design is up because of the cost of embedded software." Software is now the bottleneck in semiconductor product design.
In the 1980’s and early 1990’s EDA was a hot market, but due to poor leadership and market forces the industry has hardly moved over the past decade. I don’t expect this to change and this has ramifications in the semiconductor industry.
The EDA industry is an oligopoly dominated by three firms; Cadence, Synopsys, and the much smaller Mentor Graphics. These firms have historically outsourced their R&D by purchasing smaller EDA design firms. This has produced un-attractive returns over the past decade. So these firms are attempt to in-source innovation and R&D. This, of course, is easier said than done. If creating an innovative culture was simple these firms would have done it a decade ago.
So, the EDA venture firms have been shuttered and there is little venture capital money supporting new EDA firms. There have been no IPO's since 2001. The market leaders’ stock prices have been stagnant the past ten years and currently there are zero industry analysts focused on the EDA market. So where is the innovation occurring?
Each EDA firm offers a set of fragmented tools used that designers can use. Even if a designer purchases software from a single EDA firm they will get multiple tools that rarely interface efficiently. As more features are added to chips the amount of design software also increases. As integrated circuit device complexity increases designers are looking for a greater level of integration from their design tools.
The two largest EDA firms Cadence and Synopsys are in the process of integrating their various development tools, many obtained through acquisition, into integrated software development platforms. As IC designers’ focus more on embedded systems they are requiring an end-to-end design platform. However Synopsys and Cadence tools are still fragmented because the tools were independently designed and built; often from different firms with dramatically different software architectures. The time consuming process to re-design and integrate these tools are costing these firms valuable time in the competitive, fast moving semiconductor market.
The move to a platform strategy is the correct approach for Synopsys and Cadence, but it does not come without significant execution risks. Their largest customers are still building Application Specific Integrated Circuit (ASIC) chips from the ground up and these vendors may look to move to more nimble tool based providers like Mentor Graphics and Magma Designs.
Very high performance ASIC's may require a tool approach for the foreseeable future due to the rapid pace of innovation. Balancing the needs of the many smaller firms designing semiconductor devices with the largest customers, like Intel and IBM, is becoming more difficult as the divide between these two groups’ requirements grows. As ASIC complexity has increased semiconductor design has increasingly moved to a software focus and EDA vendors have struggled to keep up with the latest production processes.
The move to more complicated embedded systems enabled by smaller manufacturing processes may be an inflection point for most designers to move to integrated design platforms and also move to a higher level of abstraction offered by programmable logic devices (PLDs).
While a complete ASIC flow includes coordinating with multiple suppliers of services, silicon, tools, and IP using a PLD platform is a one-stop shopping experience.
By vertically integrating the chip’s hardware design and offering a software design platform to the end product designer that includes framework IP (intellectual property), the PLD vendors have moved the level of abstraction up a layer; increasing the efficiency of designing digital products.
While the two largest EDA vendors are each struggling to integrate their disparate offerings into a unified platform two companies already offer an integrated development environment. These two companies, Xilinx and Altera, dominate the re-programmable logic device (PLD) and structured ASIC market. They have and will continue to pilfer design wins from the EDA vendors and ASIC vendors.
It takes long term strategic planning to build a platform strategy making the switch difficult for the EDA vendor’s with their antiquated tool strategy. The PLD vendors have historically had a platform strategy. Migrating to a completely new strategy has inherent execution risks. While the two major EDA firms are focusing their resources on strategy re-direction the PLD vendors are focused on building huge IP (Intellectual Property) libraries for their well tested, well understood, integrated platforms.
In addition, the EDA vendors have struggled to gain a foothold in the FPGA industry as FPGA customers use free or heavily subsidized development tools from the FPGA vendors.
Because analog chip designs have longer life cycles and PLD’s are geared for digital and mixed signal products Cadence, the analog chip design leader, has more time to execute on its platform strategy. Although, even with analog products it is debatable who has a better offering to build a platform. Is it Texas Instruments, Analog Devices, or Cadence?
Programmable logic vendors are far more profitability than the EDA companies. As a result over the past decade R&D in programmable logic has increased approximately four fold while R&D in EDA has merely doubled. The digital semiconductor device market is consolidating around a handful of companies, such as INTC, TXN, NVDA, BRCM, ALTR, XLNX, and MCHP. With fewer customers the EDA vendors will have an even more difficult time maintaining profitability.
Semiconductor design R&D is moving higher up the value chain; on top of specific semiconductor devices such as Xilinx’s and Altera’s PLDs; or on slower processor architectures from Microchip’s Microcontroller units (MCUs), TI’s digital signal processors (DSPs), or Intel’s x86 processors. Processor speed is increasingly a less relevant attribute when making a purchasing decision due to the deceleration in the increase in speed for each new processor generation. With the deceleration of processor speeds in recent processor generations and ASICs design complexity growing beyond the reach of most companies PLDs sit in a sweet spot in the semiconductor device market.
The big three account for 70 - 75% of total EDA revenue with sales relatively predictable because customers typically buy three year software licenses.
|
Company |
Symbol |
Market
Cap |
Sales |
Comments |
|
Cadence Design |
CDNS |
$3.1B |
$1.6B |
90% market share in custom analog EDA |
|
Synopsys |
SNPS |
$3.4B |
$1.2B |
digital ASIC leader, Intel is 10% of
revenues |
|
Mentor Graphics |
MENT |
$0.8M |
$0.8B |
Hasn’t moved to a platform strategy,
maintaining its tool strategy |
|
Magma Design |
LAVA |
$0.4B |
$0.2B |
|
Cadence currently trades for $11 a share (Apr 2008). It was $25 as recently as June 2007. It was $15/share in June 1997, with a 10 year low of $9/share in 2003. Synopsys currently trades for $23 (Apr 2008). It was $18/share in June 1997, with a 10 year low of $15/share in 2004.
Mentor Graphics currently trades for $9 (Apr 2008). It was $8 a share in June 1997; with a 10 year low of $5 a share in 2002. It gets 30% of sales from software and hardware emulation. Mentor has moved into new markets; now offering tools to design complex wiring harnesses and won a big contract with Ford. Magma Design Automation had its IPO in 2001 at $13/share (ended its first day of trading at $18/share). It currently trades for $10 a share (Jan 2008)
ASIC’s must be designed for a specific end product. These design and manufacturing costs have risen rapidly over the years as the geometry of semiconductors has gone from the millimeter to the nanometer level. The semiconductor manufacturing machines must be set for each ASIC, referred to as non-recurring engineering costs. Any bugs discovered after manufacturing an ASIC must be fixed with much slower software or the ASIC must be completely re-manufactured.
Since the complexity of designing an ASIC has increased dramatically many companies have outsourced ASIC design to third party firms which complicates IP ownership and potentially involves outsourcing what should be a core competency. As complexity, cost, and time to market has increased companies are looking to alternatives when building their digital products.
PLD’s are designed and manufactured once. Customers then write the software that these PLD’s will use. Generic or one time programmable processors are similar standardized IC that the customer uses to process the software it has written. By using a simpler PLD or generic processor it eases a company’s ability to in-source all of a product’s design while shrinking the time spent on product design and testing.
|
Characteristics |
ASIC |
ASSP |
PLD |
Custom Processor / DSP |
Generic Processor |
|
Customer’s
Development Cost |
Very High |
Low |
Moderate |
High |
Moderate |
|
Average
Selling Prices (ASP) |
Low |
Low to
Moderate |
Moderate
to High |
Low to
Moderate |
Low to
Moderate |
|
Time
to Market |
Slow |
Moderate |
Fast |
Slow |
Fast |
|
Customizable |
Yes (by
manufacturer) |
No |
Yes (by
customer) |
No |
Yes, only
one time (by
customer) |
|
Updatable
in the field (smooth
product updates) |
No |
No |
Yes |
No |
No |
|
Performance |
Very High |
High |
Moderate |
Low to
Moderate |
Very Low |
Clearly there is a tradeoff between speed (time to market), performance, and flexibility. The Cost-benefit of high-end PLD chips improves as small-geometry processes bring costs down while complexity of submicron designs increases cost of developing ASIC chips. The number of new ASIC starts has decreased from 12,000 in 1995 to 2,000 in 2002 and may not exceed 1,000 in 2008.
1. ASICs, typically require weeks to develop, plus have large up-front engineering costs
o It can take a year to design an ASIC, plus more time to design into end equipment
2. PLD’s allow designers to adapt to changes in design throughout overall design process (from prototyping to production)
o An engineer can design and program a high-density PLD in a matter of hours
o Bugs and glitches can be fixed during early production
o Features and functionality can be added while the product is deployed in the field
3. The ASIC market has far fewer design starts, but the ASIC designers typically have much larger budgets
The advantages of FPGAs over ASICs and ASSPs are design and product flexibility allowing users to reduce a products' time to market and avoid high up-front cost of developing custom solutions. The disadvantages of FPGAs are higher unit costs than custom solutions and they also use more energy affecting battery life.
One of the next big growth markets for semiconductors is in the SoC / embedded market. On what devices will the next group of innovators and inventors build their products? Xilinx and Altera (along with one-time programmable processor maker Microchip) have the early edge; with the industries goliaths Intel, IBM, and TI currently offering less flexible products and development platforms. Who the semiconductor intellectual property vendors and end-product innovators embrace will determine which platform overtakes the shrinking market for product specific semiconductor devices (aka ASICs, ASSPs, etc).
Programmable logic devices (PLDs) are semiconductor logic blocks that can be programmed after they are manufactured. The most common PLDs are Field-Programmable Gate Array (FPGAs). As integrated circuits become more complex FPGAs become more cost effective in electronic devices within communications, storage, industrial, consumer electronics and other end product markets. PLD chips will always be larger and slower than ASICs, but as more focus and energy is expended on design, more designs move to PLD.
"The cost of developing an ASIC, in general, doubles every generation. On 180nm, it was US$7-8 million, on 90nm it was in the neighborhood of US$40 million. As of yet, there are no ASICs on 65nm, but the development costs would be US$70-80 million."
-Xilinx CEO Wim Roelandts, Apr 2007
The programmable logic device (PLD) market is a duopoly with Xilinx and Altera controlling 85% of the market. From design win to customer beginning volume production can be two years so market share is a lagging indicator, but by all accounts Xilinx and Altera appear to be continuing to gain market share. An FPGA family typically reaches peak sales four to five years after introduction.
These two companies have taken a billion dollar of revenue from ASIC vendors over the past five years. The PLD market has grown a little less than 10% annually the past three years, but the market is expected to grow 10 – 15% annually over the next five years and the top two firms should grow earnings even faster. With the growth of embedded systems and the utilization of communication systems it is possible that the programmable logic market’s growth could even accelerate further. However, it may be a couple years before the PLD market growth begins accelerating.
In the Programmable Logic Device (PLD) market Xilinx and Altera are followed by Lattice with 7% market share, Actel with 6%, and Quicklogic with 1%. Over the next few years Xilinx and Altera should not lose or gain meaningful market share. These companies will grow because the four billion dollar PLD market is growing within the $72 billion semiconductor logic industry. PLD’s are receiving major designs wins in the overall semiconductor logic market specifically in products that used to be serviced by Application Specific Standard Products (ASSP) and the Application Specific Integrated Circuit (ASIC).
Every year PLD’s make more economic sense as designing and manufacturing semiconductor devices become more time consuming and more expensive. Initially just a prototyping tool, PLD’s can now cost effectively ramp up to 100,000’s of units. Some low cost PLD’s such as Altera’s FPGA Cyclone® series are now in devices selling in the low millions of units a year.
PLD’s offer a much more stable platform than other logic choices such as ASIC’s or ASSP where the devices have to be re-designed and re-tested every couple years. This can be a problem for products such as automobiles which typically are not re-designed more than once or twice per decade.
As more chips are embedded in different products the addressable PLD markets expands. The volume point where ASIC's are more cost effective over PLD's is now almost 100,000 units, five years ago it was 10,000 units. Due to this PLD is taking market share from the ASIC industry and ASSP industry.
While ways to measure semiconductor markets can vary widely the current markets that PLD can and is moving into are the ASIC market valued around roughly $15B, the ASSP market valued around roughly $15B, the very high performance DSP market valued around $3B, and the embedded processor market valued around $3B. If one was to include some of the high performance ASSP vendors such as Broadcom in the ASSP market it could be valued much higher (potentially around $45B), but the PLD vendors are not about to compete with these products.
Xilinx has a customer base more diverse across industries and while Xilinx’s competing chips often offer better performance Altera has won several recent design wins because of its focus on lower power (and correspondingly better battery life). Both use a fabless model which keeps capex requirements low and taxes low, due to products being manufactured overseas. Both companies have high margins, no debt, and around $1B in cash.
|
Company |
Symbol |
Market Cap |
Net Cash |
2008 P/E |
2007 P/CF |
ROE |
Net Margin |
Yield |
Growth |
|
Xilinx |
XLNX |
$7B |
$1B |
15 |
14 |
21 |
20 |
2% |
15% |
|
Altera |
ALTR |
$6B |
$1B |
16 |
25 |
24 |
23 |
1% |
15% |
The leading PLD vendors will stay at the cutting edge of manufacturing processes. When a fab begins a new, smaller fabrication process they like to use PLD’s as one of the first chips they manufacture because they are high volume with relatively straight forward manufacturing designs. Setting up the fabrication process for microprocessors and ASICs are much more difficult, and getting more and more difficult with each new smaller geometry. The latest FPGA products are at least one or two generations ahead of any ASIC products. It is a safe bet that Xilinx and Intel will both remain at the cutting edge in manufacturing processes.
Founded in 1984 and public since 1990, at $.83 a share when adjusted for splits. Every year for the past decade Xilinx has generated positive free cash flow, including in 2001 and 2002. It has a global customer base of over 21,000 customers (top 15 customers account for 33% of sales). It sources semiconductor wafers form multiple suppliers.
Xilinx has a little over 50% share of the PLD market. It is the 3rd largest company in the overall logic market. In 2007, a survey of semiconductor customers ranked Xilinx, for the first time, as the most important vendor, ahead of IBM or Intel. Over the past decade it has slowly been moving up the list on this survey.
Xilinx’s Virtex FPGA’s can include up to two embedded IBM PowerPC cores, offering embedded system designers the ability to build a system-on-a-chip capable of running an embedded OS like Linux. It has been making solid progress in the high performance DSP market as well as the embedded processor market. It offers an embedded development kit based on the Eclipse framework.
Xilinx currently trades at an opportunistic valuation with a 2008 P/E of less than 15. It has a ROE and net margin around 20. Including investments and subtracting a billion dollar convertible the company has a net cash position of one billion. This cash position has been shrinking as the company buys back shares. I have modeled revenue growth of 12% and earnings growth of 15% (due to share buybacks and margin improvements) over the next five years. Altera’s valuation is very similar.
Altera was founded in 1983 and has been public since 1989. It has a global customer base of over 14,000 electronics equipment makers. It uses a single source, Taiwan Semiconductor, for its semiconductor wafers; subjecting the company to any potential production issues incurred by Taiwan Semiconductor.
Xilinx and Altera have solid software design platforms that customers use to program their chips. The dynamics of the PLD industry give the top players competitive advantages due to the platform they each have created. Each platform has built a strong moat based on the number of integrated circuit design engineers that understand it (and whose careers are based on this knowledge) and the amount of intellectual property available.
Thousands and thousands of engineers make their living because of the knowledge and expertise they have of the development platforms they use. Switching costs are high for engineers to learn new design frameworks. In addition, thousands more engineers learn Xilinx and Altera's development tools every year in colleges in the USA, India, China, and the rest of the world.
Most PLD’s are built using Xilinx or Altera’s development platform where they can also use thousands of IP cores (re-usable intellectual property) built by Xilinx or Altera (depending on whose platform you are using) and offered for free (or nearly free). Their platforms also support 3rd part IP cores.
Each company plus dozens of third party companies build re-usable components (IP cores) that increase engineers’ efficiency. For ASICs, the largest impediment to IP reuse is verification, which can consume 70% of the development cycle. With FPGAs reuse is easier as the FPGA vendors pre-verify 3rd party vendors IP.
The success of a design platform largely depends on the quality of the IP building blocks. A huge advantage that the PLD firms have over the EDA firms is a much easier migrating path from one manufacturing process to the next. This allows IP cores to be easily rolled into the next generation architecture.
The ASIC market has struggled to embed and verify 3rd party IP. With Xilinx and Altera offering end-to-end design capabilities from design tools to IP cores, including free IP cores and a marketplace for 3rd party IP cores, they are creating an ecosystem where the customer’s have one neck to choke when problems arise. Creating an environment similar to IT departments' migration to a single primary provider (Microsoft, Oracle, IBM, or SAP) so they can have a primary partner to lean on when issues arise.
PLD’s allow IP cores to be embedded at the software level, not in the hardware. This simplifies the process so IP can be tested and verified by the supplier on a finite number of chips since FPGA’s are a common single chip manufactured in volume by the fab; as opposed to ASIC’s whose physical design is dictated by the IP cores themselves.
A behavioral change is needed to design embedded systems and programmable chips offer a different development paradigm. The Internet and web programming could never have moved as fast as it had if websites had to be bug-free when first released, a requirement for ASICs.
Having IP that is available, certified, and supported is a necessity. Programmers on the Java EE and .Net platforms are offered thousands of tested and supported libraries. This is a basic requirement of any enterprise development platform. IP sub-systems (objects) must offer easily understandable interfaces (API's) that allow the designer to focus only on integration and their designs differentiating factor.
A serious question mark is how can EDA vendors build an evolvable design platform incorporating IP cores that can subsist across multiple ASIC generations. FPGA's allow the platform to be based on single chip architecture, similar to how the x86 platform enabled the emergence of the PC.
Every year more IP cores are built and more engineers learn these two companies’ platforms. It would be hard for new entrants to overcome the platform network effects that have been built by Xilinx or Altera. It would be nearly impossible for a startup PLD company to compete with this breadth of knowledge. The ability of Xilinx and Altera to build developer ecosystems may eventually trump all but the highest end digital and mixed signal chips.
Who has the best platform for the next generation of semiconductor devices: Intel, TI, IBM, Xilinx, Altera, ARM Holdings, Analog Devices, Microchip Technology, etc? What platform will large companies and entrepreneurs use to build the next generation of products?
One-time programmable processors are one source of competition for embedded system solutions; with Microchip Technology’s programmable microcontrollers and TI’s DSP platform.
Many entrepreneurs love Microchip Technology’s one time programmable processors and their straightforward API. Microchip has roughly a 50% market share for 8-bit processors and over 90 of its microcontrollers ship with some analog component. It also ships a microcontroller with flash memory. It has built a good developer community with over 3,000 embedded designers attending its’ thirteen separate 2007 conferences held in four different languages in seven countries.
It released a 16-bit product in 2003 and a 32-bit product in December of 2007; which both use the same development tools as it’s widely used 8-bit version. Microchip is the only microcontroller vendor supporting 8-bit, 16-bit, and 32-bit products on one development environment. Its 32-bit microcontroller, the PIC32 Family, uses the MIPS the MIPS architecture and was on EDN Magazine’s prestigious list of “Hot 100 Products of 2007”.
Microchips generic processors will never come close to the performance of PLDs, but it should continue to see success in products with lower performance requirements.
ARM Holding's processors are the dominant processor for mobile phones and the most widely-used 32-bit microprocessor family with over 75% of all 32-bit embedded CPUs. Arm processors are soft core IP designs that can be embedded in ASICs or FPGAs and are really complimentary to hardware provider’s offerings.
Intel’s x86 processor has a large community, but an inherently slow architecture that has not seen very much success outside of the mainstream PC and server marketplace. Intel is creating a multi-purpose "programmable Intel architecture machine" called Larrabee for high-performance computing and discrete graphics. Larrabee will be Intel's first "many-core" product" and is expected to be available in 2010. It is debatable how competitive this device will be with the devices in the market two years from now. In the meantime, how much embedded systems market share can the PLD vendors capture?
Texas Instruments’ is the leading DSP platform with over 50% market share and is also a leading player in the high performance analog chip market. TI risks losing market share from several diverse competitors nipping at its heals such as platform PLDs targeted at DSPs and mixed signal devices, lower end device providers from China such as Vimicro incorporating custom and 3rd party IP, end market specific mixed signal competitors such as Sigma Design’s DSP, and finally seeing competition for its single chip mobile phone solution (integrated application processor, baseband, and support pieces).
Analog Devices nascent and somewhat beleaguered DSP processor platform Blackfin has some strong advocates and Microsoft’s .Net Micro Framework, which is targeted at embedded systems, was recently ported to Blackfin; increasing the toolset available for devices built around Blackfin
A 3rd party EDA tools vendor may develop a strong integrated meta-platform that can act as a neutral party and potentially support building devices on top of several different vendor’s platforms. Granted this sounds more like a pipe dream.
While Altera and Xilinx have great platforms with strong communities there are concerns that that the business model for intellectual property (IP) is poorly positioned. A considerable amount of IP is given away by the FPGA vendors which undercuts some potential IP vendor’s offerings. The FPGA vendors will have to regulate themselves to providing the framework IP allowing 3rd party IP companies to provide important add on intellectual property.
As with any company developing a platform there are conflicts of interest that must be adroitly managed. Battles for platform supremacy often have winner take all outcomes with all other competitors left as niche players. The next decade will see further consolidation around very few design platforms.
Actel with 6% market share has a strong niche in anti-fuse FPGAs (one time programmable) and in flash based FPGAs. Flash based FPGAs is a segment Xilinx and Altera have stayed away from, but February 2007 Xilinx’s introduced a flash-based Spartan-3AN and was named to EDN's Hot 100 Products of 2007.
There have been many new entrants with Flash based FPGAs in recent years as venture capitalists and entrepreneurs may feel more compelled to compete in a corner of the market not dominated by one of the two industry leaders. Flash-based FPGAs have entering several new markets end markets and they may actually grow faster than the more traditional SRAM-based FPGAs.
Actel began shipping a Flash based FPGA in 1999 and they appear to currently have a strong leadership position in this good niche area of the FPGA market. Actel has a deal with ARM for putting soft versions of ARM7 processors in Actel’s non-volatile flash-based FPGAs. Altera recently began offering FPGAs that also support ARM processors.
Flash benefits from being non-volatile (e.g. it doesn't need power to maintain the chip's data), but SRAM devices will always be faster. Unlike SRAM based FPGAs, Flash based FPGAs do not have to be booted by an external device to load the software so they are more of a true single chip solution comparable to ASICs (related to power requirements and the number of components in circuit board design).
Lattice Semiconductor with 7% market share has made acquisition to attempt to increase its share in the PLD market. Having these diverse PLD software development platforms for its various acquisitions has been a disadvantage. In 2001, Lattice’s division Vantis and Agere (at the time Lucent) gave it a 25% market share in the PLD market. Lattice is strongest in the slower growing CPLD segment where it is number three in market share after being over taken by Xilinx in 2007 for the number two position. Altera is number one in CPLD market share.
QuickLogic with 1% market share targets specific applications (similar to ASSP's) and is strong in power sensitive markets.
Newer companies such as ChaoLogix, Stretch, and Cswitch are entering the programmable logic space and trying to create a disruptive, game changing technology.
A big problem for start-ups and other semiconductor companies entering the programmable logic market is place and route tooling software, where Xilinx and Altera’s design platforms are dominant. Roelandts the CEO of Xilinx has stated that "there is no 3rd party place and route software; place and route software at Xilinx runs to 20 million lines of code.” A new company also runs into the problem that initially no one knows its development platform and the lack of IP available for the startups chips.
A larger better funded player like Intel, IBM, or Texas Instruments enters the market. With the Geneseo project Intel and IBM opened sourced what could have been a competitive advantage for them (discussed in a couple pages). A risk could be a strong player like Texas Instruments, with its current powerful position in analog and DSP, purchasing Actel to get a foothold in the PLD market.
Due to the .com and telecom bubble PLD sales surpassed economical sense in 2000 and 2001. After bottoming in 2002 PLD sales have grown at a compound annual rate of about 15% and FPGA sales have grown over 20% annually. FPGA’s are now 75% of total PLD sales. In 2001, the PLD market peaked at $4.1B and in 2002 bottomed at $2.3B. Since 2002, both Altera and Xilinx have gained market share every year. The PLD market has only recently exceeded the peak experienced in 2001.
It looks like the economy is headed back into a recession, but unlike the 2001 recession, which was business capital spending induced, this recession looks like it will be driven by a drop in consumer spending. This will affect semiconductor sales, but it will not be nearly as dramatic as the drop in 2001.
The 2001 recession was caused by a glut in supply of technology products and an over investment in technology oriented startups. The 2008 recession has neither of these characteristics.
In 2001, companies like, Xilinx, Altera, Cisco, EMC, Sun Microsystems, etc sold (and often financed the sell) products to startups and other customers in the technology sector whose unprofitable business models caused them to go bankrupt when the market turned down and venture capital investment dried up. These bankrupt or dramatically downsized companies dumped their newly purchased technology products on the market. So not only had demand decreased drastically; supply increased drastically. This caused a glut in the market that took years to work through.
This time supply chains have been optimized for just in time product deliver and technology companies’ balance sheets have never been stronger often in conjunction with record low sales to inventory ratios. While there may be some re-selling of technology equipment in emerging markets it should be minimal compared to what occurred in the telecom and .net bubble of 2001. In addition, within the semiconductor industry, PLD sales should fare better than most since they require less up front risks, such as time and costs.
In March 2006, LSI Logic, the first vendor to create a structured ASIC, closed its RapidChip division exiting the structured ASIC business. This was a blow to the EDA vendors who haven't seen significant growth since the 2001 downturn and have struggled to gain a foothold in the FPGA industry. By shuttering its structured ASIC division LSI conceded that customers are moving to FPGA’s for new products as it is much faster and more economical to initially design a Xilinx (or Altera) FPGA and migrate to the same vendor’s structured ASIC in volume, than design using LSI’s ASIC-based process.
In 2001, Altera introduced their HardCopy® structured ASIC platform and they are now a leader in the structured ASIC market. Altera has a straightforward migration path from FPGA to structured ASIC. They maintain compatibility with the FPGA architecture, but reduce the die size by removing the configuration circuitry, programmable routing, and programmability for logic and memory with fixed wired interconnect. Xilinx has a similar product called EasyPath®.
In 2000, over 80% of PLDs went into communications related equipment. Today that number is just under 50% of sales and steadily dropping. Since 2000 the growth in PLDs for communications end products has significantly trailed other end markets.
In fact, if the growth in the communication sector picks up, like we think it will in 2010 or 2011, this could be a major boost to the PLD market. The other end market segments like storage, industrial, and consumer electronics should continue to see strong growth over the next decade.
Programming many-core microprocessors and FPGAs is also a problem of educational deficits and development platform shortcomings. The developer’s choice here may ultimately be decided by who builds the simpler development platform or raises the level of abstraction through powerful framework IP.
Where the innovation is made will make all the difference. Will embedded systems use Platform FPGAs, dedicated microprocessors, or DSPs? With globalization and the Internet; China, India, Russia, and Eastern Europe are producing hundreds of thousands to millions of engineers. Altera has over 30 joint laboratory and training centers in China, but it is not alone as Xilinx and other major semiconductor firms are working to get the next generation of engineers to learn and innovate on their platform. What development platform will win the minds of the next innovators?
Competition is migrating from semiconductor companies solely competing based on their IC based on the platform they offer. Microchip is a successful example of this strategy. When the innovation in the IC slows the innovation moves to the platform; and the efficiency it bequeaths on the end product designer.
The information technology departments and particularly web software developers have discovered and embraced agile project development methodology. Rapidly shrinking product lifecycles are encouraging digital product designers to embrace similar prototype and release agile strategies.
As chips become more complex their design and manufacturing
costs increase dramatically. More
requirements require a complex design which produces more functionality to test
and debug, and a more complex design triggers lower yields and higher NRE in
the manufacturing stage. Embedded
systems are being designed around SoC. Industrial,
automotive, and consumer electronics devices are increasingly becoming
digitized.
Another example of an FPGA SoC is HDTVs. Most HDTV are now designed with a stand-alone FPGA or coupling an FPGA with an ASSP as a coprocessor. These Platform FPGAs include hard-coded DSP blocks, internal memories, and standard IP blocks along with the programmable logic.
High-end convergence consumer devices are a great opportunity for FPGAs. In addition, the growth in internet traffic is will create another round of build-out in the FPGA industry's sweet spot, telecom and networking.
In 2008 products built around an FPGA will be delivered. What will be new is that for the first time some of these products will be built on an industry-standard soft-core processor architecture from ARM which ARM released in 2007 specifically for FPGAs. Now the huge base of ARM engineers can develop software for embedded systems based on a FPGA system-on-a-chip and use the massive installed base of ARM software and tools. Will adding an ARM processor and its accompanying ecosystem to the FPGA platform be the inflection point for FPGA design wins and the growth that inevitably follows?
For well over a decade processor speed has been increasing more than memory speed, because much more R&D has been poured into logic and processor design. This is because logic (Broadcom, Xilinx, Altera) and processor designers (Intel, IBM) have far better margins than memory designers (Micron, Spansion, Samsung).
It is apparent that for servers the application performance improvement curve is flattening. Intel’s generic microprocessor architecture has overtaken many RISC processors because of the shear R&D money Intel has dumped into its x86 design and manufacturing compared to RISC makers. Even so, Intel is no longer dramatically increasing the speed of its processors.
Intel is only able to keep up with Moore’s Law by adding more processors. Inefficiencies abound from using multiple processors. Programmers are not taking full advantage of these multi-core and multi-processor machines because these machines are not offering the proper development platforms and compilers for object-oriented programmers to optimize for multi-core, multi-processor environment. Additionally managing memory in these multi-core architectures is more challenging. Moore's law is running into an instruction set wall because the data stream is too large.
Physical barriers (like transistor leakage) are inhibiting the ability for semiconductor companies to continue keeping pace with Moore’s Law. To keep up with Moore’s Law microprocessor designers are forced to build up by adding more processors. Processor manufacturers’ ability to scale clock speed and shrink transistors is decelerating. They must now use parallelism.
FPGA's have moved into supercomputers as the enabler of Reconfigurable Computing. Like any change in the marketplace, reconfigurable computing also represents a risk. Companies like Stretch are creating FPGA like software configurable microprocessors, but how straight-forward will engineers find the design tools?
The architecture for a RPU (reconfigurable processor unit) system is just now becoming standardized. While FPGAs theoretically have the ability to be reprogrammed at run-time software development tools have not supported this. While some proprietary systems, such as Cray’s FPGA-augmented XD1 and SGI’s Altix with RASC option, have seen limited success it appears FPGA’s may finally be entering the standard x86 market.
The reconfigurable computing data-stream-driven model may alleviate many of the performance barriers in the existing x86 server market. As prices fall and the speed of a single processor core stabilizes it is not farfetched to see FPGA's in desktop PC's and cell phones. A PC manufacturer may decide to use a FPGA chip to implement common tasks such as TCP/IP processing.
Xilinx and Altera are working with Intel and AMD to plug directly into new multi-core processor's front-side bus architecture. We may start to see PLD products for this as early as mid-2008.
Both Intel and AMD have developed their own technique to allow PLD’s to plug into the PC architecture. In June 2006, AMD announced its Torrenza program creating an interface to plug into its proprietary HyperTransport CPU bus. Not to be left behind, Intel, along with IBM and others, in September 2006 announced Geneseo an open source project creating a set of extensions to PCI Express aiming to help graphics chips and other accelerators. Geneseo proposals aim to extend Express in four general areas: fine-grained power management, locking shared memory, hints to help coherent processor handle I/O more effectively, and efficiencies for mapping virtual to physical memory.
PC and server architectures are highly standardized alleviating many of the advantages of PLDs. However, as computers take over more real-time computing functions, like processing high definition video signals, there may be more of a need to balance the inflexible speed of ASICs and the flexible sluggishness of software for x86 processors. Some examples of new real-time PC computing functions could be replacing proprietary PBX systems with generic servers running unified communications software or streaming video from a home server appliance.
The platform aspects of Xilinx and Altera’s business model do not require that the companies offer the best of breed technology; although they often do. The PLD market is not prone to the price completion seen in the ASIC, processor, and memory markets because of the high barriers to entry caused by the required software platforms and the lock-in once a vendor selects a FPGA design platform. This makes the two dominant PLD vendors very profitable with margins over 20%.
The PLD market will see solid growth for years once several macro trends emerge causing a tipping point in PLD demand. These emerging trends are
1. The increasing complexity of new semiconductor devices
2. The new markets for PLDs due to microprocessor performance barriers and dramatically shortened end product lifecycles
3. The maxing out of the current communications infrastructure
4. World-wide growth due to globalization, potentially temporarily restrained due to a US consumer led recession
If you have any questions or comments related to this white paper please contact me at Xtensible Solutions.