Global News

November 17, 2005
Robotics Wave pushing the industry forward

UNITED KINGDOM -- Gloom and despondency still dominate the semiconductor industry. Bad news is good news; seized upon with a perverse sense of satisfaction, while good news is discounted as a mere aberration. This is now no longer a question of pessimism versus optimism; it is rapidly becoming an industry milestone. A year is a long time in the Moore’s law industry cycle and the current doom-inspired procrastination endemic in far too many companies is condemning them to long-term decline. In the words of an internationally famous CEO: "The most useful and positive aspect of today’s pessimistic market environment is, does it allows companies like ours to achieve higher growth rates via market share gain 'make lunch or be lunch'." We could not agree more.

Each year companies bet the ranch on the next-generation wafer fab, knowing if they failed they would probably kill the company. A US $30 million fab was just as unaffordable in 1970 - with a total semiconductor market (TAM) of only US $2400 million - as a US $3000 million fab is supposedly today - with a US $2,40,000 million total semiconductor TAM.

Yet fabs were afforded, and the semiconductor market grew, fuelled as much by Moore’s Law’s product enabling ability as the industry visionaries’ faith that saw problems only as opportunities. The industry’s belief in itself was never challenged or questioned. Sadly, this is far from true today.

Since the 2001 crash, industry has never really regained its confidence, neither too have the industry watchers and investors. Constantly berated by an industry-insensitive stock market, few companies are showing the grit and determination that made this industry great. In short, the world’s largest semiconductor firms are in danger of becoming a club of established static companies, allowing the true dynamics of the industry to flourish outside in the younger generation of companies and the new geographic entrants.

Significant growth ahead

The annual IC unit growth trend shows much less year-on-year volatility than the equivalent IC value growth, and a much more stable long-term trend. The IC market value grew at a CAGR of 15.7 percent over the period 1984-94, slowing dramatically to 7.3 percent over the period 1994- 2004. Looked at in units, however, the comparable growth rates were 7.2 percent and 10.0 percent, respectively. By this measure, the IC market growth is accelerating, not slowing, and the real problem is ASP not growth. This also serves to justify the conviction that there is significant future growth in prospect for the industry, and that the real industry issue that needs to be addresses is not lack of growth but the intensity of competition. Over the long-term, the average number of ICs per wafer start is virtually constant, with feature size reductions translating into ever more complex chips.

With IC units currently growing 10 percent per year, at no net increase in overall ICs per wafer start, industry needs to add 18 fully-loaded 30k wafer start per month, 200mm fabs per year just to keep pace. Adding new capacity is thus not the problem, just the fact that investment tends to bunch and fabs come on line non-linearly. Over capacity: what over capacity? The current trends are good.

Market trend: Robotics wave

Even though the all-pervading mobile phone and PC handle all our communications and information handling needs, they cannot carry out more labour-intensive chores like cleaning our floors or tidying the home or office. We have waited over fifty years since the basic concepts arose for products that will save us from such drudgery. At last, the market need, technology and economic justification have come together to make sense for both the user and the manufacturer to create a new electro-mechanical robotic system, with the promise to provide a corresponding new chip market over the next few years.

The growth of electronics over the past fifty years has been characterised by waves of key applications. The development of the semiconductor industry was initially supported by military and space projects that culminated in the first marketable volume product containing ICs in the calculator. Following this success, industry became more market led by waves of key applications that pushed the industry forward.

The analogue wave was led by the TV and VCR, and the digital era has so far had two waves - the first starting 1980 and led by the PC, with the second wave led by both the Internet and digital consumer (including mobile phones). We are now seeing the beginning of the robotics wave, a period where applications are aimed at human labour saving and an extension of human skills, as opposed to the products for just information handling and entertainment as in the recent past.

Now is the right time!

Despite fifty years of development, the robotic market is still in its early stages, with companies still trying to define market requirements. We believe that the market will ultimately be very broad and, as with every embryonic market, there will be some products that will be winners, some dead ducks, some that never appear and others that will be confined to specialist niches. The fact that some of the world's leading corporations are active in this market will enable them to use their financial muscle and resources to make this a market with global mass appeal. Given that the mechanics, electronics and software required to make a powerful robot are now readily available, the timing is now right to develop the market.

Looking further out in time it is both types of intelligence service robots, which will bring the greatest reward for the robot manufacturers. We expect total robot OEM sales revenue, initially dominated by heavy industrial robots, to rise from US $5700 million in 2004 to US $32,600 million in 2009, a CAGR of 415 percent. Although it is dangerous to make a direct comparison between the human brain and a silicon CPU, it is useful to look at it to give some perspective. We have used MIPs as a measurement comparator despite the fact that it is too general a means of CPU rating performance, purely because it is currently the only one quoted.

Based on this measure, it would take 300 million MIPS (300 trillion instructions per second) to emulate the brain, or looking another way, today’s typical 2k MIPS PC is equivalent to 0.01 gram of brain matter, the typical brain weighing in at 1.5Kg. For completeness it must also be noted that the human eye, as well as detecting light, also computes motion and the edges of a target object, so the human body also has distributed computing power in its nervous system as well as in the brain.

Industrial robots dominating the scene

It is thus a much bigger task to replace than simple raw MIPs processing power. Luckily useful robots do not need the brainpower of a human, the brain of a fish (1,500 MIPs) is enough for guidance through the home, so useful robotics is therefore well within our grasp, both technically and commercially. And as robots increase in computing power and intelligence, the market will start to consume many more semiconductor components, including 64-bit CPUs, 32-bit microcontrollers, a variety of sensors and plenty of memory. As such, we estimate that the robotic semiconductor market will increase from a modest US $900 million in 2004 to nearly US $12,800 million in 2009, an annual growth of 72 percent.

Industrial robots presently dominate the robot market and there is an early start of some simpler robotic domestic appliances such as robotic carpet cleaners and lawnmowers. As time progresses more domestic orientated products, still rich in electronics will enter the field and these will be made in moderately high volumes thus bringing down the semiconductor ASP from US $909 in 2004 to US $329 in 2009 a CAGR of minus 18.4 percent.

Configurable microprocessors made by ARC and Tensilica are now coming into their own. By designing a chip using configurable elements, tailored for the function, the size of the chip can be cut by over 20 percent over the standard microprocessor circuit IP core route. This makes a big difference on a low-cost volume product such as an MP3 player selling in the millions. Other advantages of configuring the microprocessor before it goes onto the chip are less immediately obvious, but also important. If you configure the microprocessor you can add different features and functions that truly differentiates your chip and end product. The other advantage is that by configuring the product and making it completely different than others it makes it much more difficult to clone as the chip and the software as they are not available on the open market.

Doubting profitability of individual SoC designs

ARM has also spotted these advantages and has now added a configuration facility to its DSP extensions so that just the right amount of DSP performance can be designed in without the need of an additional co-processor. The new technology of re-configurability differs from that of the traditional configurable processor. Configurable microprocessors have an instruction set fixed in silicon, while reconfigurable circuits have an instruction set defined by software.

With digital consumer market, OEMs talking of market windows of months rather than years and 90nm, IC developments seeing design, mask and production costs rising to tens of millions of dollars -- there seems to be an issue of whether the individual SoC designs will ever be profitable.

Additionally, the move to 300mm wafer sizes will potentially yield thousands of good die for these unpredictable and volatile markets at a time when design times are moving out thus causing a much higher design risk. FPGAs can offer a solution to fast moving markets with changing specifications, but often rule themselves out on a cost and power basis. Fortunately, both the Altera Cyclone and the Xilinx Spartan series now have low-cost versions available, with both ARC and Tensilica configurable processors available as optional IP.

Exploring reconfigurable territory

One of the latest ideas is to incorporate reconfigurable circuitry directly onto a SoC applications-orientated ASIC platform. This solution is being used by STMicroelectronics by using an FPGA embeddable fabric from a French company M2000. Also Japanese companies Panasonic and Toshiba integrate British company Elixent’s reconfigurable algorithm-processing array IP with configurable processor cores from their own companies – all on a standard CMOS process.

Other companies are moving into the reconfigurable territory. PicoChip of the UK is sampling a chip for mobile and WiMAX base stations. It uses 329 processor elements for data and control planes, and 15 accelerators for error checking and correlation. Another UK company Aspex Semiconductor combines a Sparc microprocessor core with an array of 4096 processing elements in a matrix for video applications.

In Japan IPflex has an array of 376 processing elements that can be re-configured in a single cycle at 166MHz and is aimed at image processing applications. US company Stretch has taken Tensilica’s processor and added it’s own extension fabric to run additional instructions, changed in 80 microseconds, again for the image processing market.