Intel, Samsung and healthcare made headlines at the International Solid-State Circuits Conference (ISSCC)
There is a theory that Intel is more a manufacturing than a semiconductor company. As others have outsourced, Intel has continued to build new fabs and invest in the latest processes for its own devices.
It may now be doing some foundry work – it will build FPGAs at 22nm for startup Achronix Semiconductor, for example – but it sees as more important the controls over pricing and technology that comes from owning your own factories.
Poulson, the next-generation Itanium server architecture that the company unveiled at February's International Solid State Circuits Conference (ISSCC), provides some evidence to support this view. Introducing the chip, principal engineer Reid Riedlinger said that the design had been undertaken from the ground up to deliver something manufacturable well beyond initial runs on 32nm.
Poulson's predecessor, the McKinley architecture, gave Intel about nine years of use, although similar longevity for Poulson would take it beyond 11nm. This is foreseen on the ITRS semiconductor roadmap for 2015 and is the point at which new structures and materials are expected to become crucial. Given today's chip manufacturing costs, just five years might make Poulson worth the effort – even with its eight-figure non-recurring engineering cost.
Beyond that, the new architecture does offer some further clues as to Intel's future plans. The headline numbers are all about muscle. Poulson's first iteration provides 3.1 billion transistors on a high-k, metal gate process with eight cores, 700GB/s of interconnect via a ring-based interface, and an 11-stage in-order, decoupled front-end and back-end pipeline. That's in addition to 54MB of on-die cache, including 32MB of last level cache.
The cache number is huge. The eight cores take up just 720 million of the transistors and the cache dominates the rest. This raises the question, just what does Intel see as needing to be accessed so often?
Even with a number of power-saving features, Poulson initially seems to be aimed at applications where energy consumption is less of an issue – and in the server market, it >< is becoming as much of an issue as in portable devices, if only because users want to bash down utility bills.
Intel has said 'mission critical applications' are key, and Poulson also incorporates reliability, availability and scalability (RAS) features right through to radiation-hardened sequential latching. Here, a huge cache will cope well with masses of firmware and other functional requirements.
However, as the market waits on more detail about overall Poulson performance and availability, it seems reasonable to speculate on Intel's infrastructure ambitions in cloud computing and TV (E&T March 2011). For example, streaming high-quality HD video requires plenty of cache at both ends, but it is the operator who has to provide quality-of-service performance. Also, as more services are web-delivered, the number of those operators is increasing rapidly.
In the cloud, there is a trend towards the delivery of more sophisticated applications online than simply an office suite. Even chip design software itself has been tipped for such a shift.
Then, there is encryption, an issue that both streaming and cloud computing are making more pressing than ever. To run encryption through the processor in-line with content places a heavy burden on the cache. Some suggest as much as 12MB worth for video.
Seeing Poulson as an out-of-date science project may be well wide of the mark.
It was a something of a 'Physician, heal thyself' moment. The ISSCC keynote by Dr Oh-Hyun Kwon, president of Samsung's semiconductor business, rightly noted that it is hard for his industry to praise itself for 'green' products and success in lowering power consumption from smartphones to servers if it does not put its own house in order.
He noted that the industry's fabs use 40TWh of energy every year, more than the state of Michigan with a population of 10 million people. According to Dr Kwon, 'The'[power] consumption in 2008 was 47 per cent larger than that of 2001 and demonstrates a 7'per cent annual growth rate for the last two years.'
Samsung is not the first company to pick up on this dichotomy. Former STMicroelectronics CEO Pasquale Pistorio started banging the drum on energy in the 1990s. However, subsequent consolidation in chip manufacturing and Samsung's aggressive dominance of memory production mean that today, when a figure like Dr Kwon speaks, his suppliers have to listen.
His ISSCC presentation showed that Samsung is looking at the topic closely – it is as much about saving the planet as pounds, shillings and pence. An analysis of the various pieces of machinery used in the company's fabs revealed wide variations in their performance when conducting the same tasks. Specifically, improvements in efficiency of as much as 58 per cent should be available in some cases without the need for radical redesigns.
Beyond that, there is 'vampire' power. 'Approximately 72.9 per cent of the power needed for full-blown operation is still required for standby,' Dr Kwon said. 'This is because most equipment needs to stay at close-to-operation status during standby, and restart quickly.' Fab equipment is typically idle for merely four hours a day.
Addressing management processes and software could help. Samsung is looking specifically at improving the communication and control between its automated-materials handling systems and lot scheduling. Reducing the standby-to-operation ratios to below 50 per cent could cut consumption by 2.32TWh a year.
However, broad-based collaboration may also be necessary across the fab sector. Utilities tend to supply fabs with gases, ultra pure water (UPW) and other process materials on a worst-possible-scenario basis. 'For example, the range of [process cooling water] pressure specifications can be 3kgf/m2 to 6kgf/m2 to cover the most equipment requirements,' Dr Kwon said. Common standards could resolve this.
Finally, materials and thermal budgets. 'A study [by San Francisco Energy Watch] estimates that the electrical energy consumption for heating elements can be reduced by 80 per cent if the process temperature of silicon dioxide deposition is reduced from 780°C to 300°C,' said Dr Kwon.
This reduction may be necessary anyway because of difficulties emerging in the formation of shallow device junctions and increasing risks of process-induced damage to devices. At the same time, lower temperature processes can increase the impurities in the thin films used in chip production.
'In order to overcome such problems, new materials should be developed to form an impurity-free thin film at low temperature with low RF power,' Dr Kwon said.
These represent just a selection of the challenges presented by energy-efficient manufacturing. The good news is that the chip industry now has various collaborative, process-based mechanisms to address them. There is the Sematech manufacturing think tank and a wide range of alliances between companies.
Samsung itself works within one of these with its Common Platform partners IBM and contract manufacturer GlobalFoundries, as well as those in the wider Joint Process Development Agreement: Freescale Semiconductor, Renesas, Toshiba, Infineon Technologies and – because everything old really is new again – STMicroelectronics.
Lifting the BAN
The run-up to the approval of an IEEE standard often sees companies use its core content for pathfinder designs aimed to coincide with its launch. This is true within the 802.15 near-field wireless family. At ISSCC, several companies spoke about their work on the passage of that family's newest member, 802.15.6, the incoming standard for body-area networks (BANs). Now in its second draft the BAN standard is expected to receive formal approval by 2012.
The standard is designed to 'provide an international standard for short-range, low-power and highly reliable wireless communications for use in close proximity to, or inside, a human body'. It will be fundamental to long-discussed concepts such as smart plasters and implantable drugs.
According to Arthur Astrin, chair of the standard's IEEE task group, potential applications exist in: sensors, such as heart-rate monitors; stimulators (deep brain stimulation, pain management); remote control, say, tuning pacemakers; and elderly people assistance, such as fall alerts.
'They range through better healthcare, better therapies, better recovery and better comfort and convenience,' Astrin said. 'We also hope for some commercial applications.'
The specification meets a number of key criteria, low power and high reliability being two of the most obvious in a medical context. There are also measures to ensure coexistence with other wireless devices, patient security, and alarm responses of less than one second. But potentially the most fruitful aspect will be low complexity.
For example, one common media access control (MAC) layer underpins TG6. It has a familiar star topology with an optional two-hop mode. A central hub is defined and nodes connect to that. There are then three access modes – scheduled, improvised and random – according to the kind of connectivity that a doctor might require or the kind of alert the BAN may need to carry. Beyond the common MAC, there are three physical (PHY) layers, allowing for the use of ultra-wideband, narrowband or human-body channel communication.
'We've made it so there are lots of ways to simplify the design,' said Anuj Batra, of Texas Instruments who has been working with the narrowband PHY.
He explained that the narrowband flavour has been very specifically tailored to medical applications. 'We've made sure for a chip implementation that the RF is very easy and that you can build these systems in today's technology nodes.'
David Davenport, a senior engineer in the RF and Photonics Lab at GE, said this mattered because, 'We want to encourage very rapid growth in applications – lots of low-power, disposable wireless sensors.'
However, there is another aspect to the focus on simplicity across all three PHYs, the underlying idea that a standard that is simply constructed will also be easier for regulators to understand. Beyond that, the HBC version is being driven towards non-medical consumer products, such as wearable music players.
However, in the standard's truest sense, Davenport envisaged BAN-enabled hospitals where bedside nursing visits can be reduced, patients can be freed from their wards and beds, and data can be continously monitored. This isn't just meant to be easy to do and implement, but also a big cost saver.
ISSCC session chair, Alison Burdett, CTO of UK BAN specialist Toumaz, summed it up: 'For most people, the only body-worn device they might have is a Bluetooth headset. That is going to change rapidly,' she said. 'By next year's ISSCC, this standard should be ratified and we'll be seeing the first chip designs.'