Semiconductors's archive
SC08: Michael Dell Details Everyday Supercomputing
Michael Dell, chairman and CEO of Dell Inc., in a speech at the SC08 Conference in Austin, Texas, today highlighted the democratization of supercomputing thanks to the use of standards and off-the-shelf parts. That democratization, he noted, blurs the line between high-performance computing and corporate computing, which powers services such as Facebook and Microsoft’s cloud computing service (both of which are built on Dell hardware, of course).
It also means high-performance computers will be found everywhere — even on your desktop. In his speech Dell gave a boost to Nvidia and its use of GPUs in supercomputers by announcing that Dell would add 1 teraflop to its personal HPC workstations through a Nvidia Telsa card. The idea of a supercomputer on your desktop is a big theme at the show this year, with vendors ranging from Cray to SiCortex highlighting their high-performance workstations, and vendors such as Microsoft pushing new HPC software.
Moving far beyond the desktop, Dell also announced the creation of a 96-teraflop supercomputing test bed called Project Hyperion in partnership with Lawrence Livermore Laboratories and several other vendors. A teraflop is a measure of how many floating point operations per second a computer can handle. The fastest computer today is running at more than 1 petaflop, a thousand times the power of a teraflop. The goal of the Hyperion testbed is to figure out file systems, cluster management software and networking technology in a peta-scale environment. That environment is getting closer as more power can now be crammed onto fewer machines than ever before.
As an example of the increasing power, Dell pointed to server density improvements thanks to the use of blade servers and the ability to place as many multicore processors on them as possible. He gave the example of a Dell cluster built in 2003 that used x86 processors on 1,250 servers to create a 9.8-teraflop computer. In 2008 it took 155 servers to build a 10.7-teraflop computer.
As compute power has become democratized and cheaper — Dell also noted that five years ago $1 million could buy someone 2 teraflops of computing vs. 25 teraflops today — the world is finding more uses for it. That means that in addition to the traditional scientific uses such as climate change research and gene sequencing, companies use HPC to create animated films and to virtually build products before they are ever manufactured. It also means HPC is a bright spot amid a tumbling economy.
Nvidia Machine Takes a Spot on the Top Supercomputer List
For the first time ever, a supercomputer using Nvidia chips has achieved a spot on the Top 500 list of the world’s fastest supercomputers. The Nvidia-containing machine is ranked No. 29 on the list that was released late Friday; it’s a cluster built by NEC and Sun Microsystems that uses chips from Nvidia, Intel and AMD. As Nvidia and AMD, which bought graphics chipmaker ATI in 2005, push graphics processors for scientific computing, this is a big milestone. The rest of the list was pretty anticlimactic, with IBM’s Roadrunner computer narrowly beating Cray’s Jaguar computer to stay on top of the twice-annual Top 500 list. Both machines are petaflop computers, meaning they can achieve a quadrillion floating point operations a second — a record that was broken by Roadrunner back in June.
Cisco’s New Router Shows Need for New Processors
Cisco today announced a new edge router capable of moving 6.4 terabytes of data — the equivalent of 200 full length movies — per second. Om anticipated the product last week, pointing out that the influx of data traveling over the Web requires better and faster equipment to manage such complexity and traffic growth. What we also need is a different type of chip.
Routers have to process a lot of data really quickly. They are the air traffic controllers of the Internet: Each time someone types in a URL, the router has to figure out how to get the request to the correct end point. Since the number of possible routes grows every year, as does the number of times a router is consulted, old processors just can’t cut it anymore, especially at the edge where this Cisco router will sit. Instead of making chips for such devices more powerful (and more power-hungry), engineers are following in the footsteps of the server world and adding more cores.
Multicore chips are gaining in use in the embedded world for networking gear, set-top boxes and other applications. In recent routers Cisco had turned to Tensilica, a maker of specialized embedded multicore chips that can take tasks such as routing and video encoding and speed them up without requiring a lot of power. Tensilica calls its products data plane processors or DPUs. Cisco used those DPUs on its QuantumFlow processor.
Cisco is still using the QuantumFlow processor, but has its own custom-designed cores replacing the Tensilica core, inside this latest router, according to sources. However, Intel uses Tensilica cores for audio processing in its new line of systems on a chip built for video players). Other chipmakers, such as Freescale, which in June announced a new family of processors called QorIQ (say “Core IQ”), are tackling the problem of dealing with real-time data in low-power environments with more flexible, multicore embedded processors.
As real-time data processing becomes more important in areas such as reading routing tables and video and audio processing, Tensilica’s DPU cores and Freescale’s chips offer a way to process that information using less power than a general purpose CPU or even a graphics processor that might also be used for the task. Scientists at the Lawrence Berkeley National Lab are even using the Tensilica cores to try to build an energy-efficient supercomputer. In a connected world where devices have to do more but consume less, this type of design may be the way to go.
Nortel Cuts: Layoffs Come to Telco Land
Updated with information about Nortel cuts: Big job cuts are not just for startups. The grim reaper has started to take its tool on the telecom ecosystem. We have already reported about the job cuts at Nokia (600) and Motorola (3,000) and Nortel is likely to cut about 10 percent of its work force. The malaise has started to spread to component makers. For instance, JDSU cut 400 jobs, shut down seven R&D centers and three plants. Anadigics, a Warren, N.J.-based broadband wireless and wireline chip maker cut 15 percent of its work force. ST-NXP Wireless has cut its workforce by 500. Elsewhere Freescale and Verio also made some cuts. All these companies are reacting to the broader market declines and slowing demand from consumers. Vodafone, one of world’s largest telecom operators, is looking to cut jobs soon.
Update: Nortel today announced that it was cutting 1,300 jobs versus previously reported 3,000 cuts. Nortel had previously cut about 1,200 jobs. John Roese, Chief Technology Officer at Nortel along with some other senior managers is going to be leaving the company by end of this year. Roese, who participated in our Mobilize 08 conference, in my view is one of the smartest people when it comes to the broader direction of the telecom and technology industries. I think with his exit, Nortel loses a critical thinker. But right now it seems Nortel is fighting for its dear life as economic realities are crushing the company.
The original post was published yesterday.
Intel Hopes for Healthy Growth in Medical Devices
If mobile Internet devices don’t work out, Intel is also making inroads in the personal health market. The chipmaker today launched a patient monitoring device and online interface to connect doctors and their patients remotely. This is an industry Intel has targeted for years, but because of a host of reasons, a market where the company has never gained much traction.
The Intel Health Guide product gained FDA approval in July, surmounting one of the obstacles technology faces as it enters the human body, but it still has to gain market acceptance. Aside from FDA approval, medical tech needs to gain acceptance by caregivers, patients and health insurers. Getting all three stakeholders behind a new product or a new way of delivering health care is tricky. Because the end users of the device don’t typically pay for the systems, it’s a far more involved process than simply convincing a device manufacturer that your chips, software or component works best for their particular application.
Intel’s launching trials in the U.S. with Aetna, Erickson Retirement Communities, Providence Medical Group in Oregon and SCAN Health Plan, proving that Intel knows it needs to get the technology buyers on board before trying to explain the value to doctors and educate patients. I’ll keep my hopes up, because I do believe technology can deliver great benefits in the healthcare sector for everyone involved. However, I also know that getting accepted is difficult — and that acceptance only the begining. Once the stakeholders sign on to such a system, businesses and policy makers need to address questions of securing the information, as well as the ethics of remote monitoring. But the potential for huge growth as Baby Boomers age and the multiple types of chips such systems will need make it a market that’s hard to ignore.
IBM’s New Foundry Service Takes on Intel
IBM today said it will offer a new foundry service that could enable a startup to compete on the same level as Intel. IBM says it will make 45-nanometer, silicon-on-insulator chips designed by other other semiconductor companies as a contract manufacturer. This means everyone from startups to Texas Instruments can now design high-performance chips that can consume less power — ushering in new designs for consumer electronics, cell phones and maybe even servers.
As it becomes ever more expensive for semiconductor companies to build manufacturing plants to make their own chips, there are plenty of foundry services out there. However, IBM has combined two important manufacturing technologies to make this offering unique. One is the process node, which affects how many chips can be crammed onto a wafer. Smaller process nodes, such as 45 nanometer, offer better power efficiency (or performance) and better economies of scale. The other technology is silicon-on-insulator (SOI), which is more expensive than the traditional CMOS process. Foundries offering SOI technology typically do so at larger nodes, such as at 90 nanometers, where it’s hard to justify the higher cost of SOI when a chipmaker may be able to get the same cost efficiencies or performance and power gains by going to a smaller process node with CMOS chips.
If it’s an apples-to-apples comparison at 45 nanometers, it can be worth it to pay more for SOI, because a processor gains more performance, with certain types of power advantages provided by SOI that help make devices smarter and more energy efficient. ARM — the design and intellectual property firm behind many of the chips that act as the brains for cell phones, including the iPhone– has signed a partnership with IBM to offer some of its IP libraries as part of this foundry offering.
This foundry service will provide a compelling service for companies building cell phone applications processors, embedded chips for networking equipment and gaming chips, but it may not be the only offering out there for long. AMD also has the ability to make 45-nanometer SOI chips, which means as it spins out its fabs, the newly formed Foundry Co. may also sell such capacity to all comers. An AMD spokesman said he didn’t know when I asked.
This also heightens the divide in the semiconductor world between Intel and everyone else. Intel had resisted SOI technology for years — although it appears to be coming around for later generations of chips. However, by offering such cutting-edge foundry services, IBM and its partners are making it possible for other firms to keep up with Intel without worrying about manufacturing R&D. This means Intel is running two races — one in manufacturing and the other with its design. How long it can keep this up is anyone’s guess.
Ultra-wideband Decline Proves Perils of Chip Investment
Five years ago, the promise of a new networking technology known as Ultra-wideband was a living room without wires, where DVD players, set-top boxes and video accessories could connect with TVs over the air. Ultra-wideband (UWB) is a wireless personal area networking technology that can transmit large amounts of data for short distances using very little power. Over time, its promise expanded from the living room to the home office, as backers used Ultra-wideband as the basis for Wireless USB and the WiMedia standard.
So far, this dream hasn’t materialized, and the technology has failed to find a mass market. Today, we still have wires in both the office and living room, and a host of competing standards have whittled away UWB’s opportunity. In the last week, we’ve seen players exit the UWB business, and Intel announced that it has halted research on the technology. For venture firms who have invested nearly $400 million in the space, the fate of Ultra-wideband offers a cautionary tale about the perils of betting on semiconductor standards.
On Oct. 31, five-year-old UWB chipmaker WiQuest shut its doors when it was unable to raise more money or find a buyer for its technology. That led to heralds of doom for Ultra-wideband, with analysts and media blaming long-delayed product launches, expensive chips and a hostile regulatory environment. This week, Intel said it had discontinued its UWB efforts, saying the market wasn’t worth its R&D efforts. If the technology somehow manages to revive, Intel says it could buy up one of the six remaining startups in the space.
The plan makes sense for Intel because UWB, like Wi-Fi, Bluetooth and WiMAX, is a standardized technology. That means any UWB chipmaker will have the basic set of characteristics Intel needs to play in the market. Standards are a double-edged sword for venture investors. On one hand they are good for consumers and electronics makers because they enable multiple devices from different vendors to work together. Any Wi-Fi router should talk to any Wi-Fi chip in a computer, phone or camera. This helps drive consumer adoption and can lead to the creation of a huge market. Venture capital firms love this, because if a standard takes off it can build a company like Broadcom or Atheros that can generate rapid returns in a relatively short amount of time.
The other edge of that sword is that chipmakers who adhere to the standard can do little to differentiate their chips, which makes it easy to switch vendors and effectively commoditizes the product. This happened for the Wi-Fi standard back in 2000-2002, when venture firms put more than $2 billion into more than 40 Wi-Fi companies, only to see a few rise to the top. There is also the risk, inherent in all technologies, that the market won’t adopt it. This seems to be what’s happening for UWB.
Instead of seeing the technology completely die out, Eric Broockman, CEO of UWB chipmaker Alereon, argues that Intel’s retreat from the technology and WiQuest’s failure mean a shakeup similar to that experienced by the Wi-Fi market is happening with UWB. “Typically in this type of semiconductor investing there is a win-place-show mentality,” Broockman says. “One wins big, one gets acquired for a good price, one gets acquired for a not-so-good price, and everyone else goes away. That process in UWB is being accelerated by the current economic downturn.”
There were at least seven UWB chipset companies formed in the 2003 time frame. Now, many appear close to failure. WiQuest, which raised about $54 million, was one. Two others, Artimi and Staccato Communications, are both rumored to be running out of cash. Artimi has raised $31.5 million and couldn’t be reached for comment for this story. Intel Capital invested in Staccato when it was pushing UWB. That could position Staccato to end up being the company in the show category, because Intel might buy it for its intellectual property at a cheap price down the road.
Fighting for the win and place spots are Alereon, which has raised more than $70 million with a small amount coming earlier this year from SKTelecom; TZero, which raised $18 million in March led by CID Group; and Wisair, which raised $24 million in February led by Susquehanna Growth Equity. Radiospire is another player in the UWB market, but it appears to be shifting gears — or at least hedging its bets — by also making chips for transferring wireless HD video using a different standard.
Competeing standards are one of the reasons UWB is having such a hard time finding a toehold. For desktop personal area networking, Bluetooth and Wi-Fi are becoming more prominent — and have the benefit of cheaper chips. In video, UWB has conceded to Wi-Fi and specialized standards such as Wireless HD and WHDI. Those left on the playing field are quick to point out that UWB still has legs — and it might, if it finds the type of killer application that can drive adoption rates and increase chip sales to the point where they cost less to embed. But the shakeup happening here proves that chip investment isn’t for the faint of heart.
This article was also published at BusinessWeek.com