Silicon Photonics Links Traditional ICs
All of the pieces are waiting; it just needs to be done.By John Blyler, Editorial Director
The trend in electronics is for ever-faster data transfers between and within system-on-a-chip (SoC) devices. Higher speeds typically mean faster clock cycles, which translate into higher power usage and increased heat generation–-a real problem for today’s energy-sensitive data centers. Optical interconnects offer a promising alternative--especially with advances in silicon photonics, which permit the integration of electronic and optical components on the same silicon chip.
Intel is a predominate researcher in the field of silicon photonics. Past Intel senior vice president Pat Gelsinger has stated, “Today, optics is a niche technology. Tomorrow, it’s the mainstream of every chip that we build.” IBM also works in this technology space.
The theme of enabling optical technology for every computing device was a keynote discussion at the recent Tech Design Forum in Santa Clara, CA. Dr. Mario Paniccia, an Intel fellow and director of the company’s Photonics Technology Lab, highlighted recent advances including the first integrated silicon-photonics optical link operating at 50 Gbps. This demonstrated link was scalable to greater than 1 Tbps via integration (see Figure 1).
Need for Speed
Terabyte (measured in trillions or 10 to the 12) computing is the norm for many industries, such as telecommunications and data centers. Those centers now manage 15 billion connected devices connected via the cloud. Other markets, such as medical scans, already require peta-bytes (quadrillions or 10 to the 15) for data transfer and storage. Such extremely high data rates present significant challenges for traditional chips and boards that use copper (Cu) interconnects. Data rates for such systems can experience high signal attenuations, which lower data-transfer rates.
Optical interconnects would help ease some of the problems facing Cu-based systems, explained Paniccia. Optical technology is already the standard for telecommunications backhaul networks to transfer data quickly over long distances. The next logical step is to apply optical technology for chip-to-chip and board-to-board data communications. The problem for optical is one of cost. Cu interconnects between chips cost pennies, whereas board-to-board Cu connections have only been slightly higher in cost. Yet optical has been considerably higher.
One area of promise is the laser. Paniccia noted that last year marked the laser’s 50th birthday. Lasers are a requirement for opto-electric systems, such as long-distance communications. But they haven’t been widely adopted into other areas since the creation of the optical transceiver--a chip that transmits and receives data using optical fiber rather than Cu wire. What is needed is a way to combine and economically manufacture low-power laser technology into a silicon chip. But how?
Silicon Doesn’t Lase
Six things are required for optical systems: light source, guide channel, modulation, photo detection, low-cost assembly, and intelligence. The challenge with implementing these functions into semiconductor chips is that silicon has poor optical properties. At this point in the presentation, Paniccia emphasized that the goal of silicon photonics wasn’t to replace silicon technology, but rather for optical technology to offer a complementary and equivalent alternative with unique benefits. Silicon photonics must address the 5-m problem (i.e., using optics for chip-to-chip, board-to-board, and server-to-server systems). To compete with existing Cu interconnects, however, optical solutions must be low-cost and manufacturable in high volume.
Today, silicon photonics are being integrated with silicon chips with no optical material or functions within the chip. Integrated optics that connect to the chip include devices like photo-detectors, demodulators, and light channels.
One problem on the transmitting side of this chip-to-chip connection is with the laser. Because silicon doesn’t lase–or produce lasing light–other materials like indium phosphide (InP) must be bonded to the silicon. But InP serves merely as a dumb light source, notes Paniccia.
Conversely, a plus for silicon is that it can be used to create high-performance gratings to control the wavelength of lasing sources. This helps to drive down other costs. The receiving side of the chip-to-chip optical system used germanium grown on silicon for photo-detection.
One of the biggest problems with the manufacturability of optical-communications systems from one chip to another has been the packaging and assembly of such systems. To be cost-effective, the chip-to-chip system (see Figure 2) has now been assembled using printed-circuit-board-like techniques and passive-optical connections. Such an approach allows for assembly-line efficiencies of cost and production. Paniccia noted that a demonstration communications link–a 5-m link with losses of only a few decibels–ran for three days with no errors at more than a 3-petabits data rate. That translates to a bit-error rate of less than 2 to the -15 exponential. No coolers were used, so the lasers did drift a bit due to thermal fluctuation. More importantly, such an approach could be scaled upward by adding more lasers.
Using optical systems would help to mitigate some of the high-data-transfer problems faced by traditional Cu interconnects. But using optical systems would also provide an architectural advantage to certain problems by removing the distance constraint. Paniccia explained that the shared memory of connected systems could be shared instantly. Using optical solutions means that electronics can be designed without the latency added by distance while using smaller form factors with reduced thermal density (i.e., less cooling needed in data centers).
For example, today’s chips have a lot of memory stacked on the same package with the multicore processors. With silicon photonics, that memory could be moved to another board or even another room. This approach opens up new architectures for electronic design--from the chip to the board and even larger systems.
Silicon photonics is the next big step in the evolution of optical and electronic systems. But challenges remain--especially in terms of power. Today’s consumer and industrial markets want ever-increasing features with ever-decreasing power. Today’s sub-5-mW embedded processors will soon become tomorrow’s 2-mW processors. But adding optical connections will require more power.
The key, explained Paniccia, is not for optics to replace Cu-based electronics, but to use optical connections where they bring value--as in the server market. Server systems typically operate around 150 W, which could easily support the power needs of opto-electric devices. The challenge is that lasers don’t do well next to high-power systems due to thermal drift. One solution would be to package the silicon-photonics system on the top of other boards and away from the major heat sources.
The long-term goal is to bring silicon photonics into the mass market by first creating an infrastructure for the design, fabrication, test, and assembly of these systems. All of the pieces are there. It just needs to be done.
- Silicon Photonics Overview with Mario Paniccia, Justin Rattner, and John Bowers, http://www.youtube.com/watch?v=U5HhRB2W-DA&feature=player_embeddedUnderstanding.
- “Optics In A Wireless World,” WSD, April 2001, http://chipdesignmag.com/display.php?articleId=4820.
John Blyler can be reached at: email@example.com