Causing an Avalanche: The latest advance in Silicon Photonics

On December 7, Intel published world record results for a silicon-based photodetector in Nature Photonics, and we wanted to explain the results in a little more detail in this blog. Before launching into the “in’s and out’s” of avalanche photodetectors (APDs) though, it is important to provide some context for the work.

Several companies are active in the field of silicon photonics because they believe that silicon has an advantage in making the very low cost optical parts needed for large markets. These potentially include a very diverse set of applications including supercomputing, data center communications, consumer electronics, automotive sensors, and medical diagnostics just to name a few. Up until now, only a few silicon photonics products have been commercialized as companies have been working through the latter stages of the development and qualification processes. This looks set to change over the next few (?3) years as several devices exit this pipeline and go on the market.


Intel has been doing research in this area for more than 5 years and has already reported on silicon modulators, silicon Raman lasers, and hybrid InP-Si lasers (view website). Last year we also published on a photodetector made from germanium and silicon that had a bandwidth of 31 GHz. The use of Ge is important because, unlike Si, it can efficiently detect light in the near infra-red which is the standard for communications. The drawback is that so much stress is developed in pure Ge films deposited on Si that defects are introduced near the Ge/Si interface. Careful design and processing is needed to minimize the impact of these defects on the electrical performance of the device, and this will be mentioned later.

We are now reporting on a different type of Ge/Si photodetector that has built-in amplification, which makes it much more useful in instances where very little light falls on the detector. It is called an avalanche photodetector because an avalanche process occurs inside the device. First, a negative and a positive charge (electrons and holes in semiconductor terminology) are created when the light strikes the detector. The electron is accelerated by an electric field until it attains a high enough energy to slam into a silicon atom and create another pair of positive and negative charges. Each time this happens the number of total electrons doubles, until this “avalanche” of charges are collected by the detection electronics. Click on the image below to see how this works. This amplification effect (called gain) is the key to the device, and it serves as the motivation for why anyone would try to do this in silicon and not just continue to use traditional InP-based APDs. The materials properties of silicon inherently led to lower noise and better performance in this avalanche process. Another reason relates to this bit economic trivia; an individual 10 Gb/s InP APD can sell for more than $200 currently and has a semiconductor area of roughly 400×400 ?m2. Even the much cheaper 1-2 Gb/s APDs used in fiber to the home (FTTH) still sell for $3-5.

Click to view animation

It has often been assumed, however, that while silicon photonics might be lower in cost than InP-based devices, its performance would be inferior. While this is true in many cases, one of the exceptions is the area of APDs, where silicon’s material properties allow for higher gain with less excess noise than InP-based APDs and a theoretical sensitivity improvement of 3-5dB. Sensitivity is the gold standard of detector benchmarks and is defined as the smallest amount of optical power falling on the detector that can still maintain a desired (low) bit error rate. We have recently achieved a monolithically grown Ge/Si APD with a sensitivity of -28dBm at 10Gb/s and a gain-bandwidth product of 340GHz. This sensitivity is equivalent to mature, commercially-available InP APDs and the gain-bandwidth product (GBP) is the highest reported for any APD, as shown in the graph below. The GBP is important because it describes over what frequency range that the gain of the device is available. InP-based devices typically have a GBP of ~100 GHz which means that they would have a gain of 10 at 10 GHz. However, at bandwidths high enough to support 40 Gb/s, the gain falls to about 3 which is not enough to justify the cost. Our device would have a still have gain of 10 at that same point. In order to realize the full performance potential from this material system though, we need to further reduce the dark current that is coming from the defects at the Ge/Si interface, and stop the interdiffusion of Ge and Si that occurs during annealing. This intermixing is problematic since the Ge causes higher noise than if the silicon alone was in the multiplication region. If we are successful, this work will pave the way for developing low cost, CMOS-based Ge/Si APDs operating at data rates of 40Gb/s or higher in the future.


There are two other directions that we are planning to go with derivatives of this technology. The first is to move to a waveguide-based APD. This will improve the absorption at wavelengths up to about 1600 nm because the effective absorption depth can be much greater in that type of device. It also allows for integration with other optical devices, as such demultiplexers and attenuators. Secondly, we would like to reduce the operational voltage from the industry standard of ~30V bias to something more common in consumer electronics to open up a much broader user base.

We hope that you have found this summary useful. Feel free to post comments or questions.

To find out more, link to Intel Press release

6 Responses to Causing an Avalanche: The latest advance in Silicon Photonics

  1. anonymous says:

    Interesting. You have essentially built a photomultiplier at the atomic scale. I would be curious to know what sort of latency there is in the converson. Are we talking microseconds or nanoseconds? How does the latency of your present Silicon-based APDs compare with traditional Indium-based APDs?

  2. DMeier says:

    I would like to know whether you operate the APD below or above breakdown ? Is your APD similar to a single photon avalanche diodes (SPADs), a silicon photomultipliers (SPMs) or multi-pixel photon counters (MPPCs) which all operate above breakdown ?

  3. Mike Morse says:

    Good questions. There is very little latency in the conversion process; nanoseconds or lower is typical for these types of devices. I believe that this device has comparable latency to that of InP-based devices based on the physics, but we have not benchmarked this.
    In regards to the question on SPADs, we operated this device just below breakdown as is typical for the “normal” operation of an APD. We might need to improve the dark current of the device before moving to single photon counting applications where noise is critical.

  4. abe suhami says:

    apart from trying to operate in the 1300nm+ region with Si-Ge devices, does your process and structure of the photodiode has or ought to have less dark current in the 600-800 nm wavelengths with Si photodiodes, whether operated in the APD or Geiger mode?

  5. Yimin Kang says:

    The primary photoresponsivity for Ge/Si APDs is 0.55A/W at 1300nm. The responsivity at gain of 30 is 0.55×30=16.5A/W.