blog

Optical Parallelism - Comb Lasers or Laser Arrays for Next-Generation Interconnects

Written by Frank Smyth | Nov 5, 2024 2:19:11 PM

 

I was delighted to be given the opportunity to speak at the ECOC Market Focus session this year, and to share the stage with some of the leaders of our industry, including Intel, Nvidia and Broadcom. In this blog, I wanted to share my talk on the subject of increasing optical parallelism, and to add in a few additional comments. 

To give some context, it’s worth exploring what we mean by increased optical parallelism, and whats driving the need. Essentially, we are referring to techniques to overcome scalability limits in optical networks by adding more channels whether these are more wavelengths, or more fibers. The most well known example is probably DWDM, where multiple channels, each on a different wavelength, are added to a fiber link to increase the capacity. Currently, it's getting harder to keep growing those single wavelength channels, so now multiple wavelengths are needed per channel. This is one example of increased optical parallelism. Similar scalability problems are presenting in numerous network segments.

So what is driving the need? In the more traditional optical telecoms and cloud datacentre segments, there is continued exponential growth in data created, copied, and consumed, and it’s showing no sign of slowing down. Coupled with this is the fact that further significant growth in symbol rates is becoming increasingly harder to achieve, and this gives rise to the need for greater parallelism.

 

Added to all this is the explosion of AI and machine learning, where large language models have grown at a rate of 10 times each year for the past 6 years. ChatGPT alone conducts 2.5 billion sessions per month via its website, excluding API calls, with each session averaging 8 minutes and consisting of many inferences. Lower data rates and simpler modulation and processing are used in order to meet this sector’s requirement for low latency and low power so, to meet the colossal and concurrent demand for bandwidth, many, many channels are needed - increased parallelism. 

The opportunities for increasing parallelism exist through the introduction of more fibers or increasing the number of wavelengths per fiber. As a comb laser company, our focus is on the wavelength domain but such is the demand that it’s likely that both domains will need to be heavily exploited.  

The two primary options for increasing optical parallelism in the wavelength domain are:

                                  →  laser arrays, and                                                                                   → comb lasers

                                                             

Expressed simply, a laser array consists of multiple lasers, each emitting a single, independent wavelength on an individual waveguide. These can range from an array of DFBs (either individual or on a single chip) to an array of iTLAs in a multi-wavelength transceiver.  

In comparison, a comb laser is a single laser emitting multiple, precisely spaced, phase coherent wavelengths on a single waveguide. Comb lasers exist in various forms, including mode locked laser diodes, kerr micro-combs, and gain switched combs, each suitable for particular applications, and all with their own pros and cons. 

So what are the potential applications for comb lasers where greater parallelism is required? In my talk at ECOC, I simplified the myriad of transceiver and interconnect technologies into two broad categories: 

  • coherent transceivers in the C-band which covers optical transport, metro network, and datacentre interconnect, and 
  • intensity modulation, direct detection (IM/DD)  transceivers in the O-band which covers the intra-datacentre network. 

Coherent C-band exists as two main architectures: 

  • pluggable coherent transceivers, which have always been based on a single wavelength, and  
  • high-performance embedded transceivers, which have predominantly had a single wavelength, but with some examples of two wavelength, or even higher, approaches.  

The challenge driving higher parallelism within coherent C-band is that symbol rate increases are getting harder to achieve, while component bandwidths are forcing changes to new materials. This drives a demand to higher wavelength counts, and we expect to see more 2- and 4- wavelength transceivers emerging - and possibly 8-wavelengths over time. The benefits of a comb laser in these systems include: 

  • lower cost
  • smaller footprint 
  • simplified control 
  • lower power demand as a result of fewer modules  

In addition, the unique features of optical combs can be exploited to simplify the digital signal processing, enable non-linear impairment compensation, or to allow unique features such as receiver bandwidth slicing and transmitter channel stitching. 

 IM/DD O-band can, again, be simplified into two large categories: 

  • For cloud networks, pluggable transceivers and co-packaged optics (CPO) already exist at 4 - 8 wavelengths and, certainly for CPO, are likely to push higher
  • The Optical IO segment is more focused on high performance computing and AI, and is already at 8 - 16 wavelengths, with some earlier stage solutions targeting up to 80 wavelengths

The key demand for, and benefit of, optical combs in this intra-datacentre space is in relation to the densification of the wavelength grid. Today, many of these systems use CWDM4 – a 4-wavelength grid with particularly large (20nm) spacing between the lasers for the prevention of interference. The difficulty here is that dispersion issues preclude the addition of more wavelengths with this spacing, yet the industry needs more capacity and therefore more channels. This need has driven the CW-WDM multi-source agreement which defines significantly more dense wavelength grids of 800GHz, 400GHz, 200GHz, and 100GHz spacing in the O-band. Whilst traditional distributed feedback (DFB) arrays can readily achieve 800GHz and 400GHz, manufacturing tolerance and aging-related drift makes intensification of the wavelength grids at or below 200GHz very challenging. Conversely, comb lasers have a precisely defined wavelength spacing that cannot drift and it is this that makes them extremely attractive for this grid densification requirement. 

However, it must be recognized that there are challenges with comb lasers too, chief amongst which is their optical power. Because the full optical power is shared across many wavelengths, it is challenging for combs to compete against laser arrays where each wavelength has the full power. 

Pilot Photonics has pioneered an approach that overcomes this limitation, and offers the best features of both comb laser and laser array. The result is an optical signal with the power per wavelength of a laser array, but with the stability, low linewidth and phase coherence of an optical comb. The architecture has been integrated into a monolithic Indium Phosphide (InP) chip and demonstrated in multi-wavelength coherent optical transmission systems. Each coherent wavelength can carry up to 60Gbaud DP-32QAM signals delivering more than 2 Terabits per second from a single InP laser chip and package. 

Now, thanks to the support of our European Innovation Council programme COCOPOP, we are also focusing on delivering a 16-wavelength version in the O-band for datacentre applications. 

Successful integration of this technology will enable the industry to satisfy bandwidth demands whilst simultaneously facilitating reductions in power requirements and size, the result of which will be increased data transfer and capability. 

Further information is available here: 

 And for more information or to discuss your project or needs, please contact: 

sales@pilotphotonics.com