Q&A with Kotura's CTO: Integration styles and I/O limits
Wednesday, October 17, 2012 at 8:20AM The second, and final part, of the Q&A with Mehdi Asghari, CTO of silicon photonics start-up, Kotura.
Part 2 of 2
"When do the big players adopt a new technology and go from an electrical to an optical solution? In my experience, usually when they absolutely have to."
Mehdi Asghari, CTO, Kotura
Q: Silicon photonics comes in two integration flavours: the monolithic approach where the modulators and detectors are implemented monolithically while the lasers are coupled externally (e.g. Kotura and Luxtera); and heterogeneous integration where III-V materials such as indium phosphide are bonded to the silicon to form a hybrid design yet are grown on a single die (e.g. Aurrion). Does one approach have an advantage?
A: I have a III-V background and converted to silicon photonics over 15 years ago. The key issue here is what are you trying to do? Why are we going from III-V processing to silicon? Is it the yield and process maturity or the device performance for actives?
If it is the former, then heterogeneous integration does not really solve the problem since you are still processing III-V devices and are likely to need multiple fabs to do it. If it is the latter then you should stick to III-V wafers.
The fact is that silicon provides passive performance that is far superior to III-V while the active performance – the detector and modulator - is good enough. In fact our germanium detectors could be better and our electro-absorption modulators can be lower power and exhibit a broader working spectral range.
We have seen repeatedly that being good enough is all that silicon has to show it can do to win and that it is certainly doing.
"It is not enough to offer a 10%, 20% or even a 50% cost saving when you are offering the customer a brand new solution that comes with all the risks and unknowns associated with that technology."
Kotura has developed components for telecom (variable optical attenuators, and the functions needed for a 100 Gig coherent receiver) yet its focus is on datacom. Why is that?
We started in telecom as we looked for low hanging fruit that could give us a good margin and an easy start in our early days. This is important for a start-up with a new technology. The well-entrenched incumbent technologies are hard to displace.
You have to find an application with a clear value proposition to get started. Once you have established yourself, your supply chain and manufacturing infrastructure, you can take on more challenging and larger market opportunities.
We see certain areas in datacom that are not well served by either the short reach optics or the telecom grade solutions. Extended reach data centre is one key area where short-reach optics based on VCSELs cannot cover the reach needed and conventional telecom solutions are inherently over-engineered and do not meet the power, cost and size needed.
We think silicon photonics can play a key role here as a starting point in datacom. A key advantage of our platform here is that we can do WDM [wavelength division multiplexing] and hence offer 100 Gig on a single fibre (per direction). This is a major cost saving for longer reaches (>>50m) deployed in such links.
There are some big system players with silicon photonics (Cisco Systems, Alcatel-Lucent) and several small merchant silicon photonics players, such as the companies mentioned in the previous question, which must develop products to sell while funding the development of their technologies. How do you expect the silicon photonics marketplace to evolve, especially now that the technology is being more widely embraced?
For silicon photonics to succeed commercially, we need a multitude of vibrant and successful players in the field. Some of these can be start-ups that lead the innovation in technology and manufacturing but others can be larger organisations that have invested to service an internal need or leverage an existing dominance in the market.
There is room and a necessity for both. It takes a village to raise a child. One single company will not turn silicon photonics into a successful commercial reality.
Cisco Systems has been talking about its proprietary CPAK transceiver. Here is an example of a system vendor using in-house silicon photonics for its own use. Why do you think about such a development? And is Kotura being approached by equipment vendors that want to work with you to develop a custom design?
It is not new for a large company like Cisco to try and make sure that is has its own proprietary components to go into its systems to protect their product and their margins. We see that in all industries.
In terms of other companies coming up with their own proprietary solutions, we do see more and more of this - and a lot more this year than last year - especially when you come off the telecom bandwagon and into the datacom environment: data centres and high-performance computing.
That is because the customer is in charge of the entire environment, the two ends of the link, they can leverage more value from the solution you have to offer without worrying about standards. This is one way for systems companies to leverage value from components.
People are starting to see that the conventional technologies they have deployed are hitting a wall. When they are deploying a new solution they are rethinking their hardware strategy, and how they leverage it to add more value and differentiation to their system.
New ways to architect systems are becoming possible. If you are able to avoid limitations such as distance between the processor and memory, the router and switches and so on, you can come up with a very different architecture for your system and solution.
When do the big players adopt a new technology and go from an electrical to an optical solution? In my experience, usually when they absolutely have to. Most people don't adopt a new solution until they really, really need to; when the value proposition completely outweighs the risk.
It is not enough to offer a 10%, 20% or even a 50% cost saving when you are offering the customer a brand new solution that comes with all the risks and unknowns associated with that technology.
You have to offer them something new, to enable a new application, to add value by enabling a feature, something they can leverage in their product.
When you say systems people adopt new technology when they hit a wall, can you highlight examples of these hurdles?
When you look at the adoption of optics coming from the copper-dominated connectivity, it is very interesting.
Originally, for optics to work its way into the copper world, it had to hide itself and look like a copper solution. People had no idea how to create connectors and they were worried about fibre. So it was disguised as a copper solution.
As customers have got used to it, we can now come out and be more open. We can now do more innovative things with optical transceivers. If you look at the adoption rate, it is being accelerated by customers' demand such as 25 Gigabit signalling.
We can see that the processors and the ASICs - a switch from Broadcom or a processor from Intel or AMD - they are running into I/O [input/ output] density bottlenecks. The chip area is pretty constant, the packages are about the standard size, the number of pins are going beyond what they can support, they have to ramp up the pin rate to about 25 Gigabit-per-second (Gbps), while there are also some 10Gbps pins.
But the number of 25Gbps pins are becoming so high, potentially many hundreds, that they are not going to be able to trace them into the PCB (printed circuit board). The PCB can only take a 25Gbps signal for about 4 inches (~10cm) and then you need serdes [serialiser/ deserialiser) and repeaters.
You may imagine a current router or switch ASIC having ten 25Gbps pins and 100 10Gbps pins. The 10Gbps pins I can take to the edge and use 10Gbps transceivers; and the ten 25Gbps pins I can still do something about it. I may need a lot of electronics and serdes, and use pre-emphasis and equalisation.
But the next generation, when it becomes 100 25Gbps pins, you just cannot do that at the board level. That is where we will start to have to use optics close to the chip.
Will they go for very compact transceivers that sit next to the ASIC or would they try and co-package it with the ASIC?
My perception is that the first generation will be next to the ASIC. People will not integrate an unknown technology into a multi-billion dollar business, they will hedge their bets and have an external solution that offers them some level of assurance that if one solution does not work, they can change to another. But once they get used to it, they can start to integrate these in a multi-chip module solution.
What are the timescales?
I see transceivers next to the ASICs being deployed around 2017-18, maybe a bit sooner, with the co-packaging around 2018-20. People are already talking about it but usually these things take longer.
For part 1 of the Q&A, click here
Further reading:
Reader Comments