A busy week in Scotland
ECOC 2023 was a fabulous show with many issues up for debate. Progress was shown in 800 gigabit and 1.6 terabit optics, linear drive pluggable optics (LPO), co-packaged optics and optical input-output (I/O), VCSELs, high-speed coherent optics and DSPs. There was also much discussion about AI, and how the optical industry should progress.
Stepping back, a more fundamental theme evident at ECOC is a debate of serial versus parallel. This is a topic well-known to the optical industry: serial is always the desired approach and when a parallel design is adopted, the aim is to improve the serial to then simplify the parallel design. This, however, is getting harder.
At ECOC, progress in enabling 224G electrical signals was evident but will 112G electrical lanes end up being used for much longer than expected? Coherent optics has reached a baud rate of 200GBd, but how much longer will it increase before multi-wavelength (parallel) transponder designs are used? The answer depends, it seems, on whether the design is pluggable or embedded, and where in the network it is used. How best to scale optical I/O and deal with fiber count is another issue. Parallel channel bandwidth density and fiber management issues also have consequences on the ultimate parallel systems: computing clusters for training large AI models made up of thousands and tens of thousands of GPUs. Optical fiber is also part of this debate. Does it make sense to exploit fiber bands (E, S, C and L) – one form of parallelism - or go straight to parallel core and multi-mode fibers?
Linear drive pluggable optics were prominent at ECOC, continuing the buzz that started at OFC in March.
CIG presented the most extensive set of LPO test data shown so far. The company showed procedures for tuning transmitter settings to compensate for signal distortion along copper traces. Tuning one lane is sufficient to set the parameters of the transceiver. Will tuning of transceivers on a switch board simplify testing of these modules in production? If so, this could further reduce the cost of LPO.
CIG also showed significant variation in LPO performance on switches from two different vendors. Figure below shows test data from a more problematic switch, where tuning was not sufficient to meet the ER spec. Optimizing switch designs for LPO may be the next challenge for switch and ASIC vendors. This will be a challenge with 200G lanes. There is no data yet on the next stage for LPO: 8x200G for 1.6-terabit modules but we should see first results next year.
Transceiver suppliers are enthusiastic about the prospects for LPO, but none of the end users have confirmed plans on deployments of these modules. Nvidia is deploying them in internal AI clusters, but it has not commented on offering systems equipment with LPO to end users. LPO modules should start being used next year given the significant savings in power consumption and we may increase our forecast to be published in December 2023: Market for AOCs/DACs and LPO/CPO.
NewPhotonics has a silicon photonics chip that works with a DSP-based pluggable module or a LPO design. The Israeli startup first came to prominence at OFC 2023 with its chip working with Credo’s Dove 8x106Gb/s PAM4 DSP IC, modulating two electrical channels to send a 224Gb/s optical lane signal. At ECOC, the company demonstrated a fully working chip (transmitter and receiver paths) sending an Intel serdes signal at 224Gb/s over 12km of single-mode fiber without using a DSP. NewPhotonics says its technology can scale up to 3.2Tb/s and support LPO at 200Gb/s and beyond.
NewPhotonics says its technology also enables “Linear Drive plus”, performing gearbox and equalization functions normally requiring a DSP. This is noteworthy as until now there has been a question mark as to whether LPO would only be suitable at 800G (8x100G) and not 1.6Tb/s (8x200G). Now there is at least one way to develop 8x100G/4x200G (800G) modules and 16x100G/8x200G (1.6T) LPO modules.
The startup’s chip uses a time-division-multiplexing approach, modulating the signals and sending them in time slots. NewPhotonics has developed a way to generate parallel channels. At the receiver the signals are unpacked, and some optical signal processing is used to recover the data.
Full text of the research note is available to LightCounting clients at: https://www.lightcounting.com/login