Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
It literally says in the acknowledgments "Polariton Technologies acknowledges NLM Photonics for providing the organic electro-optic material."
waiting with bated breath for the publication of DoA's epic saga "The Vulcan and the Used Car Greaser - the Trials and Tribulations of Darkwave Tragic" - gonna be a smash hit I'm sure
yeah you're not wrong at all - drive voltage is generally proportional to Vpi, but to compare apples to apples it's important to note what's being quoted. For a given drive voltage, Vpi could be e.g. 3-4 times higher. Specsmanship is a popular discipline even in the scientific community...
Good die selection needs to be done in any case for any technology, that's what wafer-level testers are for. What the yield will be like is anyone's guess I suppose.
Nah, generally very different applications.
Don't confuse Vpi with Vpp (drive swing). To achieve PAM4 target ER, you need to drive only a fraction of Vpi.
On the LPO point: If EO modulators turn out to be true enablers for LPO (over Si and InP), then that could be the real intercept at 200g, but support for LPO has been flagging over concerns of interoperability. For AI clusters that's largely a don't care though. In fact Nvidia kind of kickstarted the whole thing.
I'm pretty sure it was a regular MZM; the drive voltage is on the high side, so there is certainly room for improvement on power. As the DSP is the biggest hog, moving to the next CMOS node is the obvious next step. Removing the driver could save a few Watts, but the problem is the timeline - this stuff will start to be rolled out in the 2nd half of this year. There's room for iterations but the end users don't have a lot of appetite for technology risk. As long as the optics and switch fit in the shelf-level power envelope it won't be a hard requirement.
No, the point is that the big prize is at 400Gbps per channel, not 200Gbps.
The Intel demo was 100GBd PAM4, not coherent. The Marvell and Acacia products are coherent, 800Gb/s per wavelength. Marvell also has PAM4 products, mainly DSP. The point is that those modulators have similar bandwidth, despite the different type of modulation.
No, obviously incorrect. This demo was on the show floor for everyone to see. Marvell and Acacia have silicon photonics modulators (in the coherent space) at 120-130GBd that are starting to ship in volume.
Just one comment on the role of the foundry: No foundry guarantees chip-level yield. In principle, they don't even know what they are making (in terms of functionality, performance), what it's for, how to test it, or what the selection criteria are. Selecting KGDs (hence yield) is very much the foundry's customer's responsibility.
What the foundry guarantees are that their process adheres to their stated specifications and tolerances as long as the customer adheres to the design rules. The customer needs to take those rules and specs into account in the design of their chip, such that the design is robust with respect to the process variations, in order to obtain an acceptable yield at chip level.
Intel demo'd 200Gbps/channel pure silicon photonics modulators at OFC. We discussed that here. In addition, as they have the capability of 300mm-wafer-level integration of InP, they could switch to EAMs if needed.
Of course Nvidia is working on such 800G transceivers, they have been shipping tons of them already and are now #2 after InnoLight in that category, according to Cignal AI. Don't forget that in these short reaches (sub-100m) multi-mode VCSEL-based optics still dominate.
Nokia is primarily in the telco / service provider space, I don't think they can compete in this short-reach space.
In this industry, NDAs are intended to protect the exchange of Confidential Information, i.e., information that is considered confidential/proprietary by the party owning that information. They are usually mutual in nature, so that Confidential Information can flow both ways. Usually the initial intent is to exchange primarily technical information to assess whether a further, deeper collaboration makes sense, and they typically precede any business agreements.
The existence of such an NDA is not Confidential Information that is proprietary to one of the parties exclusively (there is no Disclosure of said information from one party to the other), and as such is not covered by the NDA itself.
These NDAs typically also include wording to indicate that the NDA in and of itself is not to be construed as a business partnership or business association between the parties.
However, when a business association does come into existence, there may well be terms that prevent the disclosure of said association to third parties, but then either the NDA needs to be amended or superseded by terms and conditions of the business contract.
It's hard to view that change as anything else than transparent pandering to the AI hype. First, it's a red herring because as far as the modulators are concerned it's a don't care. Second, the Ethernet market is much larger than IB. Third, even Nvidia itself is helping dig IB's grave with their Spectrum X Ethernet switches. They have a notable Ethernet switch line up in addition to IB from the MLNX acquisition.
Easily the most interesting part of the presentation though! clearly 4 channels.
Has that design been taped out? When are first wafers expected to fab-out? Timeline for first test results?
Yeah, sloppy wording... it's a transceiver PIC reference design, not a transceiver module reference design. AMF doesn't do modules.
Would need to look it up, but I think that 2.5B was referring to module sales, while this is PIC sales. LC does a better job than most, of course there's assumptions, projection and extrapolation, but they certainly do not throw darts at the wall.
Latest from LightCounting on market shares of GaAs / InP / SiPh / TFLN et al. in transceivers:
Sales of Silicon Photonics chips will reach $3 billion by 2029
Ha, great question! The wafers are round because they are cut from a solid silicon cylinder called an "ingot". The ingot is cylindrical because it is created through a rotational "pulling" process from sort of a crucible of molten silicon. There is always some loss of silicon wafer area because it isn't possible to completely tile the wafer with the rectangular dies. As the starting wafer is not that expensive, it's overall not a big deal.
Possibly, but that would come with strings attached. Hence my earlier comment about customer NRE, which could be put towards not only product but also process development (but without strings attached on the process side).
When that happens we're all monumentally fucked
Silicon photonics will not be a concern (there's good capacity in the US, EU and other parts of Asia), but for CMOS that's a huge exposure, which is why TSMC is busy building fabs outside Taiwan (incl. a $40B investment in Arizona if I'm not mistaken).
The main point I was trying to make is that because - other than Intel - none of the major chip/module/equipment vendors in comms have their own fab, acquiring a company such as LWLG is not a natural fit, but they could work with their respective foundry partners instead to ensure access to the tech.
In the indium phosphide space the situation is a bit different though, think Infinera, Lumentum, Coherent...
Alright 1 final thought for the day - I think the choice to partner with a smaller specialised foundry is the right one. Yes they are small but they're a credible player. With such a partner at least you get the time of day. Show that your stuff works in one foundry, port it to another one as a second source and then for these communications markets you're set volume-wise.
Please put to bed this notion of 5/6/7 foundries - it would be completely unnecessary and an utter waste of resources. All this process development is on LWLG's dime at the moment and does not come cheap.
Technically speaking it's the patent office's job to verify that any new patent application is not already covered by an already granted patent and reject or require amendments to applications that do not pass muster. Clearly this process is not perfect and in fact patents can be challenged even after having been issued. The more nefarious case is when somebody is simply infringing without given any indication of doing so - it's then entirely up to the infringed patent's owner to demonstrate that. This is often not practicable.
Genuinely have no idea but "deal" is a pretty broad term - one thing that's puzzling is why so far there never seems to have been any NRE-based revenue. Does anyone have any insight into that? Was that a strategic choice to not pursue? An agreement to (co-) develop a chip counts as a deal, even if there's no upfront commitment to take it into production.
Takeover - the big chipmaking companies are all fabless, with the notable exception of Intel, who is now clearly separating out their foundry business as well. The big transceiver assembly houses are also fabless and generally just buy the components or develop based on foundry offerings. An acquiring party would probably be either a SiPho foundry, one of the more vertically integrated players in communications who would know what to do what a new material, or indeed an actual materials company.
No, that's not at all what that means, how do you get to that conclusion?
First, O-band is used predominantly for the shorter reaches uo to 10km (mainly 500m / 2km PSM and CWDM). For longer distances C band is more suitable because of the lower attenuation.
Second, all competing technologies also work in the O band (with the possible exception of VCSELs although there are R&D results in that area).
No worries! Should be no problem to port to O-band, some components need to be adapted, but from the underlying physics the key Vpi*L metric should inherently be somewhat better (lower) at the shorter wavelength.
can't quite find the post but I recall someone asked about having a "moat" based on poling-related patents. Tricky one - patent enforcement requires detectability of infringement. While it may be possible to detect that a wafer or PIC has a polymer on it and that the polymer has been poled, it'll be hard to detect unequivocally how the poling was done. It's a balancing act between protecting IP through patents versus keeping it completely hidden as trade secrets.
If you can hit the required average optical transmit power, modulation amplitude, and eye quality as per 802.3dj with sub 1V peak-to-peak drive amplitude, then that is in the cards, yes.
The test setup diagrams from recent talks showed a C-band laser and indeed an EDFA (erbium-doped fiber amplifier). That's pretty typical and not a reason for concern as these bench top setups generally have excess loss that in the final system you would not have (e.g. because light is coupled in and out through grating couplers instead of edge couplers, and there are several optical power splitters). The amplifier boosts the power level but does actually add some noise as well.
They'll need to confirm these results in the O-band. Also the current modulators appear to be single-ended (GSG electrode configuration). A differential version (GSSG or GSGSG) would be a big step forward because that's what most implementations use as of now. Plus of course the inclusion of a low-loss fiber coupling interface. Wirebond should still be OK but some customers may ask for a flip-chip solution.
Assuming said PIC meets the specs (optical, electrical, mechanical), the customer could integrate it in a prototype 800Gbps transceiver module, along with all the other componentry (DSP, laser, driver, TIA, PDs, microcontroller) and run test patterns or even live traffic over it. That's what you really want to get to.
What KCC is referring to is single dies with multiple modulators on them, i.e. a 4-channel modulator implemented on a single PIC. That's what the product should look like and what the customers need to build transceiver prototypes.
As the stated objective is a 4x200G PIC for 800G/1600G IMDD pluggable transceivers, it has to be O-band. This is not the same market as AOCs by the way, which are usually based on VCSELs (850nm).
C-band is for longer distances, usually coherent + WDM. No inherent reason this market could not be serviced, but it's a quite different PIC architecture implementing dual polarisation IQ modulation. There has been some talk about coherent O-band communications for shorter reaches, but nothing too concrete so far.
This is one gap that requires clarification, as it appears from the recent presentations that the device experiments they showed were conducted in the C-band.
Right now, LWLG is a customer of AMF, not the other way around. Presumably LWLG owns IP to certain process steps specific to addition of the polymer, poling, and modulator design. It doesn't seem to make sense to sell the polymer to AMF only to buy it back at a markup as long as all those wafers go back to LWLG.
This changes of course when this LWLG-enabled process becomes available to third parties through AMF, in which case LWLG can start collecting royalties.
Through packaged modulator sales by Polariton, that is likely the case.
OpenLight is based on the Juniper / Aurrion heritage of integrating 3-5 on Si, so this all makes sense. The fab is Tower, but the process is reportedly different from Tower's open foundry offering. Best to partner directly with Tower as an add-on to the standard foundry offering, IMO.
Of all the silicon photonics platforms, this is among the hardest to integrate with because it comes with a CMOS stack on top.
No, this is a silicon-only solution. I'm sure Marvell (like many others) is evaluating alternative modulator technologies though but for the next speed (400Gbps per channel).
I'm reading that as the list of parties that visited the OFC demo.