Silicon nitride: the hard part isn't how low you get the loss

Yesterday someone in the group forwarded TSMC's SiN-on-SOI work from OFC 2025, and I clicked through it. They use low-temperature PECVD to deposit silicon nitride, so as not to bake the adjacent silicon active devices. The cost is stated honestly: low-temp PECVD film quality and uniformity aren't as good as high-temp LPCVD, and you need another round of process optimization to make up for it.

That one sentence is worth more than all the device data in the paper. The silicon nitride waveguide road has been walked for forty-odd years — guiding light by the late '70s, propagation loss in the telecom band pushed below 0.3 dB/cm by the late '80s. At the spec level it stopped being a problem long ago. The real watershed has always been how wide the process window is, and whether it's repeatable.

When I do side-by-side evaluations for clients, the first question I ask a process platform vendor is what the yield curve looks like, how mature the PDK is, how many tape-out batches they've run — loss comes later. Companies doing silicon nitride processes, like Guoke Guangxin, are selling exactly this. A platform with pretty parameters that needs an engineer babysitting every batch won't get clients to pay.

One-line judgment: the commercial value of this line isn't the best value in the lab, it's the variance on the production line. Whoever drives down the variance gets the orders.