techjeet

I kinda wish I was born in the 1980s so I could figure out some of the early CMOS circuits as a grad student

@hackradios @insane_analyst Fiber aside, the fiber coupler and the microLEDs lens add significant crosstalk

@techjeet5 @insane_analyst The way these LED links usually work is by dedicating a fiber per lane, so there's no optical xtalk. On the other hand you can quickly see why this approach would be a packaging and reliability nightmare...

DSPs gonna have a 1000000x1000000 crosstalk cancellation pre-FEC lmao

me when I eliminate serdes but need to serialize and deserialize

@SinghJyotirmai Hilarious that all of this was just meant to be an ace in the hole against the Ottomans

Lmao

@photonspotting 😭😭

@techjeet5 Plus Resonac and Resona Semi, what idiots honestly

Whoever named Semtech and Samtec and Santec can't keep getting away with this

@insane_analyst Lmfao. I do understand their argument but I just want to know one singular instance where people chose AEC over optics.

Bro have mercy on Credo bulls lmao

Does anyone know of any instance of AEC copper replacing optics?

Cool thread on my family history

Actively looking for a job in high-speed analog IC (preferably in SERDES/optics). If anyone has any recommendations, I’d be happy to chat

1. All CPOs use external laser sources (for the exact reliability and thermal control problems you brought up). No serious modulator-based CPO vendor has proposed otherwise. 2. FEC/DSP is needed for scale out environments where fiber length and coupling losses are significantly larger. Scale up inter-rack solutions are where CPO can be really effective. Even if couplers add loss, the latency benefit is still good enough to justify increasing laser power 3. By no means is 15% “relatively small gain”. If you can squeeze any power/size improvements out of PHY, you’ll take it

Copackaged optics is not reaching widespread deployment in data centers for the next 10 years. I don’t care what the roadmaps says !!! Everyone who is shilling this is either lying to you or doesn’t actually work in a data center. LASERS FAIL ALL THE TIME. That’s LITERALLY facts ask anyone who is running large data centers they need to swap pluggable transceivers DAILY. When your laser dies in a CPO system, your switch ASIC is a brick and is going to take down all the fucking traffic! Not just some network performance degradation it’s going to turn into a fucking brick you need to buy a new one and it costs like 300k dollars lmaooooo. And the “efficiency” argument many people will say is not as good as it looks on paper. Your optical engine still needs DSP chips for FEC and equalization. The lasers need thermoelectric coolers sitting millimeters away from an ASIC dumping 300+ watts of heat. The TEC power draw eats the efficiency gains that people like to tout in their pitch decks. Not to mention PHY’s are like 15% at most of data center power draw? You’re taking that big of a risk for a relatively small gain. Then there are coupling losses that are locked in at assembly forever... which just makes manufacturing this reliably at scale a nightmare. Bro not every PHY is the same based on how you placed the fiber like give me a breaaaaaaak. Ask anyone who actually has a PhD in photonics what they think they are going to tell you the same shit. Everyone telling you different than me needs to financially benefit from their investments. I don’t!

Excellent intro to phase rotators rfinsights.com/synthesizer/ph…

@techjeet5 who cares just do software bro

Broadcom just dropped 400G/lane optical DSP broadcom.com/company/news/p…

This plot of $GFS monolithic (fusion) sipho+45nm elec node is all thou need to know. Wonder how bad return loss is. This is why Ayar Labs and Lightmatter are fleeing GFS in favor of TSMC COUOP (hyb bond). $TSEM also has hyb bond offering. Only GloFo does monolithic approach.

400G/lane technology has arrived! Scale AI workloads with the first 400G/lane optical DSP: Taurus BCM83640. Debuting at #OFC2026, it enables low-power 1.6T transceivers and paves the path for 204.8T networks. Learn more: brcm.tech/taurus

@SemiAnalysis_ Yeah pretty much lmao

Expectation vs Reality: Life as a SerDes engineer

@bubbleboi No, this is PAM4

400G per lane? Fuck are they using PAM9?

@jwt0625 Deming Chen the goat

what a diverse group of authors

@Willschips Nvm misunderstanding on my part

@techjeet5 Define ChipFoundry tools