推文

@carrioresearch · 2026-10-15 15:16

$META $LITE $COHR $AAOI 🚨 META: CO-PACKAGED OPTICS CUT LINK INTERRUPTIONS 6X VS PLUGGABLES SAN JOSE - Two Meta ($META) optical engineers used a session at the OCP Global Summit (Oct. 12-15) to close out an 18-month evaluation of 51.2T co-packaged optics (CPO) switches, where the optics sit on the switch package instead of in pluggable modules on the front panel. Meta ran CPO units and fully retimed pluggables side by side in the same lab at the same operating temperature, with the pluggables as the control group, in two phases: the first to find failure modes under stress, the second to build statistics on more mature systems. - More than 1.8 million system hours of CPO data, which Meta put at over 200 years of system operation - Annual link interruption rate: 1.22% for retimed pluggables, 0.47% for CPO in phase 1 (2.6x better), 0.2% in phase 2 (6x better) - 0.04% in phase 2 with one external laser source (ELS) failure mode taken out, 30x better than pluggables - That failure mode was root-caused to a surface-mount component on the laser driver. It drove 85% of CPO failures in phase 1 and 83% in phase 2, and Meta says a design change fixed it and was verified - No unserviceable failures in 1.35 million system hours of phase 2 testing. Optical engine hardware was 0% of phase 2 failures - Transient or unknown interruptions were 37% of pluggable failures and 0% on CPO. DSP firmware was another 36% of the pluggables', and CPO has no DSP to fail - Average bit error rate on CPO ran almost two orders of magnitude worse than on the pluggables, yet CPO interrupted less - At 100G per lane, CPO cut optics power 64% against retimed pluggables, linear pluggable optics (LPO) 45% "Having a very good bit error rate of 10^-9 or 10^-10 doesn't mean that the link is going to be reliable," one of the engineers said. Meta's caveat was just as direct: reliability is "vendor and implementation specific," and one generation from one vendor does not prove every future integrated optics platform will beat pluggables. Meta's slide put the conclusion in one line: "The 51.2T validation proved viability of integrated optics technology for high volume manufacturing." The next target is a 102.4T CPO or NPO product, a 1.5U, 100% liquid-cooled switch with 64 data ports on MMC16 front-panel connectors, 500 m reach to spec and up to 2 km, and 60% lower optics power than retimed pluggables. The laser modules are field-replaceable and built to the OIF ELSFP standard "to enable open ecosystem and multi-sourcing." "Field serviceability with the external laser source alleviates a lot of risks in the serviceability, at least for the first one or two generations. We think that's the right approach," the engineer said. The fiber array units inside are routed at the factory and not touched again in the data center, and the liquid cooling also keeps dust away from the optics. Asked from the stage whether Meta's next step is CPO or near-packaged optics (NPO), the answer was "Both." CPO has the shortest electrical path and scales best from 200G to 400G-plus per lane, but Meta called moving to a single CPO vendor "a huge risk." NPO saves almost as much power and lets Meta mix optics vendors in one footprint, but linear drive gets harder at higher lane rates such as 448G, and adding retimers to fix that erodes the power and cost savings. I put these numbers next to the linear pluggable data Meta's engineers showed at the same summit, which used the same baseline of retimed pluggables at 1.0. LPO came in at 2.0x fewer interruptions in live production from April to August, and CPO at about 6x in the lab. The two setups differ, so the ratios don't stack, but both ways of dropping the DSP beat the module Meta runs today. In LPO the lasers stay inside every module. In CPO they move to a faceplate module that Meta now wants from multiple vendors.

曝光 956 · 评论 0 · 点赞 11 · 书签 2 · 曝光/时 900.3671207007069