ST Micro Skips In, Arm In Arm With AWS, Bearing A Chip For 1.6 Tbps Pluggable Optics

Developed in collaboration with Amazon Web Services, ST Micro detailed a new photonic integrated circuit (PIC) on Thursday that it says will support pluggable optics capable of shuttling bits around the datacenter at up to 1.6 Tbps.

The chips, which are slated to ramp up in production later this year at ST's fab in Crolles, France, aim to address the growing bandwidth requirements of AI clusters, which often feature large scale-out compute fabrics linking thousands of GPUs together.

According to ST, the PIC100 will be fabricated using its BiCMOS process technology – such as the upcoming 55nm BiCMOS B55X node – and is designed to efficiently convert electrical signals to optical ones and back again.

At launch, the fab says the chips will be capable of carrying 200 Gbps of bandwidth per lane — that's twice what existing pluggable optics can do today, as far as we're aware. ST says its customers, who take the chips and put them in products you and I can use, are already working on 800 Gbps and 1.6 Tbps pluggables based on the designs.

In addition to higher aggregate bandwidth, the chipmaker also claims the chip is efficient and runs cooler than conventional pluggables.

Heat, ST explains, is problematic for these devices as it can lead to performance degradation or increased failure rates, if temps aren't kept in check. To mitigate these challenges, the chipmaker has implemented special waveguides and fiber couplers designed to minimize losses and reduce operating power.

Along with beating the heat, or so it hopes, ST is also working on designs capable of up to 400 Gbps per lane, which could eventually pave the way for 3.2 Tbps optics.

Meanwhile, the biz reckons this technology can be used for optical interconnects for GPUs — though we suspect the optics would need to get a lot faster to compete with existing die-integrated photonic interconnects from the likes of Ayar Labs, Intel, Broadcom and others.

• A big AI build has 'stalled' and won't happen this year as funds and GPUs prove elusive
• HPE says blocking Juniper buy is a sure Huawei to ensure China and Cisco thrive
• Openreach tests 50 Gbps broadband – don't expect it anytime soon
• AI ambition is pushing copper to its breaking point

Before you get too excited about the prospect of 1.6 Tbps datacenter optics, it may be a while before we have networking equipment that can take full advantage of them. One of the challenges is that 200G serializer-deserializers, aka SerDes, are still relatively new, and as for early 2025, we're not aware of any switches or ASICs you can buy today that make use of them.

But, even when these next-gen switches do arrive, you can expect their 1.6 Tbps ports to primarily be used for aggregation links with other switches in the datacenter. 

Servers won't be getting ports anywhere near that fast for a long while longer. In fact, we're only now seeing the first-generation 800-Gbps-capable NICs hit the market with Nvidia's Connect-X8, which we'll note is equipped with 3x the number of PCIe lanes you'd see on a typical network adapter.

Cards with a standard x16 slot are going to be limited to 400 Gbps of bandwidth until Intel, AMD, or one of Arm's CPU partners ships a PCIe 6.0-compatible processor. And even then, your server will likely need to be more than a rack width or two away from the switch for optics to make sense over active copper cabling. ®