Broadcom, Cisco, Inphi and the Timing for Co-Packaged Optics: A CIR Research Note

In the past few months two important announcements have been made showing different ways forward for the Co-Packaged Optics (CPO) business.  One of these announcements came from a duo of Cisco and Inphi. It shows perhaps a certain amount of caution about CPO.  For various reasons, Broadcom is a tad more optimistic.  However, announcements from three such important and prestigious companies show a solid confidence in CPO going forward.  This contrasts with the relative lack of respect for on-board optics (OBO) which after years of trying never really won the respect of the industry. Still, as we suggest below, enthusiasm for CPO needs to based on realism derived from the history of switching chips.

Broadcom:  In mid-January, 2021 Broadcom announced two next-gen switching platforms based on CPO.  The 51.2T “Bailly” switching ASIC is impressive but hardly a surprise.  The conventional wisdom is that CPO will come into its own with 52.T and the release of Bailly in 2023 meets industry expectations for 52.T switches.

From a strategic point of view, what is more interesting is the announcement of Broadcom’s 25.6T Humboldt switching ASIC which is expected to hit the market at the end of next year. Based on what we heard when CIR was doing interviews for our CPO report from late last year , the market for CPO at 25.6T is still a little tenuous.  So, the question arises as to why Broadcom plans to go for CPO so soon and at such a “low” switching speed.

Several possible answers suggest themselves to us.  The most obvious answer to the question above is that some big users (or one big user) – are/is asking for CPO 25.6T ASICs for their next generation gear.  Component and chip makers frequently launch a new line after some guaranteed orders arrive.  In the past this was true of some optical engines for example.  Another explanation is that Broadcom wants to keep ahead technologically of other switching chip makers – Broadcom is now usually considered the one to beat in switching chips and presumably hopes to keep things this way. A 25.6T chip may also enable Broadcom to get baked into switch designs at an early stage and some designers based at OEMs may see the power advantage of CPO as being sufficient reason for its use even at 25.6T

Cisco and Inphi:  Meanwhile, Cisco and Inphi have gotten together to provide a CPO solution starting at 51.2T, with first deployment in 2024.  In the Cisco/Inphi vision 25.6T, switches will not need CPO, but will make use of 800G pluggables instead, but – or so the story goes — by the time the 51.2 generation comes into being, we will need CPO it is claimed

This was the almost universally accepted story that CIR heard when we were interviewing for our CPO report last year.  However, it seems that Cisco/Inphi intends something bigger here – an “open environment” as they call it, which sounds like we are headed down the road to some kind of MSA, bringing in other switching chip firms as well as transceiver firms and OEMs looking to the future.

Skepticism still needed: As noted above, these developments must be considered positive for CPO; COBO never had it so good.  But a bit of healthy skepticism is also needed here.  Anyone who thinks the datacom world will easily withdraw from the pluggable effort is naïve to say the least – pluggables are so convenient after all and OBO and CPO simply don’t offer good alternatives to pluggability.

The ultra-futurists on such matters also forget their history.  The I/O for SerDes was once thought to be impossible at more than 10G and – even more relevant to our story here – at one time some people were saying that 12.8T switching would need CPO or something like it. Obviously, nothing like this scenario happened.  So, you shouldn’t bet your life savings on CPO in switches.  At least not quite yet – it is undeniable that silicon core switching power is increasing at maybe three times SerDes power and the limits of conventional approaches to switching chips will be felt by the industry at some point soon.