Refresh the Old to Free Up Power, Cold Plate, and Immersion Cooling, and AI takes Center Stage
OCP 2023 in Santa Clara was a great show.
Let’s start with the packed keynotes, which all discussed what we will do to solve the power problem, except Google. We will need more efficient silicon designs and new cooling technologies to lower the amount of power. Cold Plate and immersion were showcased, and many other vendors discussed technologies, such as a diamond substrate.
We need to do more DCI to reach smaller and more remote locations that do have power and need to refresh the massive hyperscaler installed base with new servers that are more power efficient. All the vendor and hyperscaler keynotes highlight the need for more CAPEX. Power will be an issue in the future, and we need to get past this wall, or we can’t grow compute as quickly as the industry would like or need to.
Where do these developments leave us going forward?
There will be a significant investment in the server’s disaggregation, new server cooling approaches, and more distributed data centers. This is good for Marvell and will drive robust adoption of ZR/ZR+ modules in the current 400G and future 800G generation.
AMD highlighted the need to upgrade the existing infrastructure with next-generation processors to free up space and power. We agree. The question is timing and how big of a pull-in we get. We also expect regional preferences to change. Some T1 locations may become more AI-focused and push less latency-sensitive workloads to remote locations.
While there were fewer new equipment racks compared to previous years, we enjoyed the Meta rack and the Credo racks in the OCP experience center, which showed past, present, and potential future rack topologies. A ton of Ethernet ports and some CXL and CXL switching as vendors look to disaggregate.
We already look forward to OCP in 2024!