Datacenter servers will face a double whammy this year as CPU supply constraints pile on top of an already severe memory shortage. Even so, shipments are still expected to grow at a double-digit rate.
The latest Cloud and Datacenter Market Snapshot [PDF] from analyst firm Omdia says that server CPUs are also experiencing supply constraints that are likely to see prices rise, with a knock-on effect on overall system costs.
Omdia claims this is at least partly due to the difficulty of shifting production volume across different process nodes. As CPU roadmaps have broadened, vendors are simultaneously manufacturing processors on different process nodes such as 3 nm or 5 nm.
“The nature of processor design and fabrication is such that shifting volume across process nodes is complex and time consuming,” the firm says, adding that lower-than-expected yields may have also contributed to the shortage.
The upshot is that Omdia believes server CPU prices could increase by 11 to 15 percent as a consequence of the shortfall in supply. And this is a conservative estimate, based on the fact that the biggest customers – the hyperscale cloud operators – will typically have fixed their prices beforehand through long-term agreements with vendors.
Any CPU issues will add to an already challenging IT supply chain situation driven largely by the memory shortage. The cost of DRAM is expected to nearly double this quarter, as The Register reported just days ago, while NAND flash used in solid-state storage is also forecast to rise by more than 30 percent.
DRAM is in particularly short supply because of high demand and because the memory chipmakers have repurposed manufacturing capacity to focus on making more of the high-margin products such as high-bandwidth memory (HBM) used in GPU accelerators for AI processing.
Omdia says that it is concerned that memory availability could disrupt server manufacturing output, which might in turn affect the completion dates for ongoing datacenter projects.
Despite this, the analyst firm still foresees a 12 percent growth in server shipments this year, thanks to a refresh cycle covering general-purpose servers that is also boosting shipments. It regards memory availability as much more of a risk than CPU availability.
Meanwhile, when it comes to racks packed with AI servers and GPUs, Omdia estimates that at least 71,000 racks will ship this year with more than 100 kW of IT load, largely driven by Nvidia’s NVL72 integrated system.
Thanks to the seemingly insatiable demand for AI infrastructure, Omdia forecasts that this will increase by 56 percent next year, and we will see the first ultra-dense racks with upwards of 200 kW of equipment.
But it isn’t likely to end there. Last year, Google detailed plans for future 1 MW IT racks at the Open Compute Project (OCP) Summit in Dublin. ®