After earlier missteps in executing its server CPU strategy, Intel has been working aggressively to reestablish its credibility in the market. Competitive pressures from AMD, along with a surging Arm in the cloud space, have turned the company from being the only significant datacenter player to having to find its footing.
A couple of years back, CEO Pat Gelsinger boldly committed to delivering new technologies to the market quickly. Specifically, he committed Intel to advancing by five manufacturing nodes in four years. For those unfamiliar, a node typically takes about two years to achieve, so Gelsinger’s strategy is extremely aggressive. At the Hot Chips 2023 event that just ended, Intel unveiled its plan to deliver two new server CPUs in 2024. I’ll detail Intel’s plans below and offer my own thoughts on them.
Different cores for different needs
There are three classes of servers in the server market. At one end is the single-socket/dual-socket server designed for scale-out, powering many distributed workloads across the cloudified datacenter. These servers are about core density—squeezing as many cores into a rack unit as possible to handle a massive number of lighter workloads.
In the middle is the traditional two-socket server that has powered the enterprise’s virtualized workloads for generations. HPE’s ProLiant DL380, Dell’s PowerEdge R760 and Lenovo’s ThinkSystem SR650 are all examples of Intel-based two-socket servers that populate the racks in server rooms and data centers around the world. These servers strike the balance between power efficiency and performance.
At the far end of the continuum are the performance workhorses. These dual and four-plus socket servers are about delivering the most performance per rack unit to enable workloads with rich compute support. Workloads in areas such as ERP, analytics, high-performance computing (HPC) and AI will utilize every bit of computing resources available.
Each one of these server types is critical to the enterprise and cloud datacenter. However, the design of the CPU cores that support these servers is different. For example, scale-out servers want as many cores as possible, but the cores do not need as much performance, nor does the supporting microarchitecture. Meanwhile, the big caches and acceleration capabilities required for high-performance workloads are unnecessary for distributed application support.
Because of these disparate workload needs, Intel announced the release of two new cores in 2024—Granite Rapids, focused on performance workloads (P-Core), and Sierra Forest, focused on scale-out and efficiency (E-Core).
Forbes Daily: Get our best stories, exclusive reporting and essential analysis of the day’s news in your inbox every weekday.
Both cores are built on the same platform, so there are similarities between the two—the same socket, memory and firmware. This should enable server manufacturers to more easily design motherboards to accommodate them. And, of course, both cores are built with the same Intel 3 process, which should lead to considerable performance lifts and power savings over the current fourth-generation Xeon processors (code-named Sapphire Rapids).
Overall, the details on what we can expect from a speeds-and-feeds perspective are a little scant. So far, Intel has revealed it will support up to eight sockets, up to 12 channels of memory bandwidth, I/O up to 136 lanes of PCIe 5.0/CXL 2.0 and up to six UPI links for CPU interconnect. But the details stop there.
While I’d like to see greater detail, I get what Intel is doing by playing its cards close to the vest. Sierra Forest is scheduled to launch in the first half of 2024, and Granite Rapids will follow. Intel did reveal some time back that Sierra Forest would support up to 144 cores per socket, but all we know for Granite Rapids it that it will support more than the 60 cores per socket supported in Sapphire Rapids. I’m certain Intel is sharing more details around early specs and performance numbers with its customers under non-disclosure agreements. Still, there is no incentive to tip its hand to the competition.
While Sierra Forest is designed for getting the highest number of power-efficient cores on a socket, Granite Rapids is designed for performance. As such, Sierra Forest will be available in single-socket or dual-socket configurations, while Granite Rapids will be available in up to eight-socket configurations.
If you’re wondering why up to eight sockets, the answer is simple: the market wants it. Many enterprise applications still require these higher socket configurations to support things like resource planning, analytics and some HPC workloads. And Xeon is the only game in town for that, given that AMD focused squarely on the single- and dual-socket server market.
Sierra Forest is designed to scale from the enterprise to the cloud
With Sierra Forest, Intel has designed a CPU that can scale from supporting virtualized workloads in typical enterprise datacenters to the racks that populate the datacenters of the hyperscalers. By stripping out some of the performance-focused architectural elements (i.e., Matrix Engine) built into its performance core, the company can focus on what’s important—core density and power.
As Sierra Forest boasts a significant uplift over Sapphire Rapids, what we will see deployed in the enterprise datacenter is more likely in the 32-core per socket range. These 144-core chips are targeted at the hyperscale deployments.
Granite Rapids: AI, supercomputing, analytics and more
There is no question about where Intel is targeting Granite Rapids: compute-intensive workloads (ahem, AI). While the value prop for Sierra Forest is a little nondescript, Intel has been a bit more aggressive with Granite Rapids. Intel positions itself as already owning the performance crown in AI with Sapphire Rapids, and only creating more separation with Granite Rapids.
While compute-intense workloads often get bucketed into one category, in fact their performance characteristics can vary significantly. Some workloads require a lot of memory throughput, while others require the fastest possible cores. For still others, it’s all about accelerating performance through CPU enablement or discrete add-ons such as GPUs.
What Intel is doing with Granite Rapids makes It well-positioned to meet the needs of these different types of workloads. Memory bandwidth, memory capacity, high-performing cores, lots of I/O to support accelerators and accelerator engines are all critical to driving the diverse set of requirements for the enterprise high performance workload. And Intel has detailed improvements in all of these.
Intel needs to achieve a couple of key things in 2024 to set itself up for longer-term success in the server market. First, it has to deliver on the story. Two cores on the same compute platform that handle the range of workloads being deployed across the enterprise and the cloud? That’s a compelling prospect. But the cores must perform as advertised, and the numbers must impress.
Second, the company must deliver Sierra Forest and Granite Rapids on schedule. The recent challenges experienced with delivering Sapphire Rapids put a blemish on the Intel’s record as an execution engine. That has hurt its reputation, especially as its chief competitor has continuously hit its schedules and milestones.
CEO Pat Gelsinger and his executive vice president and general manager of data center and AI, Sandra Rivera, have put the train back on the tracks, and it looks like 2024 will be reflective of the renewed discipline and focus at the company. But of course, we’ll have to wait and see.