Intel And AMD: One-On-One In Servers
The third generation of Intel Xeon Scalable “Ice Lake” processors pulled up the gap with AMD EPYC: support for 64 PCIe Gen 4 lanes (was 48 x PCIe Gen 3) and eight DDR4-3200 memory channels (was 6 x DDR4-2933). The list of AMD EPYC processors is three times shorter than the Intel Xeon SP 3rd Gen line with its variety of models (50+), but they support 128 PCIe Gen 4 lanes, and the core ceiling is still higher (64 versus 40 for Xeon).
One Socket Versus Two
The more cores a processor has, the less they need for multiprocessor platforms. If one CPU covers the need for computing power, why chase two? Single-socket platforms are more straightforward and cheaper than two-socket platforms. It is convenient to build a scale-out infrastructure on them. 128 AMD EPYC lanes are enough for a graphics server with four x16 PCIe GPUs and a storage server with 24 x U.2 SSD x4 PCIe.
The AMD EPYC P-series single-socket processors have become a powerful driver for developing hyper-converged systems, GPU computing, and NVMe storage. AMD has not only pushed Intel into existing markets but created a new trend for the single-socket future of servers. Intel followed AMD’s lead two years ago and released its answer, the U-series Xeon Gold.
“single-socket” processors do not require additional production efforts – they are obtained from their older brothers by disabling the interface responsible for combining several CPUs in one system. They sell P- and U-processors at reduced prices – expecting that the expansion of the sales matrix will take away the “lost” profit.
Server CAPEX includes the cost of software licenses, which are usually tied to sockets and processor cores.
VMware calculator reminds that the license covers up to 32 cores of one CPU, and VMware vSphere Enterprize is licensed per socket. It turns out that the arrangement of infrastructure on single-socket servers with 32-core processors is much cheaper in terms of license fees than on two-socket servers with 16-core processors. Leading VMware bloggers acknowledge AMD’s contribution to the rise in popularity of single-socket servers.
Windows Server 2016 Standard base license covers 16 physical cores. You have to pay extra for more cores. SQL Server is licensed in the core. This stimulates not to produce multi-socket / multi-core but to pay more attention to objective estimates of the required resources: to take processors with a higher clock speed, to unload disk and network I / O.
One Socket, 32 Cores: Which To Choose From?
The second-generation Intel Xeon did not have 32-core single-socket CPUs. Now it is. VMware users will be interested in comparing 32-core Intel Xeon Gold 6314U and AMD EPYC 7543P:
The processors are comparable in price. AMD’s clock speed is 20% higher, which gives a noticeable advantage in practical applications. They have an equal share of DDR4-3200 memory channels. Both processors support PCIe Gen4, but AMD with 128 lanes leaves more freedom for the periphery. Intel Xeon understands the Optane PMem 200 persistent memory (with a brutal fate that never became in demand).
Companions And Fellow Travelers
The support group for single-socket servers based on AMD EPYC has formed, matured, and includes leading manufacturers over three generations of processors. The user has plenty to choose from: in addition to Dell PowerEdge R6515, HPE ProLiant DL325, or Lenovo ThinkSystem SR635 Rack, there are offers from ASUS, Gigabyte, Supermicro, Tyan.
It is worse with single-socket servers on Intel Xeon. Major League disregarded the U-series and the second and third-generation Xeon Scalable, offering only dual-processor systems. Of course, they work with one CPU, but “involuntarily one-legged” servers are not an attractive solution. The offerings of single-socket platforms for Intel Xeon from second-tier manufacturers are not enough to squeeze out a competitor.
Let’s assume that Intel has extended the subscription to seats in the “one-socket box,” secured a formal presence in the still foreign territory, and will play on the loyalty of consumers in anticipation of the next station on the shared server path.