Four-Socket Servers: When Two CPUs Are Not Enough

Two sockets covers most of what a general-purpose server does. The reason to go to four is rarely "we need more cores" - it is memory capacity, memory channels, or a workload that has to see all of it as one coherent address space. Framed that way, the four-socket decision becomes easy to make and, more often, easy to reject.

What the fourth socket actually adds

Memory controllers live on the processor, so memory capacity scales with socket count. That is the entire argument. An HPE ProLiant DL580 Gen9 carries ninety-six DIMM slots because the Xeon E7-4800 and E7-8800 v3/v4 processors reach memory through four SMI2 links per socket, each driving a buffer that fronts two DDR4 channels with three DIMMs apiece - eight channels and twenty-four slots per processor, ninety-six across four sockets. Those slots belong to the processors, not to the chassis: each socket owns twenty-four of them, so a DL580 Gen9 with two processors installed is a forty-eight-slot machine until the other two arrive.

The DL580 Gen10 drops the buffered design in favor of direct six-channel Xeon Scalable controllers with twelve slots per socket, forty-eight in total. It reaches the same 6 TB ceiling with 128 GB modules, but only with M-suffix processors - standard first- and second-generation Xeon Scalable parts stop at 768 GB per socket, which caps a non-M chassis at 3 TB no matter how many slots are free. Check the processor SKU before you size the memory.

PCIe lanes scale the same way. A DL580 Gen9 offers nine PCIe 3.0 slots against six on a two-socket DL380 Gen9 of the same generation. If you are attaching several dual-port HBAs, multiple high-speed NICs and accelerators to a single host, the slot count alone can justify the chassis.

What does not scale is storage. A four-socket box is a compute and memory chassis, not a storage chassis - the drive cages are modest, and the design assumes capacity lives on an array or in a separate node.

The numbers side by side

	DL380 Gen9	DL580 Gen9	DL580 Gen10
Sockets	2	4	4
Processor family	Xeon E5-2600 v3 / v4	Xeon E7-4800 / 8800 v3 / v4	Xeon Scalable
Memory channels per socket	4 (direct)	8 (behind 4 SMI2 links)	6 (direct)
DIMM slots	24	96 (24 per processor)	48 (12 per processor)
Maximum memory	3 TB (24 x 128 GB LRDIMM, v4)	6 TB (96 x 64 GB)	6 TB (M-suffix CPUs; 3 TB otherwise)
PCIe 3.0 slots	Up to 6	9	Up to 16
Chassis	2U	4U	4U
Power supplies	Up to 2	Up to 4 x 1500 W	Up to 4

Note what the memory column does. Reaching 3 TB in a two-socket chassis requires the largest modules available in every one of twenty-four slots. Reaching the same 3 TB in a DL580 Gen9 takes 32 GB modules in ninety-six slots - smaller, more common parts, and a service event replaces one of ninety-six rather than one of twenty-four. When memory capacity is the requirement, socket count is how you get there.

Count the interconnect links before you count the cores

Four sockets can be wired two ways, and the processor SKU decides which one you get. The Xeon E7-4800 and E7-8800 v3/v4 families have three QPI links per socket, which produces a fully connected mesh: every socket is one hop from every other, and remote memory latency is uniform.

Xeon Scalable is less uniform. Gold 5100 and 5200 series parts have two UPI links each. Put four of them in a chassis and you get a ring, not a mesh - four of the six socket pairs are adjacent, and the two diagonal pairs communicate through an intermediate socket. Gold 6100 and 6200 series and Platinum 8100 and 8200 series parts have three UPI links and form the fully connected topology. Both configurations boot and both are supported; they do not behave the same under cross-socket traffic. Check the link count on the specific SKU before assuming you have a mesh.

Where four sockets stop paying off

Software that is not NUMA-aware. More sockets means more remote memory. An application that does not pin threads and allocations to a node can run measurably slower on four sockets than on two, because it spends its time crossing interconnect links instead of hitting local DRAM. If your workload has no documented NUMA behavior, test before committing.

Licensing. Windows Server is licensed per physical core with a minimum of eight core licenses per processor and sixteen per server, and most commercial database and hypervisor products count cores or sockets too. A four-socket host with ninety-six cores - four twenty-four-core E7 v4 processors, one hundred ninety-two threads - carries a licensing footprint that frequently outweighs the hardware decision. Do that arithmetic first.

Failure domain. Consolidating a large fraction of your estate onto one chassis means one maintenance window, one firmware update and one hardware fault affect all of it. Two two-socket nodes with the same total core count often serve the business better, even when they serve the workload worse.

Facilities. Four 1500 W supplies expect a 200-240 V feed to deliver rated output, and a fully populated 4U chassis needs a proper rail kit and two people to rack. This is data center or colo territory, not a wiring closet.

The clean case for four sockets is narrow and real: in-memory databases and analytics that must address several terabytes as one space, large-scale virtualization where per-host memory is the binding constraint, and any application whose license is charged per instance rather than per core. Outside that, scale out.

Where to go from here

Decide the memory target first, then the processor SKU and its interconnect link count, then the slot allocation. If those three point at four sockets, the DL580 Gen9 and DL580 Gen10 are the workhorses of that segment and cover most requirements between them.

Once you know the processor model, DIMM population and adapters you need, build the configuration on the Build Your Server page and send it to us for a quote. If you are still comparing a single four-socket host against a pair of two-socket nodes, tell us the workload and the memory footprint in the request - that comparison is worth having before the hardware is chosen, not after.