Most servers scale out: you add nodes. Some workloads scale up instead, demanding more cores and far more memory inside a single coherent system than two sockets can provide. The ThinkSystem SR850 V3 is Lenovo's answer for those jobs, a four-socket server in a dense 2U that delivers large core counts and a very large memory footprint without leaving the rack height of an ordinary host. This guide explains where the SR850 V3 earns its place, how to size it, and when consolidating onto one four-socket box beats running several two-socket servers.
What the SR850 V3 is for
The SR850 V3 is a four-socket, 2U scale-up server. The reason to buy four sockets is rarely raw throughput; it is the combination of a very large coherent memory pool and high core density inside one operating-system image. That is exactly what in-memory databases such as SAP HANA, large consolidation estates and demanding mixed enterprise workloads need, where partitioning across many small servers would break the application model or add unwanted complexity.
Fitting four sockets into 2U rather than the traditional 4U is the SR850 V3's distinctive trick: you get scale-up capability at a rack density much closer to an ordinary 2U host, which is valuable where floor space and rack units are constrained.
Memory is the headline
Four sockets exist primarily to support memory. The SR850 V3 exposes a large number of DDR5 channels across its processors, and the goal is to hold the working set, indexes and caches of a big in-memory database entirely in RAM. Size memory to the database's certified requirement plus headroom, and populate DIMMs to keep every channel balanced, because an unbalanced four-socket configuration wastes bandwidth on a grand scale.
Capacity planning here is genuinely different from a general host. Use our memory and RAM guidance to plan a balanced, high-capacity layout, and treat memory as the first design decision rather than an afterthought once cores are chosen.
Cores, NUMA and licensing
Four populated sockets mean four NUMA domains, so topology awareness matters. A scale-up database should be sized and pinned to keep its working set NUMA-local where possible, since remote-memory access is a silent tax on latency. Choose processors for the workload rather than maxing core count, particularly because in-memory database and enterprise software licences are frequently per-core, so cores are a recurring cost as well as a capital one.
The discipline is the same one we apply to database hosts: the fewest, fastest cores that meet the requirement usually beat the highest core count on total cost. Match the SKU to the workload with our processors guidance before finalising the spec.
When SR850 V3 beats several SR650s
The four-socket case rests on whether the workload genuinely needs a single large coherent system. If you are running an in-memory database that must fit one OS image, or consolidating a sprawl of workloads that benefit from a shared large memory pool, one SR850 V3 is simpler and often cheaper to operate than several two-socket SR650s plus the clustering needed to knit them together.
If, on the other hand, your workload scales out cleanly, several SR650s give you a smaller failure domain per node and finer-grained growth. Weigh consolidation simplicity against blast radius. For the broader scale-up versus scale-out framing, our Dell vs HPE vs Lenovo comparison sets the platform context.
Resilience for mission-critical roles
A scale-up box typically carries workloads you cannot afford to lose, so insist on the platform's full RAS feature set: ECC memory with patrol scrub and advanced memory protection, redundant hot-plug power on separate feeds, redundant fans and licensed XClarity for remote management. The whole point of consolidating onto one large server is undermined if a single fault takes everything down.
Decide the high-availability model early, because it shapes how many systems and what storage you buy. Build and price an SR850 V3 to your memory and core targets in our Lenovo configurator.