Compute
Compute is the servers themselves: processor, memory, local drives, network cards and accelerators. Different servers exist because the constraints are different. A dense 1U server may suit scale-out workloads, while a larger chassis gives more room for memory, drives or accelerators.
What I work on
I do not design the ProLiant hardware. I work on which CPU, memory, storage and accelerator options customers can choose, which combinations stay buildable, and what changes when a platform or a part reaches the end of its life.
Three servers, three shapes
HPE ProLiant Compute DL360 Gen12
A compact 1U, two-processor server that HPE positions as a hybrid cloud platform for enterprise applications.
HPE product page ↗HPE ProLiant Compute DL380 Gen12
A 2U, two-processor server. The extra height buys room for many more drives, up to 36 in the EDSFF E3.S format.
HPE product page ↗HPE ProLiant Compute DL380a Gen12
A 4U, two-processor server built around GPUs, with space for up to ten double-width accelerators. HPE positions it for AI inference.
HPE product page ↗The four choices inside a server
They compete for the same chassis, expansion slots and power budget.
CPU
The general-purpose processor in the server.
The processor also fixes how many memory channels and expansion lanes you get, how much power the server draws, and often what the software on top of it costs.
More cores is not automatically better. Many virtualization and operating-system licences are priced per core, so the cheapest processor on the quote can be the most expensive server to run. I try to get that into the configuration decision, not the renewal conversation.
Memory
The working space a server uses while applications run.
The number of memory slots follows the number of processors, not the number of cores. How a capacity is reached, with fewer large modules or more small ones, changes speed, cost and room to grow.
A design built on the largest modules depends on a narrow set of parts. One built on smaller modules can fill every slot and leave no headroom. I care about which routes to a capacity are validated, because that is what lets a design move when memory pricing or availability does.
HPE publishes this for the ProLiant Compute DL380 Gen12: with Intel Xeon 6 processors it supports DDR5 at up to 6400 MT/s with one module per channel, and 6000 MT/s with two. The same slots and the same capacity target, but a different speed depending on how you fill them.
Local storage
The drives inside the server itself, rather than a separate storage array.
Drive format, endurance and how the drive connects can change which chassis you need and how much expansion is left for everything else.
The move from 2.5-inch drives to the smaller EDSFF E3.S format is a good example. It is not only a different drive: the server needs a different drive cage, so a drive choice becomes a chassis choice.
Accelerators
GPUs and similar processors that handle highly parallel work, most visibly AI training and inference.
Adding them pulls power, cooling, expansion slots and often the network into the same decision. The accelerator ends up shaping the server around it.
That is why accelerators sit inside compute. My part is tying each accelerator to a small number of validated server pairings, so slots, power and cooling are part of the product definition rather than something discovered during integration.
See the hardware inside a server
kushagragoyal.com/stack/compute