CPU / processor

“Eight cores at five gigahertz” is a bit like “five seats and a big engine”: it tells you something real, but not how fast you cross town with a full car and a headwind.

The short answer

A CPU (central processing unit) executes the instructions of programs. Its spec sheet lists cores, threads, clock speeds, cache and a power figure, and each of these describes a limit under certain conditions, not what you get on your desk. What you actually get depends on the workload, cooling, power limits, memory and, in a virtual machine, on what the host hands out.

How it works

1. A CPU understands a fixed set of instructions. Intel documents its architecture in the Software Developer’s Manual, split into volumes: architecture (Volume 1), instruction set reference (Volume 2) and system programming (Volume 3). Extensions add instructions; Intel’s spec sheet for the example processor below lists, for instance, AVX2 and VT-x (a hardware aid for virtual machines).
2. Package, core, thread. The Linux kernel documentation describes the topology in three levels: the package is the chip in a socket, a package holds one or more cores, and a core can run more than one thread. For Linux, a thread is a logical CPU, and that is what nproc and the task manager count. The documentation also notes that sockets are irrelevant to the kernel as such.
3. The spec sheet counts both. Intel’s sheet for the Xeon E5-2699 v4 lists 22 cores and 44 threads and marks Hyper-Threading (Intel’s name for running two threads per core) as supported. So 44 threads does not mean 44 cores. Per core, two threads share the same execution hardware, which is why two threads do not run twice as fast as one (own classification, see below).
4. Clock speed has two numbers. The same sheet lists a base frequency of 2.20 GHz and a maximum turbo frequency of 3.60 GHz. Intel’s article on thermal design power says that a processor can exceed its rated power with turbo or AVX workloads only for a limited time, until thermal throttling or a power limit steps in. Treat the turbo value as a peak, not as a constant.
5. TDP is a power figure, not a speed figure. Intel defines thermal design power (TDP) as the power under the maximum theoretical load. The sheet gives 145 W for this processor. It is the number a cooling design is planned around; it says nothing about performance (own classification).
6. Cache sits on the chip. The sheet lists 55 MB of “Smart Cache”. Cache is a small, fast memory on the processor that holds data it needs again soon, so it has to wait less for main memory (own explanation).
7. The CPU decides what the rest of the machine can do. The sheet also fixes the memory: DDR4 at 1600 to 2400 MT/s, four channels, up to 1.5 TB, ECC supported, and a maximum memory bandwidth of 76.8 GB/s. The same applies to expansion: PCI Express 3.0 with up to 40 lanes. A board can only offer what its processor offers.
8. The board limits the processor too. HPE’s QuickSpecs for the ProLiant DL360 Gen9 list the E5-2699 v4 as an option, allow up to two processors, do not support mixing processor models, and state that v4 processors require PC4-2400T memory.
9. In a virtual machine you see the guest’s view. The manual page of lscpu says that in a virtual environment the output shows the guest’s configuration, not the physical hardware.

In practice

Read what the system reports, and keep “logical” and “physical” apart:

lscpu              # model, sockets, cores per socket, threads per core
lscpu -e           # one line per logical CPU
nproc              # number of logical CPUs available to the current process

An example from Bytearium’s own inventory (reading of 14 Oct 2026): the ESXi view of the DL360 Gen9 shows “Xeon E5-2699v4 @2.20GHz” and “44 CPUs”. The inventory deliberately does not turn that into cores and threads. Using the spec sheet figures, 44 fits one processor with Hyper-Threading (22 × 2) as well as two processors without it (2 × 22); how many sockets are populated and whether Hyper-Threading is on is not recorded. That is the first thing to look up in the BIOS or in iLO before anyone quotes “44 cores”. The inventory also lists 28 threads for the G-PC; the processor model is not recorded.

Cores you can’t use

More cores only help if the work can be split. A single-threaded task (many game servers, a build step, a database query without parallel plans) runs on one core and is bounded by that core’s clock and cache, however many sit idle next to it. Before paying for a 32-core part, look at what your typical load actually does: top or htop showing one core at 100 % and the rest asleep is the sign that clock speed, not core count, is your bottleneck. For a homelab that mostly runs VMs and containers, many moderately clocked cores are the better buy; for one latency-sensitive service, fewer fast ones are.

Checklist

• Read the exact model name and look up its spec sheet at the maker, not a table from a shop.
• Compare cores and threads separately and write both down.
• Treat the turbo frequency as a peak and the base frequency as the number to plan with.
• Check processor, memory and board against the board maker’s compatibility list before you buy.
• Run lscpu inside a virtual machine and look up the host’s hardware separately.
• Note how many sockets are populated and whether Hyper-Threading is on, together with the model.
• Plan cooling and power supply from the TDP, with margin.