Think of a workstation or server CPU as a distribution warehouse. Cores are the loading docks, memory channels are the highways to the storage yard, and PCIe lanes are the delivery bays for graphics cards, SSDs and network cards. Cache is the staging floor beside the docks. TDP is the building's power and cooling budget. With these chips, the highways and bays often matter as much as the dock count, because heavy workloads stall when data cannot arrive fast enough.
The class of system the chip targets: enthusiast desktop (HEDT), workstation, or server.
When it matters: It predicts platform cost and features. Server parts often trade clock speed for core count and expect rack cooling; workstation parts balance clocks, lanes and memory for a desk machine.
The physical connector the CPU fits, such as sTR5 or SP5. It must match the motherboard.
When it matters: Trap: related chips can share a socket but need different boards. Threadripper and Threadripper PRO both use sTR5, but boards differ in memory channels and lanes.
Which motherboard chipsets officially support this CPU.
When it matters: A chip may fit a board's socket but lose memory channels or lanes on a lesser chipset. Check the board maker's CPU support list and BIOS version before buying.
Whether two or more of these CPUs can run together on one motherboard.
When it matters: Dual-socket servers double cores and memory, but workstation apps often scale poorly across sockets. Trap: single-socket-only parts will not work in a dual-socket board's second socket.
Skip it if: You are building a single-CPU workstation.
The number of independent processing cores on the chip.
When it matters: Rendering, simulation, compiling and virtual machines scale with cores. Reality check: many apps stop scaling well beyond 16 or so cores, and video editing gains taper off around there.
Skip it if: Your main work is gaming or lightly threaded apps: a mainstream desktop chip will be faster.
How many instruction streams the CPU can run at once, usually two per core with SMT (simultaneous multithreading).
When it matters: More threads help heavily parallel work and many virtual machines. Trap: some software licenses per core or per socket, and some apps handle very high thread counts poorly.
Skip it if: You compare chips from the same family: threads follow core count.
Whether each core can run two threads at once. Intel calls this Hyper-Threading.
When it matters: It adds throughput in rendering and server work at no cost. Some security-sensitive or real-time setups turn it off, so check whether your software vendor recommends that.
Skip it if: You run general workstation software with default settings.
The core design generation, such as Zen 5 or Granite Rapids. It sets how much work each core does per clock.
When it matters: A newer architecture often beats an older chip with more cores. It also decides instruction support such as AVX-512, which some scientific tools need.
The guaranteed minimum core speed, in GHz, at the rated TDP.
When it matters: High-core-count chips have low base clocks because every core shares one power budget. It sets the floor for sustained all-core jobs in a tight power or cooling setup.
Skip it if: Your system has ample cooling: all-core boost is the better guide.
The highest speed a single core or a few cores can reach, in GHz.
When it matters: It drives snappiness in lightly threaded tasks such as modeling viewports and CAD. Reality check: with all cores busy, clocks fall far below this figure.
The typical speed when every core is busy, in GHz.
When it matters: This is the most useful clock for rendering, encoding and simulation. It still depends on cooling and power limits, so a weak cooler pulls it down.
Skip it if: Your workloads use only a few cores at a time.
The large on-chip memory shared by the cores, checked before the CPU reaches slower RAM.
When it matters: More cache per core helps databases, simulation and some engineering codes. Trap: on chiplet designs the L3 is split per chiplet, so the total is not one shared pool.
Fast memory private to each core. The figure adds up all cores' L2.
When it matters: It grows with core count and architecture, so it rarely separates two chips on its own.
Skip it if: Almost always: architecture, L3 and benchmarks tell you more.
Extra L3 stacked on the compute chiplets, used by AMD's X-suffix EPYC and X3D parts.
When it matters: Large gains for cache-hungry work such as fluid dynamics and chip design tools. Little benefit for rendering or general server loads.
Skip it if: Your software is not limited by cache or memory latency.
How many independent paths connect the CPU to RAM. More channels means more total bandwidth.
When it matters: Bandwidth-heavy work such as simulation and AI data prep benefits most. Trap: you only get full bandwidth with one stick in every channel, so a half-filled board wastes it.
Which RAM generation and module type the CPU supports, such as DDR5 RDIMM.
When it matters: Desktop RAM will not work on most of these platforms. Buy modules on the board's qualified list, matched in size and speed across every channel.
The fastest officially supported RAM speed, in MT/s (million transfers per second).
When it matters: Trap: the rated speed often applies to one stick per channel. Filling two per channel usually drops the supported speed.
Skip it if: Your work fits in cache or is limited by compute, not memory.
The most RAM the CPU supports, in TB.
When it matters: Matters for large databases, virtualization and huge datasets. Reaching it needs the largest modules in every slot, which costs heavily and may lower memory speed.
Skip it if: Your datasets fit comfortably in a few hundred GB.
Whether the CPU supports ECC (error-correcting code) memory, which detects and fixes memory errors.
When it matters: Standard for servers and long compute jobs, where one flipped bit can corrupt results or crash a system. The board and modules must also support it.
Whether the platform needs registered DIMMs (RDIMMs), which add a buffer chip to support more and larger modules.
When it matters: Trap: RDIMMs and ordinary unbuffered DIMMs are not interchangeable. Buying the wrong type means the system will not boot.
How many PCI Express connections come from the CPU for graphics cards, SSDs and network cards.
When it matters: This is the main reason to pick these platforms: several GPUs and many NVMe drives at full speed. Check the board wiring too, since not every lane reaches a slot.
Skip it if: You use one graphics card and two or three SSDs: a mainstream desktop covers that.
The PCI Express version of the CPU lanes. Each generation doubles bandwidth per lane.
When it matters: PCIe 5.0 helps fast NVMe arrays and high-speed network cards most. Current graphics cards lose little on PCIe 4.0.
Support for CXL (Compute Express Link), a standard that runs over PCIe lanes to attach extra memory or accelerators.
When it matters: It lets servers add memory beyond the DIMM slots. It needs CXL devices and software support, which are still mostly data center gear.
Skip it if: You are building a workstation, not a data center server.
The rated sustained power and heat output, in watts (thermal design power).
When it matters: Plan the cooler, power supply and case airflow around it. These chips run near their TDP for hours during renders, unlike short desktop bursts.
The highest power level the platform may set the chip to, in watts, above the default TDP.
When it matters: Raising it gives more sustained all-core speed if the cooling can remove the heat. Server vendors often set this in firmware.
Skip it if: You keep default power settings.
What cooling the chip needs, whether a cooler comes in the box (rarely), and the temperature limit before the chip slows itself down.
When it matters: Trap: large sockets need a cooler with a cold plate that covers the whole heat spreader. Server chips often assume loud rack airflow, not a quiet desk case.
Whether the CPU has built-in graphics. Most chips in this class have none.
When it matters: Without it you need a graphics card for any display output, even to reach the BIOS. Servers usually rely on a basic video chip on the board instead.
Skip it if: You will install a graphics card anyway.
Instruction sets for wide vector math (AVX-512, Advanced Vector Extensions) and for matrix math used in AI (AMX, Advanced Matrix Extensions, Intel only).
When it matters: Scientific codes, some emulators and CPU-based AI inference can speed up a lot with them. Most creative apps see little difference.
Skip it if: Your software vendor does not mention these instructions.
Dedicated hardware on the chip for jobs like compression, encryption, and encrypting memory for virtual machines.
When it matters: They free cores in data center software that supports them. Memory encryption matters when you host other people's virtual machines.
Skip it if: You run a single-user workstation.
Whether the CPU is built from several smaller dies (chiplets) and how many compute dies (CCDs) it has.
When it matters: Chiplets allow high core counts. Trap: data moving between chiplets takes longer, and low-CCD models may get less memory bandwidth than the channel count suggests.
Skip it if: Your apps are not sensitive to memory latency.