Think of a CPU as a restaurant kitchen. Cores are the cooks, threads are the extra hands each cook can use, and clock speed is how fast each cook works. Cache is the prep counter within arm's reach, so cooks rarely walk to the pantry (RAM). Power figures are the gas bill and the heat the vent hood must clear. The socket is the kitchen's footprint: it decides which motherboard fits. Compare cook speed only within the same generation, since newer kitchens get more done per step.
The number of independent processing units on the chip. Each core runs its own work, so more cores finish multithreaded jobs faster.
When it matters: Games want about 6 to 8 good cores; video export, 3D rendering and compiling keep scaling past that. Trap: on hybrid chips the total mixes fast P-cores with slower E-cores.
Skip it if: You only browse, stream and do office work: any current 4 to 6 core chip is plenty.
The full-size, fastest cores on a hybrid chip (P-cores). They handle games and anything that needs a quick response.
When it matters: Games and lightly threaded apps run mostly on P-cores, so compare this count, not the total. Two chips with the same P-cores and clocks often game almost identically.
Skip it if: The chip is not a hybrid design: then every core is a performance core.
Smaller, lower-clocked cores (E-cores) that handle background tasks and add multithreaded throughput for less power and chip space.
When it matters: They speed up rendering and compiling and keep background jobs off the P-cores. Trap: a big jump in total cores between generations is often all E-cores, which does little for games.
Skip it if: You mostly game or use lightly threaded software.
A third, very low-power core type on some laptop chips (Intel calls them LP E-cores). They run idle and light tasks to save battery.
When it matters: They help battery life during light use such as video playback. They add almost nothing to performance, so leave them out when comparing core counts.
Skip it if: You are buying a desktop CPU.
How many instruction streams the CPU can run at once. With SMT (simultaneous multithreading), each core runs two threads.
When it matters: Extra threads help multitasking and parallel work such as encoding. Reality check: an SMT thread adds a fraction of a core's speed, not a second core.
Skip it if: You are comparing chips for gaming: core count and per-core speed matter more.
Whether the cores use simultaneous multithreading, which Intel brands Hyper-Threading, to run two threads per core.
When it matters: SMT helps rendering, encoding and heavy multitasking. Intel Core Ultra 200S desktop chips dropped it and rely on more cores instead, so judge by benchmarks, not this flag.
Skip it if: You mainly game: frame rates depend far more on per-core speed and cache.
The core design generation, such as Zen 5. It sets how much work each core finishes per clock cycle.
When it matters: A newer architecture at a lower clock can beat an older one at a higher clock. Trap: some product lines mix generations under one naming scheme, so check this field.
The maker's internal name for a chip family, such as Raptor Lake or Granite Ridge.
When it matters: It exposes rebrands: two model numbers with the same codename are often near-identical silicon. It is also the easiest term to search when checking platform or BIOS support.
Skip it if: You already know the generation and socket.
The guaranteed minimum speed of the performance cores, in GHz, when running at the rated base power under full load.
When it matters: Real clocks sit well above base almost all the time. It matters mainly in power-limited designs such as thin laptops and small cases.
Skip it if: You are buying a desktop chip and have a decent cooler.
The highest speed a performance core can reach, in GHz, usually on one or two cores at a time.
When it matters: Higher boost helps games and snappy apps. Reality check: all-core loads run lower, cooling and power limits decide how long boost holds, and clocks only compare within one architecture.
The base and peak speeds of the efficiency cores, in GHz. They run lower than the P-cores by design.
When it matters: They mainly affect multithreaded throughput such as rendering. Compare them only between chips of the same generation.
Skip it if: You mostly game, or the chip has no E-cores.
The manufacturing process the chip is built on, such as TSMC N4 or Intel 7.
When it matters: A newer node usually improves efficiency. Trap: node names are marketing labels, no longer physical measurements, and a smaller number from one foundry is not comparable to another's.
Skip it if: You are comparing speed: the node does not predict performance on its own.
The large on-chip memory all cores share, checked before the CPU reaches out to slower system RAM.
When it matters: More L3 helps games most. Trap: APU versions of a chip (AMD G models) often carry much less L3 than the non-G model with a similar name.
Extra L3 cache stacked on or under the cores, sold by AMD as 3D V-Cache on X3D models.
When it matters: Large gains in many games, little in most other work. Trap: on 12 and 16 core X3D chips only one chiplet gets the extra cache, so games rely on correct scheduling.
Skip it if: You do not play games or run cache-sensitive simulations.
The sustained power level the chip is rated for, in watts. AMD's equivalent figure is TDP (thermal design power).
When it matters: It is not real power draw. Desktop chips often run far above it under load, so size the cooler and power supply from Max Turbo Power instead.
The highest power the chip is allowed to draw while boosting, in watts. Intel also calls it PL2.
When it matters: This is the number to size coolers and power supplies for. Many motherboards loosen power limits by default, so long rendering jobs can hold this level or go past it.
The lowest power level a device maker can set the chip to, in watts, trading speed for less heat and noise.
When it matters: Mostly a laptop and mini PC concern: the same chip performs very differently depending on the power level the maker chose. Check tests of the exact device.
Skip it if: You are building a standard desktop.
The highest temperature the chip allows before it slows itself down for protection, in degrees Celsius.
When it matters: Modern chips boost until they near this limit, so reaching it under heavy load is not always a fault. Hitting it within seconds usually points to a bad cooler mount or paste.
Skip it if: You run stock settings with a sensible cooler.
Whether you can raise the clock speed by hand in the BIOS. Intel marks these chips with K; nearly all Ryzen desktop chips are unlocked.
When it matters: Trap: Intel overclocking also needs a Z-series board, and AMD A-series boards block it. Modern chips already boost near their limit, so manual gains are small.
Skip it if: You will never tune clocks or voltages.
The physical connector the CPU sits in, such as AM5 or LGA1851. It must match the motherboard.
When it matters: A matching socket is not enough: the chipset and BIOS must support the exact CPU. Older boards may need a BIOS update before a newer chip will boot.
Whether this socket is still getting new CPUs or has reached the end of its line.
When it matters: An active socket lets you drop in a faster CPU later on the same board. A dead-end socket means your next CPU upgrade includes a new motherboard.
Skip it if: You plan to replace the whole PC rather than upgrade parts.
Whether a CPU cooler comes in the box.
When it matters: Many unlocked and high-end chips ship without one, so add a cooler to the budget. Bundled coolers work at stock settings but get loud under sustained heavy load.
Which RAM generation the CPU's memory controller supports, such as DDR5 or DDR4.
When it matters: RAM must also match the board. AM5 and LGA1851 are DDR5 only. LGA1700 chips support both, but each LGA1700 board takes only one type.
The fastest RAM speed the maker officially guarantees, in MT/s (million transfers per second).
When it matters: Faster XMP or EXPO kits usually run above it, which counts as overclocking and is not guaranteed. Two sticks reach high speeds more easily than four.
Skip it if: You plan to run RAM at stock speeds.
How many independent paths connect the CPU to RAM. Mainstream desktops use two; workstation chips use four or more.
When it matters: Fill at least one stick per channel to get full bandwidth. Trap: four sticks on a mainstream desktop is still dual channel, just two sticks per channel.
The most RAM the CPU officially supports, in GB.
When it matters: Rarely a limit on current desktops. Reaching it needs every slot filled with large sticks, which usually forces lower RAM speeds.
Skip it if: You need 64 GB or less.
Whether the CPU works with ECC (error-correcting code) memory, which detects and fixes single-bit memory errors.
When it matters: Worth having for servers, NAS boxes and long scientific jobs. The motherboard must support it too. Trap: DDR5 on-die ECC is not the same as full ECC.
Skip it if: You are building a gaming or general home PC.
Whether the CPU has a built-in graphics processor (iGPU), which model it is, and how many compute units (graphics blocks) it has.
When it matters: An iGPU gives a backup display if a graphics card fails. Trap: Intel F models have none. More compute units help light gaming, but only with fast dual-channel RAM.
Skip it if: You always use a dedicated graphics card and never need a fallback.
The speed of the built-in NPU (neural processing unit), an AI engine, in TOPS (trillion operations per second).
When it matters: It runs features such as background blur and Windows AI tools at low power. Copilot+ PCs need at least 40 TOPS. A graphics card runs most heavier AI work faster.
Skip it if: You do not need Windows AI features.
How many PCI Express lanes come from the CPU for graphics cards and SSDs, and their version. Each new version doubles bandwidth per lane.
When it matters: PCIe 5.0 mainly helps the fastest SSDs. Graphics cards with only 8 or 4 lanes lose more on older generations. Trap: the board's slot must support the same generation.
Skip it if: You use one full-size graphics card and a normal SSD.