CPU

A central processing unit (CPU), also called a central processor, main processor or just processor, is the electronic circuitry that executes the instructions of a computer program. It performs the basic arithmetic, logic, control and input/output operations specified by those instructions, in contrast to specialized circuits such as a graphics processing unit (GPU) or a neural processing unit (NPU).

In virtual reality (VR) and augmented reality (AR) systems the CPU runs the application's main logic loop, including game and physics simulation, scene management, input handling and the preparation of draw calls that the GPU then renders. On standalone headsets the CPU is one block of a single system on a chip (SoC) that also holds the GPU, an NPU and dedicated blocks for positional tracking and passthrough; on PC-tethered systems the CPU is a separate desktop or laptop processor. Because VR renders a stereo image at a fixed, high refresh rate, sustained per-frame timing on the CPU is as important as raw throughput: a single late frame can trigger reprojection and visible judder.

Origin and architecture

The modern programmable CPU descends from the stored-program concept described by John von Neumann in the First Draft of a Report on the EDVAC, circulated on 30 June 1945 to people connected with the EDVAC project at the University of Pennsylvania. The report gave the first written description of a computer that stores its program in memory alongside its data, and laid out the components that became known as the von Neumann architecture: a memory, a central arithmetic unit, a central control unit, and input and output organs. The name has been criticized because von Neumann's draft did not credit the ENIAC engineers John Mauchly and J. Presper Eckert, who had developed key ideas of the design.

At the hardware level a CPU repeatedly carries out a fetch-decode-execute cycle: it fetches the next instruction from memory, decodes what operation that instruction specifies, and executes it using functional units such as an arithmetic logic unit, registers and control logic. A modern processor packages several such execution units, or cores, on one die so that independent streams of instructions can run in parallel. The two dominant instruction-set families are x86, used by most desktop and laptop processors that drive PC VR, and ARM, a reduced instruction set computing (RISC) design that dominates mobile and standalone devices because its streamlined instruction set and core design draw less power.

System on a chip in standalone headsets

Standalone VR and AR headsets do not carry a separate CPU package. Instead the CPU is one component of an SoC that also integrates a GPU, an NPU, a memory controller, an image signal processor and connectivity, all on a single piece of silicon. Qualcomm's Kryo CPU cores, used across its Snapdragon and dedicated XR processors, are custom designs built on the ARM instruction set.

Qualcomm announced the Snapdragon XR1 in May 2018, describing it as the first platform dedicated to extended reality, with Meta (then Oculus), HTC Vive, Vuzix and Pico already developing on it. The first-generation Oculus Quest, released on 21 May 2019, instead used the Snapdragon 835 Mobile VR Platform, a chip originally built for smartphones. The Meta Quest 2 (2020) moved to the Qualcomm Snapdragon XR2, an octa-core Kryo 585 CPU paired with an Adreno 650 GPU, which Qualcomm said roughly doubled CPU and GPU performance over the Snapdragon 835.

The Meta Quest 3 was the first headset to ship with the Snapdragon XR2 Gen 2, an eight-core CPU based on the Snapdragon 8 Gen 2 smartphone chip but tuned for VR and built on a 4 nm process. Qualcomm and reporting on its specifications state a CPU increase of more than 33 percent and a GPU up to 2.5 times more powerful than the original XR2, an NPU up to 4 times faster for peak AI performance and up to 8 times more efficient per watt, support for displays up to 3K by 3K per eye, and processing of data from up to ten cameras.

Processors used in major standalone headsets

Headset	Year	Processor	CPU
Oculus Quest	2019	Snapdragon 835 Mobile VR Platform	Octa-core (smartphone-derived)
Meta Quest 2	2020	Qualcomm Snapdragon XR2 (Gen 1)	Octa-core Kryo 585
Meta Quest 3	2023	Snapdragon XR2 Gen 2	8-core (based on Snapdragon 8 Gen 2)
Apple Vision Pro	2024	Apple M2 + Apple R1	M2: 8-core (4 performance + 4 efficiency)

Dedicated co-processors and offloading

A general trend in XR silicon is to move latency-critical, repetitive work off the CPU and onto fixed-function or dedicated blocks. On the Snapdragon XR2 Gen 2, dedicated hardware acceleration handles positional tracking, camera passthrough, SpaceWarp motion extrapolation and super-resolution sharpening; UploadVR reported that the passthrough acceleration cut end-to-end passthrough latency from around 50 milliseconds to around 12 milliseconds. Qualcomm describes the XR2+ Gen 2 as using dedicated areas of the SoC to accelerate room tracking, passthrough, spacewarp and super resolution, which offloads the CPU and GPU and reduces both power and latency.

The Apple Vision Pro, released in 2024, makes this split explicit by using two chips. An Apple M2 with an 8-core CPU (4 performance and 4 efficiency cores), a 10-core GPU and a 16-core Neural Engine runs visionOS and applications, while a separate Apple R1 chip processes the input from the headset's cameras, sensors and microphones. Apple states the R1 streams images to the displays with 12-millisecond photon-to-photon latency and has 256 GB/s of memory bandwidth. Dedicating the R1 to sensor fusion keeps that fixed workload off the M2's CPU and helps hold the motion-to-photon path short.

Role in VR and AR

Simulation and the frame budget

In a VR application the CPU runs the simulation: game logic, physics, animation state and the construction of the scene that will be drawn each frame. Meta's PC VR performance guidelines describe the CPU as involved with "simulation logic, state management, and generating the scene to be rendered," and recommend limiting per-frame script or logic execution to roughly 1 to 3 milliseconds so it fits inside the frame budget. At a 90 Hz refresh rate the entire frame, CPU work plus GPU work, must complete in about 11.1 milliseconds, and at 72 Hz in about 13.9 milliseconds. The consequences of a CPU bottleneck differ from average frame-rate drops: late individual frames cause judder even when the average appears adequate, and a stall during a physics or logic step can force the runtime to substitute a reprojected frame.

Because VR engine code runs much of its simulation on a single main thread, per-core clock speed and single-thread performance often matter more than total core count for VR; a system can have many cores yet still stutter if its single-thread speed is mediocre. Meta notes that on a PC headset, owing to higher resolution and GPU load, the CPU tends to be less of a bottleneck than it is on mobile VR devices, where the lower-power SoC CPU is more easily saturated.

Draw calls and stereo rendering

A draw call is a command the CPU submits to the GPU to draw a batch of geometry, and each call carries driver overhead on the CPU. VR magnifies this because the scene must be presented to two eyes. With traditional multi-pass stereo rendering the engine runs two complete render passes, one per eye, which nearly doubles the CPU workload. Single-pass instanced stereo rendering reduces this cost: the engine issues a single draw call per object with an instance count of two, and a vertex shader decodes the instance identifier to select the eye, which halves the number of draw calls submitted on the API side and lowers CPU usage. Meta's PC guidelines recommend keeping each frame to a maximum of 500 to 1,000 draw calls to bound this CPU driver overhead.

Tracking and sensor processing

A worn headset continuously estimates its own pose and the pose of any controllers, which is computation-heavy. On standalone systems much of this inside-out tracking work and the related computer vision is now handled by dedicated blocks inside the SoC rather than by the general-purpose CPU, which lowers latency and power draw. Where dedicated hardware is absent or insufficient, these tasks fall back to the CPU and add to its frame budget. The Apple R1 is the clearest example of fully separating sensor processing from the application processor by giving it its own chip.

Thermal and power constraints

On battery-powered, head-worn devices the CPU shares a fixed thermal and power envelope with the GPU and the other SoC blocks. The shift to dedicated accelerators and to efficient ARM-based cores is driven in part by these limits, since a streamlined, power-efficient core design extends battery life and reduces heat in a device worn on the face. This is one reason standalone-headset SoCs descend from smartphone designs, where the same constraints apply, and why offloading repetitive tracking and passthrough work onto fixed-function hardware is preferred over running it on the CPU.

Current status

As of mid-2026 the CPU remains a required component of every VR and AR system, but in head-worn devices it is increasingly one part of a heterogeneous SoC rather than the sole compute engine. Standalone consumer headsets are built on ARM-based SoCs (Qualcomm's Snapdragon XR2 Gen 2 family in the Meta Quest 3 and Quest 3S, Apple's M-series plus R1 in the Apple Vision Pro), in which a growing share of latency-critical tracking, passthrough and reprojection work runs on dedicated accelerators and NPUs, leaving the CPU for simulation and application logic. PC VR continues to rely on conventional x86 desktop and laptop CPUs, where single-thread performance and consistent frame timing are the main factors that determine whether a headset runs smoothly.