What is VDI?
Virtual Desktop Infrastructure Explained

The Core Idea

In a traditional setup, every piece of compute — CPU cycles, RAM, GPU rendering — happens inside the device on your desk. With VDI, that compute moves into a data center. Your laptop or thin client becomes nothing more than a keyboard, mouse, and screen relay. The actual Windows or Linux session lives on a virtual machine (VM) in the cloud or on-premises hardware, and only pixels travel across the wire.

This inversion is surprisingly powerful. It means a $200 thin client can run software that would normally demand a $3,000 workstation, because the heavy lifting happens on server-grade hardware that dozens of people share. It also means your entire desktop state — open files, application settings, browser tabs — is preserved in the data center and survives device failures, theft, or replacements without a backup in sight.

How a VDI Session is Established

The sequence every VDI connection follows:

1. Authentication. The user opens a thin client or browser and provides credentials. These are verified against Active Directory, Azure AD, Okta, or another identity provider.
2. Brokering. A connection broker (e.g., VMware Horizon, Citrix DaaS, Microsoft AVD, AWS WorkSpaces) determines which VM to assign. For non-persistent desktops it picks the least-loaded available machine; for persistent desktops it routes the user back to their own.
3. Display protocol handshake. A remoting protocol — RDP, PCoIP, Citrix HDX, or NICE DCV — establishes a compressed, encrypted channel between the VM and the endpoint. Screen updates are encoded as video frames or vector commands and streamed to the client in real time.
4. Session. The user works normally. Keystrokes and mouse events travel to the VM; rendered frames travel back. On logout the session is suspended (persistent) or the VM reverts to a clean snapshot (non-persistent).

Where GPU Acceleration Fits In

Standard VDI works well for productivity tasks — Office, browser, ERP — because the CPU can encode screen updates fast enough. The moment a user opens a CAD model, 3D visualisation tool, video editor, or machine learning notebook, software-rendered graphics become a bottleneck. Frames drop, interfaces lag, and users revolt.

GPU acceleration solves this in two ways:

• vGPU (virtual GPU). NVIDIA's vGPU driver slices a physical GPU — typically an M60, T4, A10, or A16 — into multiple virtual functions, each assigned to a different VM. A single A10 can support up to 32 simultaneous 1 GB vGPU profiles. The scheduler time-multiplexes GPU cycles across sessions, giving every user a fraction of the hardware with zero interference.
• GPU passthrough. The entire physical GPU is dedicated to one VM. No sharing, no scheduling overhead — the full 24 GB of VRAM and all CUDA cores are available to a single session. Used for deep learning, simulation, and professional visual workstations where GPU time-sharing is unacceptable.

Cloud providers expose this through specific VM families: AWS's g4dn and g5 instances (T4 and A10G), Azure's NV and NVads series (M60 and A10 vGPU), and Google Cloud's N1/N2 VMs with attached T4 or L4 GPUs. These are the SKUs behind AWS WorkSpaces Graphics, Azure Virtual Desktop, and Google Cloud Workstations.

Display Protocols: The Hidden Variable

The quality of a VDI session lives or dies on its display protocol. Bandwidth, latency, colour accuracy, and 3D rendering fidelity all depend on how pixels are encoded and transmitted.

On-Premises vs Cloud VDI

VDI can run entirely in your own data center (on-prem), entirely in a public cloud (DaaS — Desktop as a Service), or in a hybrid of both. The choice is mostly economic and operational:

Five Reasons Organisations Deploy VDI

When VDI Is Not the Right Answer

VDI is not a universal fix. There are workloads and situations where it adds cost and friction without meaningful benefit:

• High-bandwidth local peripherals. Audio recording studios, video capture setups, and USB-connected lab instruments suffer from protocol latency and driver complexity.
• Offline work. VDI requires network connectivity. Field workers, aircraft crew, or remote sites with unreliable connectivity cannot depend on it as their primary environment.
• Sub-5ms latency requirements. Algorithmic trading, real-time simulation, and certain industrial control interfaces are sensitive to the additional hop a VM introduces.
• Small teams without dedicated IT. A 10-person startup does not need a connection broker, profile management solution, and VDI licensing stack. Standard laptops with cloud storage are simpler and cheaper.

Cloud VDI Pricing in Practice

Cloud VDI pricing has two components that many buyers miss: the compute cost (the VM running your desktop) and the Windows licensing cost. Most cloud providers bundle Windows into their VDI-specific SKUs, but if you bring your own licenses you can save 20–30% on some instance types. Spot or preemptible VMs are generally unsuitable — an interrupted session mid-work is not acceptable — so on-demand or reserved pricing applies.

GPU SKUs add a meaningful premium. An Azure NV4as_v4 (1/8 M60 vGPU) starts at around $0.23/hr, while a full NVads A10 v5 with 36 vCPUs runs to $7.20/hr. At 160 working hours per month that's $37 vs $1,152 per seat — a decision that warrants genuine profiling of whether users actually saturate the GPU.