What is VDI for HPC? GPU-accelerated remote desktops, explained

What VDI means in CLASSIC IT

Virtual Desktop Infrastructure (VDI) is the general term for delivering a desktop environment from a central server rather than from the machine in front of the user. The server does the processing; the user’s device renders only what the server sends.

For office workloads, the overhead is manageable. For GPU-intensive simulation, seismic visualization, or VFX rendering on Linux, that overhead adds cost and latency with no corresponding benefit. The hypervisor is solving a problem HPC teams don’t have.

What HPC USERS need from VDI

HPC environments run on bare-metal Linux. GPU servers are physical machines with NVIDIA A-series or RTX cards, managed by SLURM, PBS, or LSF, running applications like Ansys, CATIA, OpenFOAM, or Maya. There is no Windows intermediary in the workflow, and no reason to introduce one for remote access.

The HPC community often calls this remote visualization, which is a more accurate name for what’s actually happening.

How GPU-accelerated remote visualization works

DCV (now Amazon DCV, formerly NICE DCV), distributed by NI SP, is the protocol built for this use case. It connects to physical Linux GPU servers with no hypervisor layer and no vGPU licensing required. A single GPU server with 48GB of VRAM supports 10 to 15 concurrent full-performance 3D desktop sessions, each with independent GPU memory allocation. The exact number scales with VRAM and the application's per-session memory footprint.

Beyond raw streaming, a few protocol properties matter specifically for HPC:

• Bandwidth adaptation. Session quality adjusts automatically to available bandwidth, so users on WAN links get a usable session rather than a degraded or frozen one. RVX, an Icelandic VFX studio, uses DCV for team members connecting from mainland Europe. Their CTO described the transatlantic session quality as "the best they had seen so far": a result that holds because the protocol was designed for production WAN conditions, not LAN demos.
• Color fidelity. For simulation visualization and VFX, color accuracy is not aesthetic: it affects decisions. DCV supports 4:4:4 chroma and a lossless mode for applications where compressed color would introduce errors in what users see on screen.
• Data stays on the server. Pixels leave the server, AES-256 encrypted. The simulation files, models, and output datasets don't. For large datasets that cannot practically move over the network, and for compliance environments where data residency matters, this architecture is what makes remote access viable.
• Broad client support. HTML5 browser, native clients for Windows, macOS, and Linux, and support for up to 4×4K monitors.

Where VDI fits in the HPC workflow

Interactive VDI sessions and batch job submission are complementary, not competing.

A typical CFD workflow: submit an overnight Ansys computation through the job scheduler as a batch job. The next morning, log in to visualize the result, adjust parameters, and submit the next iteration. The visualization and parameter-adjustment step needs an interactive GPU session. The overnight compute run doesn't.

EF Portal, the HPC portal developed by NI SP, manages both from a single interface:

• Users see their queued batch jobs and their running VDI sessions in the same browser dashboard.
• They can launch a DCV remote desktop from the same interface where they submitted the job, without a separate login or a separate tool. Same authentication, same file management, same monitoring view.

That integration shifts VDI from a convenience to an operational requirement at scale. At that volume, moving cleanly between batch submission and interactive visualization without switching systems is a real throughput question, not a preference.

If your team is still working out whether an HPC portal makes sense at all, Do you actually need an HPC portal? walks through that decision with a practical self-test.

For teams whose workflow is VDI-first (without batch job scheduling) EF Views covers that scope without the HPC portal overhead.

Managing VDI access at scale

Connecting a single user to a single DCV server is straightforward. At scale (more than a handful of GPU servers, users in multiple locations, access from outside the company network) you need a management layer for authentication, connection routing, and session monitoring.

For pure VDI deployments, NI SP offers two solutions that cover different layers of the stack:

For teams that also need batch job scheduling alongside VDI, EF Portal integrates both in a single interface: covered in the Where VDI fits in the HPC workflow section above.

What to look for when you assess HPC VDI

Four criteria that most classic VDI checklists miss, because they were written for Windows environments.

• Linux-native support. If your simulation workloads run on Linux, the protocol must connect natively to Linux GPU servers. Solutions that route through a Windows virtual machine to reach Linux add the hypervisor overhead that makes classic VDI expensive for HPC in the first place.
• GPU sharing model. Each user can have a dedicated physical server, or multiple users can share a single server with independent GPU memory allocation. Shared-GPU deployments offer substantially better hardware economics when users don't all run peak-load sessions simultaneously, which is typical of most HPC teams.
• Color fidelity. For CFD results, seismic interpretation, or VFX work: confirm lossless mode and 4:4:4 chroma support before signing. A protocol that compresses color corrupts what your users see and the decisions they make from it.
• WAN performance. Remote and hybrid teams need a protocol tested under real network conditions. Run a trial from the actual locations your users connect from, not a LAN benchmark in the same building as the server.
• Self-service session management. As team size grows, the protocol alone isn't enough: look for a web-based portal with SSO (LDAP/Active Directory), DCV Session Manager support, and user-facing session lifecycle controls. Without it, every connection request becomes an IT ticket.

If you're moving from HP Anyware, our DCV vs HP Anyware comparison covers the feature-by-feature differences relevant to that migration.

Try it on your cluster

DCV installs in under 10 minutes on any Linux server with an NVIDIA GPU. A 30-day free trial gives enough time to let your actual users connect to actual workloads: the only measure that matters is session quality in their application on their hardware, not a benchmark.

For teams already running EF Portal: DCV VDI sessions integrate directly into the portal interface. Users launch their remote desktop from the same browser window they use for batch submission, once the DCV integration is configured in the portal admin settings.

SEE VDI IN ACTION

Frequently asked questions

RESOURCES

• EF Portal - HPC Portal
• EF Views - VDI Portal
• DCV Remote Desktops
• DCV Pool Manager
• Do you need an HPC Portal? Get your HPC Portal fit score.
• EnginFrame to EF Portal FAQ
• What's New in EF Portal 2026.0 (EF Portal 2026.0 Release video)
• How to install EF Portal (Guide)
• EF Portal or Open OnDemand: Which is the right portal for your HPC?
• Video Walkthrough
  • EF HPC and VDI Portal - Complete End-User Walkthrough
  • EF HPC and VDI Portal - Complete Administrator Walkthrough