CUDA, PyTorch & Driver Setup for Blackwell AI Workstations

A new RTX 5090 or RTX PRO 6000 is only as useful as the software stack on top of it, and this is exactly where many self-built workstations stall for days. Mismatched driver, CUDA, and framework versions produce the classic failures: the GPU is not detected, PyTorch falls back to CPU, or kernels error out at runtime. This guide walks through a clean, version-matched setup so a Blackwell-class card is doing real work the day it arrives, not the week after.

Understand the version stack before you install

Four layers must agree: the NVIDIA GPU driver, the CUDA toolkit, the cuDNN library, and your deep-learning framework (PyTorch, TensorFlow, JAX). The driver must be new enough to support the Blackwell architecture; the framework build must be compiled against a CUDA version your driver supports; and cuDNN must match. The single most common mistake is installing a framework build compiled for an older CUDA than the hardware needs, or a driver too old for the GPU. Decide on a compatible set of versions before installing anything.

A clean installation order

• Install a current NVIDIA driver that explicitly supports the RTX 5090 / RTX PRO 6000 (Blackwell) and reboot.
• Verify the GPU is visible with the nvidia-smi utility — it should report the card and the maximum CUDA version the driver supports.
• Install the matching CUDA toolkit and cuDNN, or let your framework's packaged CUDA runtime handle it inside an isolated environment.
• Install the framework build compiled for that CUDA version.
• Validate end to end: confirm the framework sees the GPU and runs a small tensor operation on it before assuming the stack is healthy.

Use isolated environments — always

Install your Python AI stack inside isolated environments (conda or virtualenv), and consider containers for reproducibility. Different projects often need different framework and CUDA-runtime versions, and isolating them prevents one project's upgrade from breaking another. Modern framework wheels bundle their own CUDA runtime, which means you can frequently run multiple CUDA versions side by side per-environment as long as the underlying driver is new enough for all of them.

Linux or Windows?

Both work. Ubuntu LTS is the most common choice for serious AI development — the broadest tooling support and the smoothest driver and container experience. Windows 11 Pro is a fully valid option, especially with WSL2 for a Linux-native workflow, and is often the right call where the workstation also runs Windows-only professional applications. The decision should follow your team's existing tooling, not dogma.

Why a pre-configured machine saves real time

The setup above is straightforward when you do it often and frustrating when you do not. Every hour an expensive engineer spends fighting driver and CUDA mismatches is an hour not spent on the actual work. A workstation that arrives with a validated, version-matched stack already in place removes that entire class of problem.

Nexus Compute ships RTX 5090 and RTX PRO 6000 workstations with your chosen OS — Ubuntu LTS or Windows 11 Pro — and the CUDA toolchain pre-loaded and validated against the hardware, so the machine runs real workloads on day one. We configure, test, and warranty-back each system and quote within 48 business hours.