NVIDIA RTX 5090 + RTX 5080 on Ubuntu: My AI and Rendering Workstation Tested

A dual-GPU workstation with an RTX 5090 32 GB and RTX 5080 16 GB ran 24.1% faster overall on Ubuntu 24.04.5 LTS than on Windows 11 Pro across a 60-check workload index covering 8 domains, according to HWBusters' measured test of the same Ryzen 7 9850X3D hardware. The site matched Blender 4.5.14 LTS scenes, model files, and core AI runtimes (vLLM 0.30.0, PyTorch 2.13.0+cu130) between the installed Windows/WSL configuration and the native Linux rebuild, with NVIDIA 616.92 on Windows and 595.99.02 on Ubuntu. The results describe the installed stacks, and the two cards' separate memory pools do not combine into a single 48 GB allocation.

My dual-GPU workstation is moving to Ubuntu. Before erasing Windows, I measured the AI tools, image generation, and Blender workloads I actually use, then repeated the tests on the same hardware with native Linux. Ubuntu 24.1% faster overall Measured workload index · 60 speed checks · 8 domains. The index methodology https://hwbusters.com/systems/rtx-5090-rtx-5080-ubuntu-windows-ai-rendering-workstation/9/ states its weighting, exclusions, and sensitivity; the results describe the installed stacks. After moving the BOSGAME M5 to Ubuntu, I wanted to see what Linux could do for my larger Rendering PC. This machine combines a Ryzen 7 9850X3D, an RTX 5090, and an RTX 5080. It runs local AI, image generation, and rendering, and it will operate without a monitor or keyboard. First, I saved the Windows results before formatting the drive. I also kept the exact model files, workflows, source versions, and installation records on my main PC. I didn’t make a Windows disk image; this was a native Linux rebuild, with the tools and data needed to reproduce the workloads. The results below compare the installed Windows/WSL configuration with native Ubuntu. The Blender version, scenes, model files, and core AI runtimes were matched; the NVIDIA drivers and platform-specific dependencies differ. The room temperature was reported at 28°C, with similar conditions during the Windows tests. The earlier BOSGAME M5 Ubuntu article https://hwbusters.com/systems/bosgame-m5-ubuntu-ai-performance-windows-11/ describes the same move on Strix Halo. Its numbers belong to that machine; this workstation has its own measurements. The original BOSGAME M5 and ASUS GX10 review https://hwbusters.com/systems/bosgame-m5-ai-pc-review-strix-halo-vs-asus-gx10/ explains those separate platforms. Hardware and Test Configuration | Item | Tested Configuration | |---|---| | CPU | Ryzen 7 9850X3D, 8 cores / 16 threads | | Mainboard | ASRock X870E Taichi, BIOS 4.43 | | Memory | 48 GiB DDR5; two 24 GiB A-DATA modules, configured at 6000 MT/s | | GPUs | GeForce RTX 5090 32 GB + RTX 5080 16 GB | | Storage | Crucial P310 4 TB NVMe; NTFS on Windows / ext4 on Ubuntu | | Windows | Windows 11 Pro, build 26300; NVIDIA 616.92 | | Ubuntu | 24.04.5 LTS, kernel 7.0.0-38-generic; NVIDIA 595.99.02 | | Power preset | Performance; no GPU overclock or power-limit increase | | Ambient | 28°C, owner-reported for both test periods | | Blender | 4.5.14 LTS, build 62c1db4208e8 | | Language-model runtime | vLLM 0.30.0 / Python 3.12.3 / PyTorch 2.13.0+cu130 | | Image-generation runtime | Saved ComfyUI commit / Python 3.13.14 / PyTorch 2.14.0+cu130 | The RTX 5090 and RTX 5080 have separate memory pools. A two-card model profile can shard model tensors and buffers across them, but the cards do not become one interchangeable 48 GB GPU allocation. A single-card profile is also useful: it leaves the other card available for an independent render or image-generation job. Each timed phase had exclusive use of the workstation. File transfers, package installations, screenshots, and unrelated inference were kept outside those windows. Failed trials remain in the saved evidence, and the tables select the accepted series explicitly.