RTX 5090 in Production: 7 Weeks of Real Render-Farm Data

Launch-day benchmarks told you the RTX 5090 is fast. This is the less glamorous sequel — what the card actually did across seven weeks in a live production render queue at Super Renders Farm. Observational data, not a staged benchmark. KEY NUMBERS (Blender Cycles, April–May 2026, one dual-RTX-5090 node, per-card): • 3.2× - median per-frame speedup vs our previous-generation GPU nodes, across 38 paired production scenes (range 1.6–5.1×). • ~83% - share of Cycles jobs already running an AI denoise pass (OptiX / Intel OIDN) - essentially identical on the previous generation. The new card didn't start the habit. • ~5.6 GB median / ~37 GB heaviest - peak render-device memory; the heaviest job exceeded even the 5090's 32 GB. • One driver (581.80 / CUDA 13.0), zero churn; 99.6% of ~4,900 tasks completed. CHAPTERS 0:00 Launch benchmark vs production 0:23 3.2× - the median, not the hero number 0:40 AI denoising was already the default 0:59 The VRAM tail 1:21 One driver, zero churn 1:35 Production data, not a benchmark TRANSCRIPT I'm Richard Ta - I run Super Renders Farm. The RTX 5090 shipped over a year ago, with its launch benchmarks. We didn't run another one - we put the cards in the queue on April first and watched seven weeks of real jobs, tracking four things: speed, AI denoising, the VRAM tail, and stability. First, the speedup. Across 38 scenes that ran on both the new cards and our previous generation, the median frame came back about 3.2× faster. The range was wide, though - 1.6 to 5.1×. Production never has just one number. Second, AI denoising. Here's the part I didn't expect: the AI denoiser was already doing most of the cleanup - on about 83% of our Cycles jobs, and the same rate on the old hardware. The 5090 didn't bring AI to the pipeline; it just made everything around an already-routine step faster. Third, the VRAM tail. Memory is a tail story. Most scenes barely touch it - under 6 GB. But the heaviest job wanted 37, past even the 5090's 32 - the kind of job that falls back to CPU, or won't render on the GPU at all. You size hardware for the tail, not the median. Finally, stability. The least dramatic finding is the one I'd want before buying: one driver the whole window, zero churn, and 99.6% of nearly 4,900 tasks finished. None of this is a benchmark - it's production data. Messy, honest, more useful for it. Up next: power tests, the controlled bench, and the next card. Full write-up (Computer Graphics World, SIGGRAPH 2026) Super Renders Farm: https://superrendersfarm.com #RTX5090 #RenderFarm #Blender #GPURendering #Cycles