Issue link: http://digital.copcomm.com/i/1545821
4 cgw a p r i l • m ay • j u n e 2 0 2 6 T H E D E P L O Y M E N T The RTX 5090's launch benchmarks are over a year old. What no one has published is the sequel — what the card does once it is buried in a production queue, rendering other people's scenes on a schedule it does not control. This is seven weeks of that: task records from a live farm, not a staged benchmark. We put our first RTX 5090 node into production on April 1, 2026 — yes, April 1; the queue does not do jokes. The node is straightforward: two RTX 5090s, 128 GiB of RAM, 32 logical cores at 4.3 GHz, Win- dows 11. One detail shapes every number here: the scheduler runs one task per GPU, so each card runs its own job — every figure is a clean per-card number, what you multiply to plan capacity. The first operational metric was completion: across the window — 51 days, just over seven weeks — the node finished 99.6% of its tasks, nearly 4,900 of them. The other eighteen failed; the scheduler logs the failure, not the cause, so we won't guess at one. W H A T 3 8 P A I R E D S C E N E S S H O W The comparison we trust most is not a synthetic scene but the jobs that ran on both generations in the normal course of business — same scene, same user, at least three tasks per side before a scene counted. Per-frame time is task wall-clock divided by frame count, straight from the queue. Between April and May 2026, 38 scenes cleared that bar — drawn from 1,419 individual render tasks, 503 on the 5090 node and 916 on the previous generation. Thirty-eight is not the size of our data; it is what survives a deliberately strict filter. Across those scenes, median per-frame time on a single RTX 5090 dropped about 69% versus our previous-generation RTX 3080-class nodes — a median 3.2× speedup. The dispersion mat- ters as much as the midpoint: the middle half of scenes fall between 2.7× and 3.4× (interquartile range), the full range is 1.6–5.1×, and a bootstrap percentile 95% confidence interval on the median lands at 3.0–3.3×. A methodology note, since it changes how to read the number: each scene contributes the median of its own per-frame times on each side, and the 3.2× is the median of those 38 ratios — a median of medians, so one slow frame cannot tilt it. All of it is Blender Cycles GPU work. Three honesty notes. First, this is observational data: users some times adjust settings between re-renders, and we did not freeze their scenes for science. Second, the comparison is node-vs-node — the 5090 side is one bare card, the previous-generation side runs as GPU passthrough inside virtual machines, so some of the gap is the setup, not the silicon. Third, the baseline is the 3080-class cards these jobs actually ran on, not a current card like the RTX 4090; a clean current-card head-to-head is a separate, controlled exercise. Taken together these are single-node, Cycles-only, observational numbers — they describe our queue, and should not be generalized to other render engines or hardware. What separates a 1.6× scene from a 5.1× one is partly visible in the data (Figure 2). Plot each scene's speedup against how long its frames took on the old hardware and a loose positive trend appears — Spearman p ≈ 0.34 (two-sided p ≈ 0.04), a modest, directional link rather than a predictive one. Short frames, a few seconds each, sit at the bottom: when a frame finishes in five seconds, fixed per- task overhead — scene load, sync, the old virtualization layer — is a large share of the clock, and a faster card has little to act on. Heavier frames gain more, where throughput and memory headroom come into play — but with real scatter, including one heavy scene that held at 1.6× because its bottleneck was not the GPU but, plausibly, stor- age or a CPU-bound stage. Workload explains part of the spread, not all of it: the median says one thing, the range says it depends on the scene, and no single multiplier ever describes a queue. Figure 1. Per-scene speedup, RTX 5090 vs the previous-generation RTX 3080-class nodes these jobs ran on, Blender Cycles GPU, April–May 2026 — 38-scene production sample drawn from 1,419 render tasks. Median 3.2× (bootstrap 95% CI 3.0–3.3×); range 1.6–5.1×. Observa- tional; bare-metal 5090 vs virtualized 3080-class — see text. Seven Weeks of RTX 5090 in Production: What Production Logs Reveal About Cycles-Heavy GPU Rendering FIELD NOTES FROM THE FIRST WEEKS OF RTX 5090 NODES IN A LIVE RENDER QUEUE: 38 PAIRED PRODUCTION SCENES DRAWN FROM OVER 1,400 RENDER TASKS, ONE DRIVER FOR THE ENTIRE WINDOW, AND AN AI DENOISING PASS THAT WAS ALREADY STANDARD — ON THE NEW CARD AND THE OLD ONE ALIKE. BY RICHARD TA CO-FOUNDER & TECHNICAL LEAD, SUPER RENDERS FARM SPONSORED CONTENT