Issue link: https://digital.copcomm.com/i/1545821
6 cgw a p r i l • m ay • j u n e 2 0 2 6 P O W E R A N D H E A T A 5090 is rated in the 575-watt class, but under sustained render load we measured something calmer: about 360 to 375 watts per card, peaking near 400, at 68 to 83 °C — the top card in a stacked pair running hottest, but comfortably under the card's limits. It draws real power and throws real heat; on a farm that is a line item, but a bounded, predictable one. Budget for the measured sustained draw, not the rated peak. There is an operational number folded into those watts. At a bounded ~368 watts per card, a median Cycles frame here finishes in about 24 seconds — on the order of 2.5 watt-hours for a delivered frame. Since the draw is bounded, energy and rack-time per finished frame should track render time, so a frame that renders much faster should cost proportionally less. We metered only the 5090, so read that as a sound inference about cost, not a measured figure — the like-for-like reading of the older nodes is still on the list. Figure 6. Per-card power draw on an RTX 5090 node under sustained render load: ≈360–375 W per card at 68–83 °C — well under the card's ~575 W class rating. W H Y T H I S S T O R Y I S A B O U T C Y C L E S Why does a render farm's RTX 5090 write-up lead with Blender? Because that is what our GPU queue runs. Over the trailing 90 days, GPU jobs were roughly a quarter of everything the farm rendered — the rest is CPU work, where V-Ray and Corona still dominate — and within the GPU mix, Cycles is about 74% of jobs, Redshi a clear second at about 15%. Engine mix keeps GPU numbers honest — and ours puts Cycles first, Redshi second. W H A T C O M E S N E X T Additional dual-5090 nodes are coming online now. On one — run as a controlled in-house bench, not from the queue — a second card brought a 1.3–1.6× speedup on Cycles, real but sub-linear, while throughput engines like V-Ray and Octane scale closer to 2×. The same matched-pair method waits on Redshi, our second GPU en- gine, and on a per-scene-type breakdown once we tag the queue for it — honest extensions of what is here, not re-runs of it. Seven weeks in, the operational story is pleasantly mundane: one driver the whole window, all but eighteen tasks finished, the median Cycles frame about three times faster than the previous generation. But the parts worth keeping aren't really about one card: production numbers aren't benchmark numbers, the median isn't the tail you plan for, and short jobs turn on overhead rather than raw speed — as true of whatever replaces the 5090 as of the 5090 itself. On a pro- duction farm, that kind of boring is the point. Watch the official "RTX 5090 in Production: 7 Weeks of Real Render-Farm Data" video in the CGW digital edition. A B O U T T H E A U T H O R Richard Ta is co-founder and technical lead of Super Renders Farm (superrendersfarm.com), a fully managed cloud render farm that supports Maya, 3ds Max, Cinema 4D, Blender, and Houdini across the major render engines. He has spent over a decade building and run- ning large-scale CPU and GPU render infrastructure for studios in more than 50 countries.