Guides
How to Build a VRAM Budget for Motion Graphics Workloads
· RenderBob team
Most studios size hardware by vibes: buy the biggest card the budget allows and hope. In a year when the biggest card costs $5,000 and may not be in stock, guessing is expensive. Size the farm to the jobs you actually ship.
Most studios size hardware by vibes: buy the biggest card the budget allows and hope. In a year when the biggest card costs $5,000 and may not be in stock, guessing is expensive. Size the farm to the jobs you actually ship.
Step 1: Inventory your real jobs, not your aspirational ones
List the workflows you actually ship: model, resolution, frame count, upscale steps, and any LoRAs or ControlNets. A studio's true VRAM ceiling is set by its heaviest recurring job, not the one hero shot it does twice a year.
Step 2: Measure peak VRAM, not average
Run each workflow and watch the peak, which for video usually lands during the upscale or second-sampling pass, not the initial denoise. Budget to the peak or you will OOM exactly when a client is watching.
Step 3: Separate the fits-everywhere jobs from the barely-fits jobs
Most iteration and preview work fits comfortably on mid-range VRAM. A minority of final-quality passes sit right at or above your ceiling. Sizing the whole farm for that minority is how studios overspend.
Step 4: Price the two answers side by side
For the barely-fits jobs, compare the cost of a bigger owned card (at 2026 prices, and only if you can source it) against metered cloud time for those specific passes. Often the heavy jobs are rare enough that renting VRAM by the hour beats owning it year-round.
Step 5: Design for quantization headroom
Assume NVFP4/FP8 will claw back 40–60% of VRAM on supported hardware, but pin the precision per workflow and test output quality. Do not let each artist toggle it by feel. That is how outputs drift.
Step 6: Leave an overflow valve
Whatever you buy, size baseline load to owned hardware and keep a defined path to cloud for the peaks. A fixed farm with no elastic escape hatch is a delivery risk the moment demand exceeds what you bought.
The output is a map: which jobs live on owned nodes, which precision they run in, and which jobs are cheaper to burst than to buy for. That map tells you where not to spend in a year when spending is hard.
More from the blog
- A Risk Ladder for AI in Documentary
Screenweaver's 8 October guide ranks five documentary uses of AI by risk, from archive restoration to a synthetic face, and pairs them with EU disclosure rules now in force.
- The B-Roll Gap: Generated, Selected, or Shot
Every cut eventually needs a shot that does not exist. AI can generate it or search a library for it, and stock libraries are answering with different rules. Here is how to choose.