Framestore Cuts Render Costs 60% with Composable GPU Pools

Framestore operates one of the most demanding GPU render farms in the visual effects industry, delivering work across feature films, television, and commercial productions running simultaneously.

The challenge was familiar: peak demand during a film's final render sprint consumed every available GPU. Between projects, utilization dropped to under 30%. The studio owned hardware sized for peak — expensive, power-hungry, and idle half the time.

Cerio.ai replaced the static farm with a composable GPU pool. Each production received a logical cluster — sized dynamically to its current deadline pressure — assembled from the shared physical pool. When a production wrapped its heavy render window, those GPUs flowed to whichever show needed them most.

The scheduling layer was tuned for VFX semantics: frame-range awareness, per-shot priority overrides, and graceful preemption with automatic checkpoint-and-resume so no render progress was lost when resources shifted between shows.

Security isolation ensured one production's unreleased footage remained invisible to other tenants, with hardware-level separation enforced at the virtualization layer.

Results after six months in production: compute costs dropped 60% versus the previous owned-hardware model. Render throughput increased 3× on the same physical GPU count, driven by the elimination of idle capacity between productions. Deadline misses — previously a regular occurrence during crunch periods — fell to zero across three simultaneous major productions.