Workstation Specialised

GPU path-traced rendering and data-parallel training divide almost linearly: two cards approach twice the throughput and four approach four times. Editing, CAD viewports and single-scene interactivity do not; they are served by one GPU together with the processor, and a second card contributes nothing. Establishing which of these accounts for the majority of the workload is the first step, as it determines whether a multi-GPU configuration is warranted.

A model, scene or dataset either fits within a card's memory or it does not. Where it does not, the task fails or reverts to system memory at substantially reduced speed. Two cards double throughput but do not pool memory for most workloads: a scene requiring 20GB requires it on each card. For this reason we establish model and scene sizes before core counts, and a card with greater memory frequently outperforms a faster card with less.

Four GPUs at full bandwidth require more PCIe lanes than any desktop processor provides, which is why HYPERFOCUS is built on a Threadripper-class platform where the lane count maintains every card at full width and leaves capacity for fast storage. On a desktop board a four-card configuration is bandwidth-constrained; the platform is generally the more consequential specification decision.

GeForce-class cards provide the greatest throughput per dollar and are suitable for most render and training work. Workstation-class cards justify their cost in three circumstances: certified drivers for CAD and simulation suites, greater memory per card, and cooling designed for cards installed in close proximity under sustained duty. We specify according to workload rather than tier, and a substantial proportion of our multi-GPU systems are supplied with consumer cards.