Loading page
GPU requirements are set by what you are running, not by which card tops a spec sheet. Each page below covers what a workload demands of a GPU, which models suit it, and what providers currently charge for them.
Multi-GPU nodes with HBM and high-bandwidth interconnects
Throughput per dollar, with enough VRAM to hold the weights
LoRA and QLoRA on one GPU, full fine-tunes on a node
Diffusion models fit in 24 GB; batch size buys throughput
FP64 throughput, ECC memory, and node-level interconnect
Hardware encoders, VRAM for scene data, and stream density
An embedding model and a generator, sized separately
The lowest cost per unit of work, not the lowest hourly rate