How does Kumo Engine scale?

DiskGraph stores the graph index on local SSD and reads sampled neighborhoods for the model. This allows the graph to exceed system RAM without loading the full graph into GPU memory.

The published synthetic commerce benchmark processed 50 billion nodes and 175.5 billion edges on one machine. Its 2.6 TiB index exceeded the machine’s 1 TiB RAM.

Generating context for one million seed entities took 5 seconds at two hops, 19 seconds at three hops, and 39 seconds at four hops with a warm cache. These timings exclude index preparation and model training. They depend on graph structure, sampling, cache state, and hardware.

See scale and model-input results for hardware details, deeper neighborhoods, and cold-cache measurements. For build time and resource use, see graph materialization. The tested graph size is a demonstrated workload, not a capacity ceiling.