A compute multiplier produces more useful work on the same compute.
A compute multiplier is any improvement that raises the useful output of a given amount of compute, without adding hardware. It is the equivalent of having more compute.
“We call them compute multipliers because they’re the equivalent of having more compute.”
Dario Amodei, Dwarkesh Podcast, August 2023
“Compute multipliers refer to algorithmic and methodological improvements in training and inference that increase the effective output of a given amount of compute.”
Anthropic S-1, glossary
Cedana is a compute multiplier because it acts on three axes at once:
Depth works through the full stack, from CPU/GPU and kernel through CUDA, Kubernetes/Slurm, Dynamo/llm-d, and the inference engines, while workloads run, adapting and optimizing based on real-time execution.
Scale the same mechanism applies to a single GPU, a multi-GPU node, a multi-node job, and the whole fleet.
Breadth applies to every model and across inference stacks (Dynamo, llm-d) and engines (TensorRT-LLM, vLLM, SGLang), with no changes to any of them.
Gains at each level compound, and because they cover the whole fleet rather than one engine’s share of it, the effect is multiplicative rather than additive.