Topology is now part of the scheduling problem. NVIDIA Topograph is an open source toolkit that discovers cluster hardware and network relationships, then turns them into information workload managers can use. That matters because AI factories are power-limited systems, and GPUs deliver maximum value only when the cluster keeps them fed with data.
GPU workloads exchange data throughout training and inference, so placement affects more than convenience. A poorly placed distributed workload can fragment topology domains, push traffic across shared links, reduce throughput, raise job costs, and leave GPUs consuming provisioned power while waiting for data. Expensive silicon has many talents, but waiting is not one of them.
One topology model for cloud and on-premises clusters
Topograph discovers topology through cloud APIs or on-premises fabric systems, then normalizes the findings into a common, canonical model. Its provider layer handles discovery, while its engine layer translates that model into formats workload managers already understand.
The toolkit can publish topology as Kubernetes node labels, Slurm configuration, Slinky ConfigMaps, Node Feature Discovery resources, or instance-oriented topology JSON. Operators can deploy it through Helm on Kubernetes or use native packages on Slurm clusters, with simulation utilities available for testing without production hardware.
Cloud integrations with working support include Google Cloud, Lambda, Nebius, Nscale, and Oracle Cloud Infrastructure. On-premises deployment models cover InfiniBand in Kubernetes, InfiniBand on bare metal or virtual machines, Spectrum-X or NetQ-managed fabric, and Multi-Node NVLink partitions. Operators can also add an environment provider through the open provider interface.
Topograph does not treat topology as a one-time inventory exercise. It regenerates its view on request and when watched cluster changes occur, keeping schedulers aligned with the hardware and fabric relationships they are meant to manage.
Locality decides whether bandwidth becomes useful
Topograph maps which GPUs and network components sit near each other, helping schedulers favor the closest topology domain. GPUs inside one locality domain have short, high-bandwidth paths; traffic between domains crosses more shared links and switches, increasing contention and latency for tightly coupled workloads.
NVIDIA NVLink and NVLink Switch provide high-bandwidth, all-to-all scale-up connectivity within rack-scale GPU domains. NVIDIA Spectrum-X Ethernet supplies predictable, low-latency scale-out networking across systems and racks, while modern NVIDIA Quantum InfiniBand ports can reach up to 800 Gb/s.
The bandwidth figures inside GPU domains are even less subtle. Fifth-generation NVIDIA NVLink, used with NVIDIA Blackwell systems such as GB200 and GB300, provides 1.8 TB/s of bidirectional bandwidth per GPU; sixth-generation NVLink provides 3.6 TB/s per GPU with Vera Rubin. The NVLink Switch fabric is dedicated and non-blocking, giving each GPU its own lane instead of forcing it to share bandwidth under load.
Schedulers can place workloads efficiently only when they have a current and accurate view of those relationships. Slurm and Kubernetes already support topology-aware allocation, but they can act only on the topology they observe. Topograph supplies the missing view, allowing them to favor GPUs in the closest domain and avoid bottlenecks before a job starts paying for them.
Placement inside the orchestration stack
Within the NVIDIA DSX OS cluster orchestration layer, Topograph works alongside Dynamic Resource Allocation and KAI Scheduler. Its integration with KAI Scheduler enables topology-aware gang scheduling across AI factory infrastructure, keeping groups of connected resources together instead of scattering them across distant domains.
That focus on gang scheduling matters for distributed workloads, which benefit from communication locality rather than isolated GPU availability. A scheduler that finds enough GPUs but ignores their connections may satisfy a resource request while producing a slower and more expensive job.
Topograph’s two-part design keeps the toolkit adaptable across environments: providers discover and normalize topology, while engines publish the result in the configuration formats that schedulers and cluster systems expect. The practical result is not a new scheduling theory. It is a continuously refreshed map that lets existing workload managers make less wasteful decisions.
