> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/sdgm/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/sdgm/_mcp/server.

# What are the context-size limits for Kumo Tabular and Kumo Relational?

> Understand GPU memory requirements and the separate SDM client context cap

**Neither model has a fixed model-level context-size limit.** In the standalone SDM library, practical capacity depends on available GPU memory. Inference currently runs on a single GPU.

For **Kumo Tabular**, approximate peak GPU memory as:

```text
512 bytes × number of cells
+ input memory and temporary allocations
```

As a rough reference, about **100 million cells** can fit on a single H100. Actual capacity depends on input shape and memory overhead.

For **Kumo Relational**, memory use also depends on the related-table context. There is no equivalent sizing formula available; measure peak memory with representative data to determine what fits.

The **SDM client currently caps context at 10,000 rows** for both models when sending context to an inference server. This client constraint is separate from model capacity. See the [client limits documentation](https://github.com/NVIDIA/structured-data-models-client/blob/main/docs/reference/limits.md).