[DOC]: B-Series Compatibility

[swtb3-ryder]

How would you describe the priority of this documentation request

Medium

Describe the future/missing documentation

The readme mentions:

⚠️ Important B200 Limitation Notice:

B200 GPUs are not supported for the following advanced features:

Self-Reflection to improve accuracy
Query rewriting to Improve accuracy of Multi-Turn Conversations
Image captioning support for ingested documents
NeMo Guardrails for guardrails at input/output
VLM based inferencing in RAG
PDF extraction with Nemoretriever Parse
For these features, please use H100 or A100 GPUs instead.

Why is this the case? With the B-series being quite new...it would be useful for the community to understand why these new GPUs are incapable of these tasks. Also, is it just B200? or are other B-series/GB-series systems also affected?

Where have you looked?

No response

Code of Conduct

• I agree to follow rag's Code of Conduct
• I have searched the open documentation issues and have found no duplicates for this bug report

[onestardao]

Yeah… this B200 limitation definitely raises eyebrows.
We’ve been benchmarking across multiple RAG pipelines (multi-turn, PDF extraction, vision-augmented input) — and the odd part is: some lower-tier GPUs handle these better, depending on the I/O bottlenecks and driver stack.

My gut says this isn’t a raw compute problem. It’s likely a firmware or ops-layer decision, maybe tied to how the B-series exposes certain instructions / memory segmentation via TensorRT or NVSwitch. Either way, the doc should clarify:

• Is it B200 only or all B-series (B40, GB200, etc)?
• Are these limitations dynamic (driver-dependent) or architectural?
• Will newer SDK versions eventually unlock this?

Also worth noting: In local LLM tasks (e.g. Ollama w/ code search), even A100s sometimes beat B200 in latency + memory locality for smaller models.

🔍 Keywords for future searchers: B200, B-series GPU, multi-turn conversation, query rewriting, self-reflection, PDF extraction, RAG + GPU, guardrails, vision transformer, minotorwire, NVIDIA documentation gap.

TL;DR: This is the kind of footnote that breaks an entire RAG stack in prod.

Would love to help test if someone opens the architecture specs.