AI 開發工具
NVIDIA Launches CUDA MCP and Self-Hosted Nsight Blueprint, Keeping Four-Model Backend On-Premises
Nsight AI can now connect live CUDA documentation to MCP clients such as Codex and Claude, or deploy a complete coding-assistance backend locally. The open-source Blueprint combines generation, completion, embedding, and reranking models, but the reference configuration requires a Hopper-class GPU and at least 200GB of storage.
NVIDIA has expanded Nsight AI, splitting CUDA coding assistance into three paths: the NVIDIA-hosted CUDA MCP Server, the self-deployable Nsight Copilot Blueprint, and a performance-analysis assistant embedded directly in Nsight Compute. The MCP endpoint can connect to Codex, Claude, or other compatible clients, allowing agents to search NVIDIA-maintained CUDA documentation and code examples instead of relying solely on knowledge available before a model’s training cutoff. The initial connection still requires authentication with an NVIDIA Developer account.
The more technically substantial offering is the self-hosted Blueprint, licensed under Apache 2.0. Its chat and RAG generation layer uses gpt-oss-120b NIM, while low-latency completion is handled by CUDA-autocomplete, which is based on Qwen2.5-Coder-7B. It also uses BAAI bge-m3 to create document embeddings and llama-nemotron-rerank-1b-v2 for reranking. The service layer comprises vLLM, LiteLLM, FastAPI, and Bodhi Tree RAG, exposing an OpenAI-compatible streaming interface and a `search_cuda_docs` MCP tool. The default Docker Compose configuration allows prompts, source code, and inference to remain local.
The hardware requirements are substantial. The reference deployment uses a DGX Spark with 128GB of unified memory. A conventional server requires one GPU with at least 80GB of memory to host the main model, plus a second GPU for completion, embedding, and reranking—or a single GPU with roughly 100GB or more of memory. The main model uses MXFP4 and requires Hopper or a newer architecture; the A100 is unsupported even if it has sufficient memory. At least 200GB of disk space is also required.
This release brings document retrieval, code generation, and profiler diagnostics into a single workflow, making it particularly useful for teams working with proprietary kernels. However, NVIDIA has not yet published complete ComputeEval results sufficient for comparison with general-purpose coding agents. AI recommendations also cannot replace compilation, numerical validation, or Nsight metrics. Although the Blueprint’s code is open, the downloaded NIMs, models, and CUDA documentation are each subject to separate licensing terms, and the project currently does not accept external code contributions.