--- schema: 1 type: video title: >- Running LLMs locally: Practical LLM Performance on DGX Spark — Mozhgan Kabiri chimeh, NVIDIA description: >- Moving LLM workloads from the cloud to local infrastructure requires a shift in engineering strategy. In this talk, I share my journey of serving and benchmarking open-source models (1.5B to 14B) on an NVIDIA DGX Spark workstation. Using a reproducible methodology with vLLM, I analyze real-world trade-offs in throughput, latency, and the benefits of the 128GB Grace Blackwell unified memory architecture. You will leave with a clear framework for local model sizing, an understanding of quantization performance like NVFP4, and a guide for when local compute is the right choice for your AI stack. Speaker info: - LinkedIn https://www.linkedin.com/in/mozhgankch/ --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/c5-kx2bwoCk-running-llms-locally-practical-llm-performance-on-dgx-spark-mozhgan-kabiri-chimeh-nvidia Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/running-llms-locally-practical-llm-performance-on-d-a6c50fe8fa04 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/running-llms-locally-practical-llm-performance-on-d-a6c50fe8fa04.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T15:26:18.287Z' language: null children: [] videoId: c5-kx2bwoCk durationSeconds: 616 publishedAt: '2026-04-10T00:20:13.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 13 endSeconds: 104 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=13' heldReason: rights - ordinal: 1 text: null startSeconds: 13 endSeconds: 104 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=13' heldReason: rights - ordinal: 2 text: null startSeconds: 104 endSeconds: 114 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=104' heldReason: rights - ordinal: 3 text: null startSeconds: 114 endSeconds: 195 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=114' heldReason: rights - ordinal: 4 text: null startSeconds: 114 endSeconds: 195 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=114' heldReason: rights - ordinal: 5 text: null startSeconds: 195 endSeconds: 206 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=195' heldReason: rights - ordinal: 6 text: null startSeconds: 206 endSeconds: 318 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=206' heldReason: rights - ordinal: 7 text: null startSeconds: 330 endSeconds: 411 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=330' heldReason: rights - ordinal: 8 text: null startSeconds: 330 endSeconds: 411 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=330' heldReason: rights - ordinal: 9 text: null startSeconds: 411 endSeconds: 421 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=411' heldReason: rights - ordinal: 10 text: null startSeconds: 421 endSeconds: 520 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=421' heldReason: rights - ordinal: 11 text: null startSeconds: 520 endSeconds: 532 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=520' heldReason: rights - ordinal: 12 text: null startSeconds: 532 endSeconds: 616 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=532' heldReason: rights - ordinal: 13 text: null startSeconds: 532 endSeconds: 616 watch: 'https://liquidgold.cc/watch/c5-kx2bwoCk?t=532' heldReason: rights --- # Running LLMs locally: Practical LLM Performance on DGX Spark — Mozhgan Kabiri chimeh, NVIDIA Moving LLM workloads from the cloud to local infrastructure requires a shift in engineering strategy. In this talk, I share my journey of serving and benchmarking open-source models (1.5B to 14B) on an NVIDIA DGX Spark workstation. Using a reproducible methodology with vLLM, I analyze real-world trade-offs in throughput, latency, and the benefits of the 128GB Grace Blackwell unified memory architecture. You will leave with a clear framework for local model sizing, an understanding of quantization performance like NVFP4, and a guide for when local compute is the right choice for your AI stack. Speaker info: - LinkedIn https://www.linkedin.com/in/mozhgankch/ --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/c5-kx2bwoCk-running-llms-locally-practical-llm-performance-on-dgx-spark-mozhgan-kabiri-chimeh-nvidia Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/running-llms-locally-practical-llm-performance-on-d-a6c50fe8fa04 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:01 to 1:44) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:44 to 3:26) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (3:26 to 5:18) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (5:18 to 6:51) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (6:51 to 8:40) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (8:40 to 10:16) > Held: transcript rights do not permit this chunk's text.