Index

Catalog of the wiki. The LLM keeps this up to date.

Sources

• exolabs-dgx-spark-handbook: EXO Labs’ guide to local LLMs on DGX Sparks; the model for Nima’s article.

Concepts

• memory-bandwidth: how fast a GPU reads its memory, why that sets writing speed, and how cloud batching trades speed per user for cost.

Entities

• rtx-5090: beast’s GPU; 32 GB, 1,792 GB/s.
• dgx-spark: NVIDIA’s small AI box; 128 GB, 273 GB/s.
• h100: data center GPU (Hopper); 80 GB, about 3,350 GB/s.
• b200: data center GPU (Blackwell); 192 GB, about 8,000 GB/s.

Lessons

(none yet)