Antirez asked what DGX Station can do when large-model inference spans RAM and VRAM. NVIDIA advertises up to 748 GB of coherent memory, but that capacity is split across two very different physical tiers.

The official specification gives the GB300 GPU up to 252 GB of HBM3e at 7.1 TB/s and the Grace CPU up to 496 GB of LPDDR5X at 396 GB/s. The GPU can address both through the coherent memory system. Coherence makes the larger pool usable; it does not give LPDDR5X the bandwidth of HBM. NVIDIA’s own memory guidance recommends unified memory when the working set exceeds HBM and warns against making it the default for every allocation.

A useful result from DwarfStar or another runtime needs more than “the Q2/Q3 quant fits” and one tokens-per-second number. I would want the exact model and quant, weight bytes, prompt and output lengths, concurrency, HBM and LPDDR5X residency, offloaded layers or tensors, KV-cache placement, time to first token, prefill rate, decode rate, and end-to-end latency. A same-checkpoint sweep across several spill fractions would show where the slowdown begins.

The 748 GB pool may make a previously impossible single-node run usable. That is a real result. Capacity alone cannot predict speed; the performance claim starts with where the active bytes live and how often inference has to cross the memory boundary.