Things we measured ourselves, on hardware we actually run — GPUs, LLM serving, and the plumbing that connects them. No republished benchmarks, no vendor slides.
We hung an Intel Arc Pro B70 off a mini PC through a Thunderbolt enclosure and served a 30B model from it. Decode barely noticed the cable. Cold-loading a model was a different story — and the reason why is a 256 MB porthole most people never hit.