Reclaiming the GB10 display carve-out: what 'more RAM on DGX Spark' actually buys a TP4 serving cluster
Two community techniques unlock the GB10’s idle 2 GiB display-reserved memory: a patched-driver kernel reclaim into Linux RAM (self-reported, no KV-pool benefit proven) and a DRM display-buffer route into CUDA (community-confirmed, +250k KV tokens but ~60-70% bandwidth). Neither helps glm-v53-flash-4x-noswitch today; verdict monitor-and-wait.