Synthesis history

Two substantive new items this pass: a Nexus paper proposes proactive intra-GPU PD disaggregation — performing the prefill/decode split within a single GPU through temporal scheduling rather than across separate nodes —…

One substantive new development: application-layer practitioners working with Fable 5 discovered that rendering text-heavy context as PNG images cuts inference costs by roughly 60% by exploiting fixed vision-token prici…

Two substantive additions this pass. First, attention-FFN disaggregation (AFD) has moved from a named-but-thin topic to a well-supported research area: AiDE, CMU operator-level disaggregated serving, derivations of opti…

LLM inference is being carved up through three successive hardware-level splits: phase disaggregation (routing prefill and decode to separate GPU pools), layer disaggregation (routing attention and feed-forward computat…