Compute Fabric for Edge AI Workloads

Mechanism

A distributed compute mesh for adaptation workloads that are too large for a single edge box and too small to justify a datacenter contract. The substrate intuition comes from Infinity (US Prov. 64/055,093): plesiochronous gradient consensus across heterogeneous accelerators, with elastic gradient-tensor buffering and a hysteretic phase-correction sideband signaling path electrically distinct from the high-speed serial data interface. This research line extends that substrate intuition from the rack-scale Infinity primitive to a mesh-scale topology over commodity heterogeneous edge nodes, where node membership is fluid and clocks are independent rather than co-located.

Why this matters

• The middle scale (roughly 4 to 32 heterogeneous edge nodes) is structurally underserved. Hyperscaler training fabrics assume homogeneous nodes and uniform interconnect. Cluster managers like Slurm assume cooperative tenancy and a single administrative domain. The mesh-scale design point sits in between, and neither end of the existing tooling stack is shaped for it.
• Companion to Infinity, which provides the apparatus-level synchronization primitive. Compute-fabric-mesh extends the same primitive to a topology where node membership is fluid (nodes join and leave during a training run) and trust assumptions are weaker (rented-pool rather than captive-rack economics).
• Status is Research, not pending. A filing decision is deferred until the topology layer has empirical validation distinct from Infinity's rack-scale substrate. We do not want to dilute the Infinity claim set with a topology variant that has not yet earned its own anchor.

Status and what's next

Active research. Topology characterization is in progress: how the plesiochronous consensus discipline behaves when the node set is heterogeneous in accelerator vendor, clock quality, and link latency, and when the membership set itself drifts during a single adaptation pass. Honest disclosure: there is no published benchmark on this surface at the time of writing. Expected output is a workshop preprint or a reference implementation within the next 12 to 18 months, at which point the filing question will be re-opened with concrete numbers in hand.