DECODE

An Upscale technical blog — written by engineers, for engineers. Deep dives on fabrics, silicon, network OS and the tooling behind AI clusters.

Why scale-up fabrics need load/store semantics, not packets

What changes when accelerators share memory over a fabric instead of exchanging messages, and why that decides your scale-up topology.

A walkthrough of sender-based and receiver-based congestion control in UET and when each one helps an all-reduce.

How we expose per-queue occupancy and PFC pause counters at sub-second cadence from a SONiC fabric, and what to do with them.

Reach budgets, FEC overhead and power per bit for 224G lanes, and how they decide copper versus optical links in rack-scale designs.

A pull-request workflow for switch configuration with validation, diffing and staged rollout, built on tools your platform team already runs.

How priority flow control pause propagates through a lossless fabric, how to read the counters, and the watchdog settings that stop the spread.