The Latest in GPU, TPU, NIC and Other Device Support — WG Device Management#

John Belamaric, Google & Patrick Ohly, Intel

• Why: DRA is the substrate under both talks above, and this is its status report from the working-group co-chairs. NICs are named as a first-class device type alongside GPUs and TPUs, which is the part that matters for fabric work.
• What: DRA reached General Availability in 1.34; this covers what shipped in 1.35 and what is targeted for 1.36.
• How: The named post-GA workstreams are managing device failures, groups of devices that work together across nodes, and controlled sharing of devices. Cross-node device groups are the notable one — that is a topology-aware allocation problem, not a node-local one.
• Where applicable: Anyone on 1.34+. The cross-node grouping work is most relevant if you are building multi-node accelerator or NIC pools.