This is the most operationally significant UEC design decision: the fabric does not need to be lossless. UEC endpoints handle reliability through retransmission.

The mechanism is selective acknowledgement (SACK). The receiver sends ACK packets back to the sender. An ACK carries:

• A cumulative acknowledgement offset: all bytes up to this offset have been received
• A SACK bitmask: individual packets above the cumulative ack that have been received out of order

The sender maintains a retransmit timer per outstanding message. If it does not receive an ACK advancing the cumulative ack within the timeout, it retransmits the missing segment(s) identified by the SACK bitmask.

This is structurally similar to TCP SACK but implemented at the NIC level in RDMA context -- no host CPU involvement, no kernel stack, no socket buffer. The NIC firmware manages the retransmit timer, buffers outstanding packets, and handles the ACK/retransmit state machine.

What packet loss actually costs under UEC

A retransmit adds latency equal to: RTT (round-trip time) + retransmit timer granularity. For a BasePOD with 2us hop-to-hop latency, an RTT from GPU to storage switch and back is approximately 4us. If the retransmit timer fires at the end of a 10us interval (a typical setting for lightly loaded fabrics), a single packet loss costs approximately 10-14us of added latency per affected message.

Compare this to RoCEv2: a dropped packet causes the QP to enter ERROR state, which requires application-level intervention (NCCL detects the failure, resets the QP, retransmits the AllReduce message). The recovery time is measured in milliseconds, not microseconds.

UEC's retransmit model is much more graceful: individual packet loss is handled silently by the NIC without application involvement. Only sustained loss (multiple retransmit timeouts, exceeding a configurable threshold) escalates to the application layer.

Implications for fabric design

Because reliability is end-to-end rather than fabric-guaranteed, UEC deployments can:

• Run without PFC on the Ethernet fabric. Standard best-effort queuing is sufficient.
• Use ECN marking for congestion feedback (keeping the switch-level signal), but the response mechanism is different (see Act 5).
• Tolerate occasional packet drops from hardware failures, brief congestion events, or even lossy wireless paths -- none of these cause application-visible failures.

This does not mean you want a lossy fabric. High loss rates mean frequent retransmits, which means added latency and wasted bandwidth. But a well-designed fabric with 0.001% drop rate produces negligible retransmit overhead while completely eliminating PFC configuration and its associated operational risks.