In enterprise networking, a 1ms latency spike is a footnote in a monitoring dashboard. In an AI training cluster running a large language model across 256 GPUs, that same 1ms spike causes every GPU in the job to sit idle -- waiting for one straggler. The network is not just moving data. It is the synchronisation mechanism for a massively parallel computation.
Here is how the requirements compare side by side:
| Dimension | Enterprise networking | HPC / AI networking |
|---|---|---|
| Latency target | 1-10ms acceptable | Sub-2 microseconds required |
| Packet loss | Rare, recoverable via TCP retransmit | Zero tolerance -- stalls entire job |
| Bandwidth | Shared, oversubscribed | Dedicated, non-blocking |
| Congestion handling | QoS queuing and selective drop | Prevented with lossless fabric |
| Switch role | Pure forwarding plane | Active compute participant (SHARP) |
| NIC role | Commodity, kernel-attached | RDMA co-processor, kernel-bypassed |
The right-hand column is not aspirational -- it is a hard requirement imposed by the way AI training works mathematically. The next two sections explain the math.