An architect is selecting a network fabric for a new GPU training cluster and wants a transport that natively provides credit-based flow control so that frames are not dropped under congestion, without relying on additional Ethernet extensions to achieve lossless behaviour. Which fabric technology meets this requirement by design?
- AStandard TCP/IP over 100 GbE, which retransmits any dropped segments to guarantee eventual delivery to the receiver
- BInfiniBand, which uses link-level credit-based flow control so a sender transmits only when the receiver has advertised buffer space Correct
- CPlain Ethernet with no extensions, which relies on large switch buffers to absorb traffic bursts and so avoid loss
- DFibre Channel, which is optimised to carry block storage traffic between server hosts and disk arrays
Why A is wrong: It is tempting because TCP does guarantee eventual delivery, but TCP achieves reliability through retransmission after loss rather than by preventing drops, so it is not lossless and its retransmissions add latency that throttles collective operations.
Why B is correct: InfiniBand implements credit-based flow control at the link layer, so a sender transmits only after the receiver advertises available buffer credits, making the fabric lossless by design without needing add-on Ethernet mechanisms.
Why C is wrong: It is tempting because deeper buffers reduce drops, but unextended Ethernet has no flow-control contract between sender and receiver, so once buffers fill, frames are still dropped, meaning it is not lossless by design.
Why D is wrong: It is tempting because Fibre Channel is also a credit-based, lossless fabric, but it is purpose-built for storage area networks and is not used to carry GPU-to-GPU collective communications in AI training clusters.