Skip to content

Chapter 16: The GPU Compute Network -- Packet Anatomy End to End · Part 10 of 12

Act 9 -- How the fabric choice changes the packet: comparison matrix

Every fabric technology produces a different packet, different switch lookup, different congestion control, different latency profile. This act consolidates the full comparison.

Fabric technology comparison matrix
Select a dimension to compare all four fabric technologies.
End-to-end latency (2-hop)
RoCEv2 + Spectrum-X
~1.4-1.8 us
InfiniBand NDR
~0.6-0.8 us
Slingshot 11
~1.1-1.3 us
Gaudi 3 (100GbE)
~2.0-2.5 us
IB wins due to credit-based FC (no PFC jitter), O(1) LFT lookup, and per-packet adaptive routing eliminating collisions.

Layer-by-layer comparison

Layer 1 -- Physical

All modern AI fabrics use 400G OSFP with PAM-4 modulation. This is the same physical spec across ConnectX-7 RoCEv2, ConnectX-7 InfiniBand, Gaudi 3, and MI300X with Elba. Slingshot also uses 200G or 400G optical -- the physical signalling is compatible but the protocol above it is not.

The optics may differ: DGX H100 ships with QSFP-DD-to-OSFP breakout cables for some rack configurations; Frontier uses SMF-based optical links between Rosetta ASICs. These differences are at the DAC/transceiver level, not visible to any network protocol.

Layer 2 -- Data link

FabricL2 headerAddressingRewritten each hop?
RoCEv2 / Spectrum-XEthernet II (14B)48-bit MACYes -- full rewrite
InfiniBand NDRLRH (8B)16-bit LIDNo -- LID unchanged end-to-end
SlingshotEthernet + Slingshot extMAC + routing tagYes -- MAC rewritten

Layer 3 -- Network

FabricL3 headerAddressingLookup type
RoCEv2IPv4 (20B) + UDP (8B)32-bit IPECMP hash -> next-hop (IP route table)
InfiniBandGRH (40B, optional)128-bit GIDLFT array[DLID] = port -> O(1)
SlingshotIPv4 + Slingshot extIP + routing tagSlingshot routing table

Congestion control

FabricMechanismGranularityPropagation
RoCEv2PFC pause + ECN marks + DCQCN rate reductionPer priority classHop-by-hop (PFC) + end-to-end (ECN)
InfiniBandCredit-based flow control (per-VL per-port)Per virtual lanePer-link (no backward propagation)
SlingshotHRP credit-based + adaptive routingPer destinationPer-link (no PFC storms possible)

Load balancing

FabricMethodEntropy sourceOOO handling
RoCEv2 + SLBECMP hash5-tuple + optional QPairNone -- retransmit on OOO
RoCEv2 + RSHPPer-packet sprayAll flows treated independentlyRSHP reorder buffer in ConnectX-7
InfiniBandAdaptive routing (per-packet, congestion-aware)No hash -- real-time congestion measurementNative IB transport
SlingshotAdaptive routing (per-flow, per-credit)Flow tag in Slingshot headerHRP credits maintain order within flow

Latency benchmark (2-hop fabric, 128B message)

FabricMPI latencyNotes
InfiniBand NDR 400G~0.6 usCredit-based FC, per-packet adaptive routing, O(1) LFT lookup
RoCEv2 + Spectrum-X RSHP~1.4-1.6 usAdded UDP/IP overhead, PFC latency jitter, IP route lookup
Slingshot 200G~1.1-1.3 usCredit-based FC, adaptive routing, Slingshot header overhead
Gaudi 3 (100GbE RoCEv2)~2.0-2.5 us100G limit, standard Ethernet, no Spectrum-X RSHP

When would you choose each?

InfiniBand NDR remains the lowest-latency, highest-throughput choice for pure GPU training workloads where the cluster is homogeneous NVIDIA hardware. The premium is the proprietary switch stack (QM9700 vs commodity Ethernet) and the reliance on the Subnet Manager.

RoCEv2 over Spectrum-X is the choice when the cluster must support mixed workloads (GPU training + inference + storage on the same fabric), when EVPN multi-tenancy is needed, or when the organisation has strong Ethernet operational expertise and wants BGP-based routing with familiar tooling. The latency penalty (~1 us per hop) is acceptable for most AllReduce patterns.

Slingshot is the choice when building at scientific HPC scale (tens of thousands of nodes) where the traffic pattern is mixed HPC + AI and where the credit-based flow control is a hard requirement due to the scale of MPI collective operations.

Gaudi 3 (100GbE integrated) is the choice when the budget requires it or when the workload is inference-oriented (lower bandwidth requirements, more sensitivity to per-GPU cost than raw training throughput).