AMD Infinity Fabric scale-up link (CDNA3: 128 GB/s per direction)

Fields

id

interconnect/amd-infinity-fabric-link-cdna3

Kind

xgmi

Version

CDNA3 (MI300X/MI325X)

Bandwidth (GB/s, per link, unidirectional)

128

Bandwidth basis

Per the MI300X product page: 'Peak Infinity Fabric Link Bandwidth 128 GB/s'. The CDNA4 whitepaper confirms CDNA3 links provide 'up to 896 GB/s peak aggregate peer-to-peer (P2P)' across 8 links, and '1,024 total peak aggregate theoretical GPU transport rate bandwidth'. 896 / 7 active links = 128 GB/s per link per direction. The 1024 figure includes protocol overhead. Per-direction bandwidth is 128 GB/s; bidirectional aggregate is 256 GB/s per link.

Link count

7

Topology

Fully connected 8-GPU mesh (each GPU connects to 7 others). No switch. 7 links per GPU for GPU-to-GPU communication. 8th link in the board spec is for scale-out (external). Scale-up bandwidth: 128 GB/s per direction per link.

Scale-up

yes

Scale-out

yes

Switching

None - direct point-to-point links between GPUs

notes

CDNA3 Infinity Fabric links operate at 128 GB/s per direction. The MI300X product page lists 'Infinity Fabric Links: 8' and 'Peak Infinity Fabric Link Bandwidth: 128 GB/s'. The ROCm microarchitecture reference states 'seven high-bandwidth, low-latency AMD Infinity Fabric links' for the 8-GPU node. The 8th link is for scale-out (external connectivity). Total aggregate P2P bandwidth: 896 GB/s (7 links x 128 GB/s). Total peak aggregate theoretical transport rate: 1024 GB/s (includes protocol overhead). This is the scale-up network for tensor parallelism and pipeline parallelism within a single 8-GPU node.

way

both

Notes

CDNA3 Infinity Fabric links operate at 128 GB/s per direction. The MI300X product page lists 'Infinity Fabric Links: 8' and 'Peak Infinity Fabric Link Bandwidth: 128 GB/s'. The ROCm microarchitecture reference states 'seven high-bandwidth, low-latency AMD Infinity Fabric links' for the 8-GPU node. The 8th link is for scale-out (external connectivity). Total aggregate P2P bandwidth: 896 GB/s (7 links x 128 GB/s). Total peak aggregate theoretical transport rate: 1024 GB/s (includes protocol overhead). This is the scale-up network for tensor parallelism and pipeline parallelism within a single 8-GPU node.

Cross-references

This record references no other records — it is a leaf. It is cited by 1.