AMD Infinity Fabric link (4th generation)

PER LINK, PER DIRECTION - THE DIRECTION AMBIGUITY IS NOW RESOLVED, 2026-10-03. AMD's CDNA 4 white paper (net2-amd-cdna4-whitepaper) states the convention explicitly for this exact family: 'The Infinity Fabric links in the AMD CDNA 4 architecture run up to 20% faster generationally at 38.4Gbps for a total link bandwidth of 76.8GB/s IN EACH DIRECTION', and endnote MI350-007 gives 'up to 153.6 GB/s peak total aggregate theoretical transport data GPU peer-to-peer (P2P) bandwidth per AMD Infinity Fabric link'. The 153.6 = 2 x 76.8 identity proves AMD's published per-link 'aggregate' is BIDIRECTIONAL and the per-direction figure is half. APPLYING THAT TO CDNA 3: the CDNA 4 white paper's own comparison sentence gives MI325X/MI300X (CDNA 3) 'up to eight AMD Infinity Fabric links providing up to 896 GB/s peak aggregate peer-to-peer (P2P) ... per GPU OAM module'. 896 / 8 links = 128 GB/s BIDIRECTIONAL PER LINK - which exactly matches the 'Peak Infinity Fabric Link Bandwidth: 128 GB/s' on AMD's own MI300X board table (link-amd-mi300x-specs). THE MI300X '128 GB/s' IS THEREFORE THE BIDIRECTIONAL AGGREGATE PER LINK, AND THE RECORDED PER-DIRECTION VALUE IS 128 / 2 = 64 GB/s, which is the value already stored here and now rests on two independent first-party confirmations instead of a guess. Cross-check the 7-link figure: AMD's MI350 ROCm page (net2-rocm-mi350-microarch) says the node uses '7 AMD Infinity Fabric links operating at 38.4 Gbps' to form a fully connected 8-GPU topology, and 7 x 64 GB/s per direction = 448 GB/s per direction per GPU against MI300X's 5.3 TB/s of HBM - a ratio of about 12x, versus roughly 7x on NVLink 4, which is the quantitative form of the inference consequence in notes. CDNA 4 PER-LINK FIGURES FOR THE NEXT GENERATION, both first-party: 76.8 GB/s per direction (153.6 bidirectional) per link, 16 bits wide at 38.4 Gbps, with 7 links giving 537.6 GB/s per direction and 1,075.2 GB/s aggregate P2P per OAM.

Die-to-die links inside a 3D-via Chiplet package, plus package-to-package links wiring up to 8 OAM modules into one platform. The MI325X platform integrates 8 fully connected MI325X OAM modules on an industry-standard OCP design via 4th-generation Infinity Fabric links, giving a fully-connected 8-way scale-up domain.

No switch part for scale-up; the domain is built from direct point-to-point links. Scale-out beyond the platform is handled by external networking, not by Infinity Fabric.

AMD's scale-up fabric and the closest analogue to NVLink in the x86 ecosystem. Inference consequence: an 8-way fully connected OAM platform gives collective bandwidth comparable in shape to an NVSwitch 8-GPU domain - no partial mesh, no bisection penalty for small groups - which makes 8-GPU tensor parallel viable without a switch part. DIRECTION AMBIGUITY, NOW RESOLVED 2026-10-03 (was contested; both readings were previously retained): AMD's MI300X page gives 128 GB/s peak Infinity Fabric link bandwidth with no direction stated (link-amd-mi300x-specs); under the bidirectional reading the per-direction value is 64 GB/s, under a per-direction reading it is 128 GB/s (link-amd-mi300x-specs, same source). RESOLUTION: AMD's CDNA 4 white paper states the convention directly - 76.8 GB/s 'in each direction' against a 153.6 GB/s 'peak total aggregate' per link - and its own CDNA 3 comparison figure of 896 GB/s aggregate P2P across 8 links reproduces the MI300X board table's 128 GB/s exactly, confirming 128 is BIDIRECTIONAL. So 64 GB/s per direction is correct and the record is no longer contested. A 64 GB/s per-direction Infinity Fabric link is roughly 7x below NVLink 4 per direction, so multi-GPU inference bandwidth headroom on 8-way MI300X is materially tighter than on 8-way H100 and pipeline-parallel sharding is the more natural fit than very wide tensor parallel. MI300X also exposes a 256 MB last-level cache, which changes how much traffic reaches the fabric at all. No separate xGMI record was created: no primary AMD page retrieved stated an xGMI rate or generation, and rather than invent one, note that xGMI is AMD's die-to-die link and its published rates belong to the Infinity Fabric family pages.