inference/research
verifiedhigh · 0.90updated 2026-10-05

AMD Instinct MI100 product page - Infinity Fabric link count and peak link bandwidth rows

spec-sheet

Fields

id
sources/w7a-amd-mi100-product-page-if-link-table
Publisher
AMD
Kind
spec-sheet
Accessed
2026-10-05
notes
Board specification table rows verbatim: 'Bus Type: PCIe 4.0 x16 , PCIe 3.0 x16', 'Infinity Fabric(TM) Links: 3', 'Peak Infinity Fabric(TM) Link Bandwidth: 92 GB/s'. CRITICAL: the row is labelled only 'Peak Infinity Fabric Link Bandwidth' - it does NOT say per-direction, and the 92 is BIDIRECTIONAL (2 x 46), not the GPU-to-GPU total. It is easily misread as a host figure because it sits in the same table as the 'Bus Type' PCIe row. Reconciles with the CDNA 2 white paper endnote MI200-43: MI100 has three links providing 276 GB/s peak theoretical P2P per card, and 276 / 3 = 92 exactly - so the 92 is per link, bidirectional, and the 276 is the 3-link total. Supported Technologies lists only 'AMD Infinity Architecture', with NO 'Nth Gen' label, so the MI100-era fabric cannot be given an AMD generation number from this page.

Notes

Board specification table rows verbatim: 'Bus Type: PCIe 4.0 x16 , PCIe 3.0 x16', 'Infinity Fabric(TM) Links: 3', 'Peak Infinity Fabric(TM) Link Bandwidth: 92 GB/s'. CRITICAL: the row is labelled only 'Peak Infinity Fabric Link Bandwidth' - it does NOT say per-direction, and the 92 is BIDIRECTIONAL (2 x 46), not the GPU-to-GPU total. It is easily misread as a host figure because it sits in the same table as the 'Bus Type' PCIe row. Reconciles with the CDNA 2 white paper endnote MI200-43: MI100 has three links providing 276 GB/s peak theoretical P2P per card, and 276 / 3 = 92 exactly - so the 92 is per link, bidirectional, and the 276 is the 3-link total. Supported Technologies lists only 'AMD Infinity Architecture', with NO 'Nth Gen' label, so the MI100-era fabric cannot be given an AMD generation number from this page.

Cross-references

This record references no other records.