Skip to content

Chapter 0: How We Got Here -- The Hardware Story · Part 5 of 15

Act 3B -- What AMD does instead: Infinity Fabric and xGMI

The DGX H100 uses NVLink and NVSwitch for intra-node GPU communication. This is the NVIDIA solution -- proprietary, fast, and deeply integrated with the CUDA software stack. But when you encounter AMD Instinct GPUs in your infrastructure -- MI300X, MI325X, or the newer MI350 series -- the intra-node interconnect works on a fundamentally different architecture. Understanding both is increasingly necessary, because enterprise AI deployments are no longer uniformly NVIDIA.

Ch0 · Hardware Foundations

Intra-node interconnect: NVIDIA vs AMD

This compares scale-up bandwidth inside one GPU server. The external Ethernet fabric remains the same operational problem across both vendors.

Topology shape
NVSx4G0G1G2G3G4G5G6G7
NVSwitch hub
Scale-up bandwidth
DGX H100900 GB/s
MI325X336 GB/s
MI350X1075 GB/s
DGX B2001800 GB/s
Selected platform
Memory / GPU80 GB HBM2e
Memory bandwidth3.35 TB/s
Scale-up topologyNVSwitch hub
External NICConnectX-7 IB/RoCEv2

18 NVLink 4.0 links per GPU routed through NVSwitch crossbars.

External fabric view

Switches still see the same kind of scale-out problem: 400GbE ports, RoCEv2 transport, PFC-protected lossless queues, and ECN feedback. The switch does not care whether traffic originated from ConnectX-7 or Pollara 400.

Same DSCP / QoS intent on Ethernet fabricsSame RoCEv2 operational tuningUALink is AMD's future multi-node scale-up direction

The Infinity Fabric stack

AMD's answer to NVLink is Infinity Fabric, a family of interconnect technologies that spans everything from die-to-die communication inside a single chip package to server-to-server links across a rack. This is not a single protocol -- it is a layered architecture:

Infinity Fabric (intra-package) handles communication between chiplets within a single GPU die stack. AMD's Instinct GPUs use a chiplet design: multiple Accelerator Complex Dies (XCDs) are bonded together on an I/O Die (IOD). The Infinity Fabric between these chiplets provides the high-bandwidth, low-latency path that keeps the chiplets behaving like a single coherent GPU to the software above.

xGMI (External Global Memory Interconnect) is the Infinity Fabric variant that connects multiple GPU packages together within a single server. This is the functional equivalent of NVLink -- it is the multi-GPU scale-up interconnect that allows 8 GPUs in a server to communicate directly without going through PCIe or the host CPU. Like NVLink, xGMI creates a fully-connected mesh topology where every GPU can directly access every other GPU's memory.

Comparing the generations: MI325X, MI350X, and what changes

The interconnect specs shift meaningfully across AMD generations. As a network engineer, these numbers matter because they tell you the bandwidth budget available for tensor parallelism within a node -- and therefore what spills out onto the external fabric as data parallelism traffic.

AMD Instinct · Generational comparison

MI300X -> MI325X -> MI350X

The important point for a network engineer is how AMD's node-local scale-up bandwidth changes across generations, because that changes what spills onto the external fabric.

Chiplet layout
I/O DieInfinity Fabric backboneXCD0XCD1XCD2XCD3XCD4XCD5XCD6XCD7
xGMI progression
MI300X288 GB/s
MI325X336 GB/s
MI350X1075 GB/s
MI355X1075 GB/s
Selected generation
Memory256 GB HBM3E
Memory bandwidth6 TB/s
xGMI links8 links
Aggregate xGMI336 GB/s
Intra-package Infinity Fabric~3.5 TB/s
TDP1000W

Memory increases and xGMI connectivity improve, but it is still fundamentally a CDNA 3-era platform.

NVIDIA context

H100 NVLink 4.0 is 900 GB/s, B200 NVLink 5.0 is 1,800 GB/s, and MI350X xGMI Gen4 is about 1,075 GB/s. That means MI350 closes the H100-era gap meaningfully, but Blackwell has already moved the target again.

The external fabric lesson stays unchanged: all of these nodes still scale out over RoCEv2-capable Ethernet.

MI325X (CDNA 3, shipping 2024--2025)

The MI325X is a CDNA 3 architecture GPU using 8 XCD chiplets on an OAM (OCP Accelerator Module) form factor. Each MI325X GPU carries 256 GB of HBM3E memory with 6 TB/s memory bandwidth.

For multi-GPU communication, the MI325X uses xGMI with 8 links. Each xGMI link runs at approximately 64 GB/s raw, yielding around 48 GB/s practical per-link throughput after protocol overhead and CRC correction. Across all 8 links to its 7 peer GPUs, an MI325X achieves roughly 336 GB/s aggregate peer bandwidth -- though workload throughput is capped by the slowest link in the mesh, giving practical AllReduce bandwidth of approximately 310--330 GB/s per GPU in production.

An 8-GPU MI325X platform (Universal Baseboard 2.0 / UBB 2.0) provides:

  • 2 TB total HBM3E capacity per node
  • 48 TB/s aggregate memory bandwidth across the node
  • Fully-meshed xGMI topology between all 8 GPUs

MI350X / MI355X (CDNA 4, shipping 2025)

The MI350 series represents a significant generational jump. AMD rebuilt Infinity Fabric for CDNA 4, delivering 5.5 TB/s of intra-package interconnect bandwidth between dies -- internal bandwidth that the MI325X did not have at this level.

The inter-GPU (xGMI) link count increases from 8 to 7 Gen4 links, but the per-link bandwidth increases substantially, yielding 1,075 GB/s aggregate GPU-to-GPU bandwidth per GPU -- more than 3x the MI325X figure. Each MI350X/MI355X carries 288 GB of HBM3E with 8 TB/s memory bandwidth.

An 8-GPU MI350 platform provides:

  • 2.3 TB total HBM3E capacity per node (MI350X) / 2.3 TB (MI355X)
  • 64 TB/s aggregate memory bandwidth across the node
  • Redesigned fully-meshed Infinity Fabric topology

The MI355X (liquid-cooled) variant runs at up to 1,400W TDP, compared to MI350X at 1,000W and MI325X at 1,000W.

The parallel to NVLink generation jumps

If this pattern sounds familiar, it should. The NVIDIA generational jumps follow the same rhythm:

GenerationNVIDIA Scale-UpAMD Scale-Up
H100 eraNVLink 4.0 - 900 GB/s intra-nodeMI300X/MI325X - ~336 GB/s (xGMI)
B200 eraNVLink 5.0 - 1.8 TB/s intra-nodeMI350 - ~1,075 GB/s (xGMI Gen4)

The bandwidth gap between NVIDIA and AMD at intra-node scale-up is real and significant. NVLink 4.0 at 900 GB/s versus MI325X xGMI at ~336 GB/s means tensor parallelism within a node is substantially faster on a DGX H100 than on a comparable MI325X node. This has direct consequences for model partitioning strategy: workloads that rely heavily on intra-node tensor parallelism may need to restructure their parallelism plan on AMD hardware.

The MI350 generation closes this gap meaningfully. At ~1,075 GB/s aggregate, MI350X node-level bandwidth approaches NVLink 4.0 territory, though NVLink 5.0 in Blackwell has already moved the goalposts to 1.8 TB/s.

Scale-out: the external fabric perspective

Where the architectures converge is on the external fabric -- the network that connects nodes to each other. Both NVIDIA and AMD platforms use standard Ethernet for scale-out:

  • NVIDIA DGX nodes: ConnectX-7 NICs running InfiniBand NDR or RoCEv2
  • AMD MI325X/MI350X nodes: AMD Pollara 400 NICs (or OEM equivalents) running RoCEv2 or UEC (Ultra Ethernet Consortium)

The AMD Pollara 400 NIC is AMD's scale-out networking solution, targeting the Ultra Ethernet Consortium specification as an alternative to NVIDIA's Spectrum-X ecosystem. From a fabric perspective, AMD-GPU nodes look essentially the same as NVIDIA-GPU nodes at the external switch: 400GbE ports, RoCEv2 traffic, DSCP marking, PFC and ECN requirements. Everything you configure on the Ethernet fabric for an NVIDIA cluster applies unchanged to an AMD cluster.

This is why ProDeploy's PFC, ECN, congestion control, and switch configuration content is directly applicable regardless of which GPU vendor's hardware populates your compute nodes. The fabric does not care whether the RDMA traffic originates from a ConnectX-7 or a Pollara 400 -- it sees UDP/4791 with DSCP 26, and it manages it the same way.

UALink: AMD's answer to NVLink Networking

NVLink Networking (Chapter 12) extends NVLink beyond a single chassis to connect multiple DGX nodes at the scale-up interconnect level. AMD is developing an analogous open-standard capability through the UALink consortium (Ultra Accelerator Link).

UALink 1.0 defines a 200 GT/s per-lane interconnect that can connect up to 1,024 accelerators in a pod, using 4 lanes per GPU for 800 GB/s total connection bandwidth. This is an industry consortium standard rather than a proprietary NVIDIA technology -- AMD, Google, AWS, Intel, Microsoft, Cisco, and others are all consortium members.

UALink hardware (switches, GPU implementations) is expected in the 2026/2027 window. For production deployments today, AMD scale-up remains xGMI within a single 8-GPU node, with RoCEv2 for all cross-node scale-out. UALink is important to understand as the technology that AMD intends to use when NVLink Networking-style multi-node scale-up is needed.

What changes for your diagnostic work

If you encounter an AMD Instinct cluster, the external fabric diagnostics remain almost identical to what this course teaches. The internal node diagnostics use different tools:

# AMD equivalent of nvidia-smi for GPU health
rocm-smi

# Check xGMI link topology
rocm-smi --showtopo

# Check xGMI link bandwidth state
rocm-smi --showtopoweight

# AMD equivalent of ibstat for RDMA state
# AMD nodes still expose RDMA interfaces via rdma-core
ibstat
rdma dev show

# RCCL (AMD's NCCL equivalent) test for AllReduce performance
# Replace nccl-tests with rccl-tests on AMD nodes
rccl-tests/build/all_reduce_perf -b 8 -e 8G -f 2 -g 8

The rocm-smi --showtopo output will display the xGMI mesh weight between GPU pairs. A weight of 15 indicates a direct xGMI link. A weight of 0 indicates the GPUs are on separate sockets with no direct link. Any weight above 15 indicates a longer path with intermediate hops -- which will degrade collective operation performance in the same way a congested external fabric hop degrades NCCL AllReduce.

The ROCm stack (AMD's equivalent of CUDA + MLNX-OFED) is installed separately. RCCL (ROCm Collective Communication Library) is the AMD equivalent of NCCL. Its tuning variables, collective algorithm selection, and performance measurement workflow mirror NCCL closely -- Chapter 8's NCCL discussion maps directly to RCCL with minor naming differences.


The practical summary: AMD Instinct GPUs use Infinity Fabric / xGMI for intra-node scale-up instead of NVLink. The MI325X delivers ~336 GB/s per-GPU aggregate peer bandwidth (versus 900 GB/s on H100 NVLink 4.0). The MI350 series closes much of this gap at ~1,075 GB/s. For the external fabric -- the network you configure and operate -- both platforms are indistinguishable: both run RoCEv2 over 400GbE, both require identical PFC/ECN configuration, and both generate the same AllReduce traffic pattern at the switch level.