Networking Hardware for AI: NICs, DPUs, Switches and Optics
For network and infrastructure engineers who must specify, cable and validate the NICs, DPUs, switches and optics of GPU cluster fabrics at 400G, 800G and 1.6T.
Level
Intermediate
Study time
14 hours
Lessons
21 in 6 modules
Exam
20 questions
About this course
Collective communication across thousands of GPUs turns the network into part of the computer, and most of what goes wrong in an AI fabric is physical: the wrong optic, a marginal link, a mis-split cage, an oversubscribed rail or a NIC on the wrong PCIe root. This course teaches the hardware of AI networks from the SerDes up. It starts with lane rates and encodings (PAM4 at 100G and 200G per lane), the difference between scale-up, scale-out and front-end networks, and how NICs connect to GPUs over PCIe and through PCIe switches.
You then work through the device classes current in October 2026: NVIDIA ConnectX-8 and ConnectX-9 SuperNICs, BlueField-3 and BlueField-4 DPUs, AMD Pensando Pollara 400 and Vulcano, and Broadcom Thor Ultra; InfiniBand (Quantum-X800) and Ethernet switch silicon (Spectrum-X, Broadcom Tomahawk 5 and 6 and Jericho, Cisco Silicon One) in the systems sold by NVIDIA, Arista, Cisco and others; and the Ultra Ethernet 1.0 specification. The optics module covers OSFP and QSFP-DD form factors, DR, FR and SR optics, DACs, ACCs and AECs, linear-drive optics and co-packaged optics such as Spectrum-X Photonics and Quantum-X Photonics. The course closes with topology and radix arithmetic, cabling and breakout planning, and link bring-up and validation with real tools. Shipping status is stated as of October 2026 and announced-but-not-shipping products are labeled. The course ends with an automated exam and a Wirewalk Institute certificate of completion and assessment; it is not a vendor certification.
What you will be able to do
Distinguish scale-up, scale-out and front-end networks and explain which hardware belongs in each
Explain SerDes lane rates, PAM4 signaling, FEC and their effect on port speeds, reach and latency
Compare NICs, SuperNICs and DPUs, including ConnectX-8 and ConnectX-9, BlueField-3 and BlueField-4, Pollara and Thor Ultra, and place them correctly in a GPU server's PCIe topology
Compare InfiniBand and Ethernet switch platforms for AI, including Quantum-X800, Spectrum-X and Tomahawk-based systems, and describe what Ultra Ethernet changes
Select transceivers, cables and breakouts for a given reach, port speed and power budget, and explain the trade-offs of linear-drive and co-packaged optics
Calculate switch radix, tier count, oversubscription and optic count for a rail-optimized GPU cluster
Bring up and validate links and fabrics using port counters, optic diagnostics and RDMA bandwidth tests
Syllabus
6 modules · 21 lessons · about 14 hours
1
Module 1: AI Network Fundamentals at the Physical Layer
Why AI fabrics differ from data center networks, and the signaling that sets port speeds.3 lessons
Scale-Up, Scale-Out and Front-End: Three Networks in an AI Cluster
SerDes, PAM4 and FEC: From 100G to 200G per Lane
RDMA Hardware Basics: RoCEv2, InfiniBand Verbs and GPUDirect RDMA
2
Module 2: NICs, SuperNICs and DPUs
The adapters that connect GPUs to the fabric, and where they sit in the server.4 lessons
ConnectX-8 and ConnectX-9 SuperNICs: Ports, PCIe Switching and Congestion Control
DPUs: BlueField-3, BlueField-4 and Infrastructure Offload
AMD Pensando Pollara and Vulcano, Broadcom Thor Ultra and Other Ethernet AI NICs
NIC Placement: PCIe Roots, GPU Affinity and One NIC per GPU
3
Module 3: Switch Silicon and Systems
InfiniBand and Ethernet switch ASICs, the systems built on them, and the standards shaping AI Ethernet.4 lessons
InfiniBand for AI: Quantum-2 and Quantum-X800
Ethernet Switch ASICs: Spectrum-X, Tomahawk 5 and 6, Jericho and Silicon One
Ultra Ethernet 1.0: Transport, Congestion Control and What Hardware Must Support
Switch Hardware Anatomy: Radix, Buffers, Cages, Power and Cooling
4
Module 4: Optics and Cabling
Form factors, optic types, copper options and the move toward linear-drive and co-packaged optics.4 lessons
OSFP and QSFP-DD: Form Factors, Cage Splitting and Breakouts
Optics by Reach: SR, DR, FR and LR at 400G, 800G and 1.6T
DAC, ACC and AEC Copper: Reach, Power and Rack Layout
Linear-Drive and Co-Packaged Optics: Spectrum-X Photonics, Quantum-X Photonics and Tomahawk CPO
5
Module 5: Topology and Capacity Planning
Rail-optimized fat trees, radix arithmetic and bills of materials for real cluster sizes.3 lessons
Rail-Optimized Fat Trees and Multi-Plane Designs
Radix, Tiers and Oversubscription: Worked Arithmetic
Building the Bill of Materials: Switches, Optics, Fibers and Spares
6
Module 6: Bring-Up, Validation and Troubleshooting
Getting a fabric from cabled to trusted, and finding the bad link before a training job does.3 lessons
Link Bring-Up: Port Mapping, Optic Diagnostics and Error Counters
Fabric Validation: ib_write_bw, perftest and NCCL Tests
Troubleshooting Flapping Links, BER, Congestion and Mis-Cabling
Exam and certificate
The exam unlocks once you have completed every lesson. It draws 20 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 45 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.
Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.
The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by NVIDIA, Broadcom, AMD, Cisco, Arista, Intel, Ultra Ethernet Consortium, IEEE, OSFP MSA, or QSFP-DD MSA. Product and company names are trademarks of their respective owners.
A plain-English, one-hour walk through an AI data center: what a model asks of hardware, what is inside a GPU server, how GPUs talk over NVLink, InfiniBand and Ethernet, where the data lives, and why AI racks need so much power and cooling.
A plain-language tour of the hardware that runs AI: why AI is limited by compute, memory bandwidth and data movement; what GPUs, HBM and precision formats are; how servers, networks, storage, power and cooling fit together; and how to size a simple inference or training system and weigh cloud against buying.
An engineer's tour of the modern data center GPU: how SMs and CUs, warps and wavefronts, tensor and matrix cores and the on-chip memory hierarchy work; what FP64 through FP4 and sparsity really buy you; how HBM, PCIe, NVLink, Infinity Fabric, UALink and CXL move data; and how power, cooling and reliability features shape real deployments. Reviewed against vendor sources as of October 2026.