Guides & Resources — AI Networking & Fabrics
How to Read a ConnectX Part Number: MCX4121A-ACAT to MCX75310AAS-NEAT Decoded
The letter after the dash is the speed: X 10GbE, A 25GbE, C 100GbE, E EDR/HDR100, H 200G, N 400G NDR. The full ConnectX-4 to ConnectX-7 decoder, the pairs buyers confuse, and a 16-card lookup table with prices.
Why Your 200G ConnectX-6 Tops Out Near 100G in a Dell R740xd or R640 (PCIe Gen3 Explained)
A PCIe 3.0 x16 slot carries about 100-110 Gb/s of real traffic, so a 200G ConnectX-6 in an R740xd or R640 runs at half speed. The math, card-by-card effects, how to check your slot, and what to buy for Gen3, Gen4 and Gen5 servers.
ConnectX Card Showing ib0 Instead of Ethernet? How to Switch a VPI Card With mlxconfig
A ConnectX VPI card that shows ib0 or won't link to an Ethernet switch is set to InfiniBand. Set LINK_TYPE_P1=2 with mlxconfig and reboot. Step-by-step, VPI vs EN part numbers, and the cable that goes with each mode.
HDR100 Explained: Two ConnectX-6 Nodes per QM8700 Port, and When You Need Full HDR
HDR100 is 100Gb/s InfiniBand on two of an HDR port's four lanes, so a 40-port QM8700 can serve up to 80 nodes. How it differs from EDR, the switch setup, the parts list, and when full 200G HDR is worth it.
NVLink 6: The Switch That Turns a Rack Into One GPU
NVIDIA's NVLink 6 switch moves 3.6 TB/s to every GPU and 260 TB/s across a Vera Rubin NVL72 rack. Here is what it actually does, what it makes possible, and why it forces a rebuild of memory, power, cooling and cabling around it.
What Is NDR InfiniBand Actually Used For? ConnectX-7, Quantum-2, and Real AI Cluster Fabrics Explained
NDR InfiniBand, carried by MMA4Z00-NS400 and MMA4Z00-NS optics, is the scale-out fabric behind NVIDIA's own DGX H100/H200 SuperPOD reference architecture. Here's the real problem it solves, who buys it, and how to size an optics order for it.
Lossless Ethernet for RoCE: Why Cable Quality Affects Tail Latency (and What to Buy)
Short answer: in a lossless RoCEv2 fabric, tail latency is driven less by average bandwidth and more by how often links have to retransmit or trigger flow control. Cable quality is one of the few variables you fully control — a low-BER passive DAC on...
800G AI Fabric Buying Guide: Copper Where You Can, Optics Where You Must (2026)
Short answer: at 800G, copper DAC still wins on cost and power for short runs inside a rack (typically up to ~2m), but reach and signal integrity limits push you to active optical cables or transceivers for anything longer. Here's how to choose, with real...
What Cable Do I Need? The Complete Guide to GPU Cluster Interconnects in 2026
What Cable Do I Need? The Complete Guide to GPU Cluster Interconnects in 2026 Building a GPU cluster is mostly a hardware procurement problem — until you hit the cabling. Then it becomes a compatibility research project that can stall a build for days. This...
InfiniBand vs. Ethernet for AI Clusters in 2026: Which Fabric Is Right for Your Stack?
If you've been following the AI infrastructure space, you've probably heard the debate more than once: InfiniBand or Ethernet? For most of the last decade, InfiniBand was the default answer for serious AI workloads — high bandwidth, ultra-low latency, and a mature ecosystem built specifically...
Beyond the Hype: Is RoCE v2 Finally Winning the AI Networking War?
If you’ve spent any time in a data center recently, especially one focused on AI infrastructure or high-performance computing (HPC), you know the "Interconnect War" is reaching a boiling point. For years, InfiniBand was the undisputed king of HPC and AI training clusters. But as...