Hardware library
Every part of an AI cluster, explained
What each component does, where it fits, what it works with, and the mistakes that stall a build. Need it sourced or financed? Every page connects you to us.
Compute
GPU servers
HGX systems from A100 to B300 and what separates each generation.
HGX H100The workhorse Hopper node: eight H100 SXM5 GPUs on NVLink 4, 640 GB of HBM3, and the most mature software stack of any HGX system.HGX H200141 GB HBM3e per GPU. The memory upgrade over H100 that wins on long-context inference and large-model serving.HGX B200First-generation Blackwell: eight B200 GPUs with 180 GB of HBM3e each, NVLink 5 at 1.8 TB/s per GPU, and native FP4.HGX B300Blackwell Ultra: eight B300 GPUs with 288 GB of HBM3e each, NVLink 5, and 800G ConnectX-8 networking.HGX A100 80GBThe proven Ampere workhorse: eight A100 80GB GPUs on NVLink 3. End of life at the OEM level, and the lowest-power HGX node.RTX Pro 6000 BlackwellBlackwell in a PCIe card: 96 GB of GDDR7 with ECC, for inference, fine-tuning and visual computing in standard servers and workstations.
Fabric
Switches
InfiniBand Quantum vs. Spectrum-X Ethernet, and how rail-optimized fabrics are built.
Quantum-2 QM9700The NDR InfiniBand leaf and spine switch behind most Hopper training clusters. 64 ports of 400G in a 1U box.Spectrum-4 SN5600The 800G Ethernet switch at the core of Spectrum-X AI fabrics, for operators standardizing on Ethernet instead of InfiniBand.Quantum QM8700The HDR InfiniBand switch behind A100-era clusters. 40 ports of 200G in a 1U box, or 80 ports of 100G when split.Quantum-X800 Q3400The XDR InfiniBand switch for 800G-per-GPU Blackwell Ultra clusters. 144 ports of 800G in one liquid-cooled chassis.
Optics
Transceivers and cables
400G/800G OSFP, DR4 vs. SR8, DACs and AOCs, and what plugs into what.
800G OSFP 2×DR4The optic that connects GPU servers to the switch fabric in most 400G-per-GPU clusters. One module carries two independent 400G links over single-mode fiber.400G OSFP DR4The NIC-side optic in NDR clusters. It plugs into a ConnectX-7 and runs one 400G link over single-mode fiber to half of an 800G twin-port module in the leaf switch.800G OSFP 2×SR4The short-reach, multimode option for switch ports. Cheaper per link than DR4 when servers and leaf switches sit in the same or neighboring racks.
Network
NICs and DPUs
ConnectX-7, BlueField-3 and where each belongs in a node.
ConnectX-7The standard GPU-node NIC for Hopper-generation clusters. HGX H100 and H200 systems typically carry eight, one per GPU, for the compute fabric.ConnectX-6The HDR-generation NIC behind A100 clusters. DGX A100 carries eight single-port cards, one per GPU, for the compute fabric.ConnectX-8 SuperNICThe 800G NIC for Blackwell Ultra. HGX B300 and GB300 systems use it for the compute fabric, one per GPU.BlueField-3A NIC with its own Arm cores. As a DPU it runs storage, security and tenant isolation off the host CPU; as a SuperNIC it is the host side of a Spectrum-X Ethernet fabric.
Storage
Storage
Parallel file systems and the NVMe tiers that keep GPUs fed.
Write-ups for this category are in progress.
Facility
Power and cooling
Rack density, PDUs, liquid cooling and what a Blackwell rack demands.
Write-ups for this category are in progress.