GIGABYTE G893-SD1-AAX5 HGX B200 8-GPU Server

The NVIDIA HGX B200: eight B200 SXM GPUs at 180GB of HBM3e each, 1,440GB pooled per node, 64 TB/s of aggregate bandwidth, and fifth-generation NVLink at 1.8 TB/s per GPU through NVSwitch, the 1.8 TB/s figure on the spec sheet is that per-GPU fabric rate. FP4 compute through the second-generation Transformer Engine reaches 144 petaFLOPS per node.

In one place, the host. This SD1 build runs dual Intel Xeon Scalable with AMX and 32 DDR5 slots in a capacity-first layout; the ZD1 runs dual AMD EPYC with up to 192 cores per socket and 24 slots in a bandwidth-first layout. Chassis, GPU complex, power, expansion, and cooling are shared. Intel estates and AMX-aware pipelines land here; AMD fleets and core-density-hungry preprocessing land on the ZD1, and MillionMiner quotes both.

Three things. AMX matrix extensions across every core accelerate the host-side AI work an 8-GPU node generates, tokenization, embeddings, preprocessing. The 32-slot memory layout gives capacity-planning room that 24-slot designs cannot. And fleet standardization: Intel estates keep their tooling, licensing, and validated images, adding Blackwell without certifying a second platform.

Per GPU: 180GB of HBM3e against the H200's 141GB, roughly 8 TB/s of bandwidth against 4.8, and NVLink doubled to 1.8 TB/s. Per node: 1.4TB against 1.1TB, plus FP4 precision that NVIDIA rates at up to 15x real-time trillion-parameter inference against the H100 generation, a vendor figure resting on physical changes. Teams inside the Hopper envelope still buy Hopper economics; teams pressing its ceilings step here.

n three situations: inference fleets at frontier scale where FP4 roughly doubles tokens per watt, training programs where the doubled fabric and bandwidth compress timelines with business value, and workloads already pressing the H200's per-GPU memory. Below those thresholds, the Lenovo and ASUS H200 systems remain the rational buy, and the quote models both paths.

Base networking ships onboard, dual 10GbE through the Intel X710 controller. The compute fabric is specified per deployment: eight single-slot PCIe 5.0 positions take one 400 Gb/s adapter per GPU on Quantum-2 InfiniBand or Spectrum-X Ethernet, the GPUDirect RDMA topology that moves gradients across nodes without touching CPUs. Four dual-slot positions remain for storage and management networking.

Thirty-two DDR5 RDIMM slots, eight channels per socket with two modules per channel. The trade: one module per channel holds peak speed, two per channel trades some speed for slot count and capacity ceiling. This build takes the capacity side, which suits very large host memory and staged expansion. The rule is unchanged, system memory at or above the 1.4TB GPU pool, and the validated maximum and your fit-out are confirmed on quote.

Eight 2.5-inch Gen5 hot-swap bays, NVMe or SATA, for the local dataset tier, plus two M.2 slots that keep the boot environment off the data path, the separation this catalog recommends on every training node. Exact drive fit-outs are specified with your quote, and MillionMiner ships units pre-populated when the storage plan is known.

Twelve 3000W 80 PLUS Titanium supplies in 6+6 redundancy define the envelope, with the GPU complex alone capable of drawing eight kilowatts under sustained load before the host counts. The 8U design cools it on air, keeping deployment inside ordinary data centers. MillionMiner confirms the exact draw of your configuration during the quote, with hosting in its own facilities as the alternative.

Submit your workload, host preference, and deployment details through the quote form. A MillionMiner specialist confirms configuration, the SD1-versus-ZD1 routing if still open, destination eligibility under the US export controls that apply to Blackwell-class accelerators, and delivery. Every system is tested before shipment and shipped worldwide DDP with duties handled.