GIGABYTE G593-SD2-AAX1 HGX H100 8-GPU Server (5U, Dual Xeon)

The NVIDIA HGX H100: eight SXM5 GPUs at 80GB of HBM3 each, 640GB per node with aggregate bandwidth above 26 TB/s, a full NVSwitch mesh at 900 GB/s per GPU, FP8 throughput beyond 30 petaFLOPS, and MIG to as many as 56 instances. In workload terms: full fine-tuning into the 70B class within one node, large-batch coupled training, and production serving density.

Same Gigabyte G593 family, different host and configuration model. This SD2 runs dual Intel Xeon Scalable with AMX and up to 8TB of DDR5, as a fixed, fully-specified build. The ZD2 runs dual AMD EPYC with 12-channel memory bandwidth, as a barebone you configure. Intel estates, AMX pipelines, and large host-memory needs land here; AMD fleets and configure-your-own buyers take the ZD2, and MillionMiner quotes both.

It is the largest host memory in the catalog's H100 systems, and it pays where the host does heavy lifting: memory-intensive preprocessing, large in-memory datasets and feature stores, and dense virtualization or multi-tenancy alongside the GPUs. The rule this catalog applies everywhere is system memory at or above the 640GB GPU pool; 8TB clears it many times over for workloads that need it.

SXM with NVSwitch. The eight GPUs are on the HGX baseboard, joined at 900 GB/s in a full mesh, which is the genuine coupled-training topology where all GPUs synchronize at full bandwidth. PCIe H100 servers, like the Lenovo SR675 V3 here, run the cards more independently. For distributed training, the SXM mesh is the point; for independent-GPU work, a PCIe platform can be the better fit.

Two things beyond the 8TB capacity. AMX matrix extensions on Xeon Scalable accelerate host-side tensor work, tokenization, embeddings, preprocessing, the stages that feed eight GPUs. And estate standardization: organizations on Intel keep their tooling, licensing, and validated images. The named processor generation for your unit is confirmed on quote.

Eight 2.5-inch Gen5 hot-swap bays across NVMe, SATA, or SAS for flexible local storage, and dual 10GbE base networking through the Intel X710 controller. The compute fabric for clustering is specified per deployment in the PCIe Gen5 expansion. The exact fit-out is confirmed with your quote.

All three are SXM H100 nodes. The Dell is the Intel-host enterprise incumbent with iDRAC governance; the Quanta is the hyperscale ODM build; this Gigabyte leads on host memory capacity at 8TB and Gigabyte value. Estate-led buyers tend Dell, ODM-value buyers the Quanta, memory-heavy and value buyers this, and MillionMiner routes the decision.

Redundant 3000W Titanium supplies in a 5U air-cooled chassis, provisioning in the ten-kilowatt class for a full eight-GPU build plus the host. The exact PSU count and draw of your configuration are confirmed on quote, and hosting in MillionMiner's own facilities is available for teams that prefer the rack power handled.

Through PCIe Gen5 fabric adapters specified per deployment, carrying GPUDirect RDMA so gradients move between nodes without touching CPUs, on InfiniBand or high-speed Ethernet. MillionMiner advises on switch and fabric architecture as deployments grow past one node.

Submit your workload and deployment details through the quote form. A MillionMiner specialist confirms the configuration, the 3-year warranty terms, destination eligibility under the US export controls that apply to Hopper-class accelerators, and delivery. Every system is tested before shipment and shipped worldwide DDP with duties handled, with hosting in MillionMiner's own data centers available.