Gigabyte G593-ZD2-AAX1 Barebone AMD H200 8-GPU Server (141GB SXM)

Same platform, bigger GPU. Both are the same Gigabyte G593-ZD2 AMD-host barebone; this one carries the HGX H200 baseboard at 141GB per GPU and 1.1TB per node, where the H100 build holds 640GB. Identical host engineering and the same configure-your-own model; the H200 is the memory step up. Choose this when your workload strains 640GB, and the H100 twin when it fits, for the better economics.

The NVIDIA HGX H200: eight SXM GPUs at 141GB of HBM3e each, 1,128GB per node with aggregate bandwidth approaching 38 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: 70B-class full-precision fine-tuning within one node, long-context serving with real KV cache headroom, and memory-bound training and inference.

The GPU complex and chassis ship fixed, the HGX H200 baseboard, the chassis, power and cooling, and the host components are yours to specify: the EPYC CPUs, the memory population, and the storage. For teams with particular core-count, capacity, or storage needs, the node arrives matched to the workload rather than to a fixed bundle, and MillionMiner configures the open parts on quote.

Core density and memory bandwidth, the two things a GPU host must bring. Eight GPUs need many cores for parallel data loading and preprocessing, and a memory subsystem that does not starve them. EPYC 9004 on Zen 4 delivers high core counts and 12-channel DDR5 per socket, the widest memory bandwidth in its class, which is why it is a strong host for a feeding-intensive 8-GPU node.

They differ on host and configuration model. The Lenovo SR680a V3 brings tier-one estate fit; the ASRock 6U8X leads on 6U density and economics; the ASUS N8-E11 is built cluster-first around one-NIC-per-GPU fabric; the Quanta S7PH H200 leads on ODM in-rack serviceability. This Gigabyte is the AMD-host barebone you configure yourself, with EPYC 12-channel bandwidth. MillionMiner routes the decision on your estate, density, fabric, and configurability priorities.

When your workload fits comfortably in 640GB. The H100 build of this exact platform costs less and delivers the same host engineering and barebone control; the H200's 1.1TB earns its premium only for memory-bound work, 70B-class full-precision fine-tuning, and long-context serving. The quote models the crossover on your actual workload rather than guessing.

The platform runs 12-channel DDR5 per socket across 24 DIMM slots, and because it is a barebone, the memory population is yours to specify. The rule this catalog applies everywhere holds, system memory at or above the 1.1TB GPU pool, and MillionMiner sizes it to your workload at quote rather than to a fixed default.

Through the PCIe Gen5 expansion carrying fabric adapters for GPUDirect RDMA, the topology that moves gradients between nodes without touching CPUs, on InfiniBand or high-speed Ethernet specified per deployment. As a barebone, the fabric fit-out is configured to your cluster plan, and MillionMiner advises on switch architecture as deployments grow past one node.

Eight 2.5-inch Gen5 hot-swap bays across NVMe, SATA, or SAS-4, redundant 3000W Titanium power, and PCIe Gen5 expansion for fabric and additional storage. As a barebone the drive population is configured to your workload, and the exact PSU draw of a full eight-GPU H200 build is confirmed on quote.

Submit your workload and deployment details through the quote form. A MillionMiner specialist configures the EPYC CPUs, memory, and storage, confirms the 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.