Supermicro GPU SuperServer ARS-111GL-NHR, NVIDIA GH200 (Used)

It fuses the CPU and GPU into one chip. The GH200 Grace Hopper Superchip puts a 72-core Grace Arm CPU and a Hopper GPU on a single module, joined by NVLink-C2C at 900 GB/s, where a conventional server connects a separate CPU and GPU over much slower PCIe. They also share coherent memory rather than holding separate pools. It is a server architecture, not a GPU you add to one.

The Grace CPU's up-to-480GB of LPDDR5X and the Hopper GPU's 96GB of HBM3 are addressable as one coherent space that both processors see, rather than two separate pools you copy data between. For workloads that constantly move data between CPU and GPU, or that need more memory than a discrete GPU's HBM alone, that unified design removes copies and the PCIe bottleneck, which is the GH200's core advantage.

Memory-bound and CPU-GPU-coupled work: large-graph analytics and graph neural networks, recommender systems with huge embedding tables, and memory-heavy inference that exceeds a discrete GPU's HBM. For these the coherent unified memory is a structural advantage. For dense multi-GPU matrix training, the HGX SXM systems in this catalog remain the better tool.

Different shape of problem. The HGX H100 and H200 systems are eight-GPU NVSwitch meshes for distributed training, with 640GB to 1.1TB of pooled HBM. This is a single Grace Hopper superchip node, one fused CPU-GPU with coherent memory. Choose the HGX systems for multi-GPU training; choose the GH200 for memory-bound, graph, and coupled workloads where unified memory wins. They complement rather than compete.

It is a secondary-market unit, inspected and verified by MillionMiner before shipment, the same framework applied to the catalog's used A100 and refurbished Hopper systems. Grace Hopper modules have no moving parts and silicon run within thermal spec does not degrade on these timescales; the wear items are fans and supplies. Condition and coverage for the specific unit are confirmed in your quote.

Economics and access. Buying used brings a different and otherwise expensive architecture within reach, for teams that want to evaluate Grace Hopper for memory-bound workloads or deploy it without new-unit pricing. The verification framework converts the condition risk into measurable evidence, and MillionMiner confirms what the unit includes in the quote.

The Grace CPU is Arm Neoverse, not x86, so the software stack targets Arm, which the NVIDIA AI ecosystem, CUDA, the frameworks, and the container catalog, supports natively on Grace Hopper. Teams should confirm their specific application stack runs on Arm, and MillionMiner can advise on the software considerations during the quote.

A 1U rackmount with dual PCIe 5.0 slots, hot-swap E1.S NVMe storage, dual 2000W redundant Titanium power supplies, and a dedicated 1GbE BMC management port. The exact storage fit-out of the used unit is confirmed in your quote.

Grace Hopper deployments scale through the node's networking and the PCIe slots for fabric adapters, and NVIDIA's Grace Hopper platforms support high-speed interconnect for multi-node scale-out. For a used single-node unit, MillionMiner advises on the networking and fabric options for your deployment during the quote.

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