Supermicro GPU A+ Server AS-8125GS-TNHR (Refurbished)

Used hardware is sold as pulled; refurbished hardware is restored first. These units are inspected, worn components are addressed, and the machine then passes the same verification applied to all secondary hardware in this catalog, full diagnostics across all eight GPUs, HBM error history, NVLink and NVSwitch health, and load testing. The specific unit's condition report and coverage are confirmed in your quote.

Fleet cycling caught up with the generation. The first large H100 deployments entered service in early 2023, the operators refresh aggressively, and the Blackwell migration is releasing those fleets into the secondary market in volume, the same arc the A100 completed a generation earlier. The supply is a byproduct of upgrades by professional operators, which is why the service history is the best kind available.

It is a memory decision. The H100 carries 80GB of HBM3 per GPU, 640GB per node, excellent for everything that fits it. The H200 carries 141GB of HBM3e per GPU, 1.1TB per node with more bandwidth, and earns its premium for 70B-class full-precision fine-tuning and long-context serving. Availability of each in refurbished condition varies with the fleet cycle, and your quote confirms what is on hand.

Lower than intuition suggests, and measurable. SXM modules have no moving parts, and silicon run within thermal spec does not degrade on data-center timescales; these machines lived nowhere else. What wears is fans and power supplies, commodity parts that refurbishment addresses up front. The residual unknowns, memory error history, fabric health, thermal behavior, are exactly what the pre-shipment diagnostics read, which turns the risk from trust into evidence.

Eight Hopper SXM GPUs in a full NVSwitch mesh at 900 GB/s each, FP8 throughput beyond 30 petaFLOPS, and 640GB or 1.1TB of pooled memory depending on baseboard, with MIG partitioning to as many as 56 instances. In workload terms: full fine-tuning into the 70B class within one node, frontier-grade serving density, and clustering over the included fabric ports.

A dual AMD EPYC configuration with up to 6TB of DDR5 across 24 slots, which clears the system-memory rule against either baseboard with margin, plus twelve PCIe 5.0 slots, twelve hot-swap NVMe bays and two SATA, six 3000W Titanium supplies, and ten fans on air. The exact installed CPUs of the unit on hand are part of the condition report in your quote.

One generation, where AI notices it. FP8 roughly doubles effective training throughput against Ampere, HBM3 and HBM3e outrun HBM2e, and the fabric carries 900 against 600 GB/s. The used A100 node is the lowest entry to genuine NVSwitch training, right for workloads comfortable in 640GB at Ampere speed; the moment FP8 or bandwidth converts to money, this node is the better buy.

When warranty posture and current supply outweigh economics. The Lenovo and ASUS HGX H200 systems and the Intel-host SYS-821GE carry new-unit coverage and factory-fresh baseboards, and procurement requiring them should buy them. Capability-per-budget buyers with verification evidence in hand land here, and AMD-standardized fleets get host consistency as a bonus.

Yes, and the configuration anticipates it: eight NICs provide one fabric port per GPU, the topology GPUDirect RDMA needs to move gradients between nodes without touching the CPU. Twelve PCIe 5.0 slots leave room for fabric upgrades, and MillionMiner advises on switch architecture as deployments grow past one machine.

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