Gigabyte G492-ZD0 HGX A100 8-GPU Server, Dual AMD EPYC 7742, 4U (Used)

These are data-center-pulled nodes from fleet cycling, hardware that spent its service life in climate-controlled professional facilities. Before shipment, every unit passes full GPU diagnostics across all eight A100 modules, HBM error-correction history checks, NVLink and NVSwitch link health verification, fan and power supply function tests, and sustained load testing. Units that fail any check do not ship. Condition details for the specific unit on offer are confirmed in your quote.

Because the largest AI operators cycle fleets. As hyperscalers and labs move capacity to Hopper and Blackwell generations, their A100 infrastructure enters the secondary market. The supply is a byproduct of upgrade cycles, not of failure, and it is the reason genuine NVSwitch training capability is now available at a fraction of its original platform cost.

Fans and power supplies, both hot-swappable commodity parts on this platform. The A100 SXM4 modules themselves have no moving parts, and silicon operated within thermal specification does not degrade on data-center timescales. The legitimate concern with used GPUs is hidden damage, and that is measurable rather than a matter of trust: HBM error history, NVLink health, and thermal behavior are readable from the hardware, which is exactly what the pre-shipment testing reads.

Eight NVIDIA A100 80GB SXM4 GPUs on an NVSwitch full mesh, 600 GB/s per GPU, pooling 640GB of HBM2e. That supports full fine-tuning of 30B to 70B class models with parallelism, large-batch training of smaller models, and through MIG, up to 56 hardware-isolated GPU instances for multi-tenant inference from a single 4U machine.

It is slightly under the rule this catalog teaches, system memory at or above total GPU memory, against the 640GB GPU pool. For many fine-tuning and inference workloads it is sufficient; for heavy preprocessing pipelines it is the first thing to upgrade. Sixteen of the platform's thirty-two DIMM slots are open, so bringing the host to 1TB is a quoted line item rather than a redesign, and MillionMiner sizes it with you against your actual workload.

Eight 200 Gb/s NICs, one per GPU, which is the topology GPUDirect RDMA clustering requires: GPUs in different nodes exchange data directly across the fabric without CPU involvement. The node arrives cluster-ready, and scaling past one machine is a switch and cabling exercise that MillionMiner advises on during the quote.

Same fundamental GPU complex across all four: eight A100 80GB SXM4 GPUs on NVSwitch. The DGX A100 adds NVIDIA's software stack and single-vendor support at appliance economics. The Supermicro AS-4124GO-NART+ is the configurable new build. The Exeton Quasar 640X is the fixed new build. This used Gigabyte node is the value floor: the same training capability at the lowest cost of entry, with condition verified by testing rather than a factory seal. Budget-led buyers start here; MillionMiner quotes all four honestly.

When memory or throughput is already the constraint. Workloads pressing past 640GB per node, training programs where FP8 Transformer Engine throughput compresses timelines with business value, or long-context serving beyond what 80GB GPUs hold all point to the H800, H200, and B200 systems in this catalog. If your models fit comfortably in the A100 envelope, the used node is the rational buy and the savings fund a lot of compute time.

Fully. Current CUDA releases and every major framework, PyTorch, TensorFlow, JAX, TensorRT, Triton, vLLM, maintain mature Ampere support, and A100 fleets still carry a substantial share of global cloud AI capacity, which keeps the ecosystem actively maintained. MIG, NVLink, and GPUDirect tooling are all long-proven on this platform.

Submit your workload and deployment details through the quote form. A MillionMiner specialist confirms the specific unit's condition report, the coverage included with it, any configuration additions such as the memory upgrade, and the delivery plan. Every system is tested before shipment and delivered worldwide DDP with duties and customs handled. Hosting in MillionMiner's own data centers is available as an alternative to on-site deployment.