NVIDIA HGX A100 8-GPU Server Used

A used NVIDIA HGX A100 8-GPU server is a complete 4U system with eight A100 SXM4 GPUs on an HGX baseboard connected through NVSwitch, giving 600 GB/s of NVLink bandwidth per GPU and 640 GB of pooled HBM2e in the 80 GB version. The host platforms that reach the secondary market are mainly the Supermicro AS-4124GO-NART and SYS-420GP-TNAR, the Inspur NF5488A5 and the Gigabyte G492-ZD2, with two AMD EPYC 7002/7003 or third-generation Intel Xeon Scalable processors, 32 DIMM slots, six or more hot-swap NVMe bays and 2200 W to 3000 W redundant power supplies. Each A100 SXM4 80 GB module has a 400 W TDP, and the eight-GPU assembly delivers about 5 PFLOPS of FP16 Tensor performance with sparsity.

These systems are bought by AI labs and universities that need a training node without waiting for new Hopper allocation, by hosting providers building GPU clouds, by rendering and simulation studios and by resellers in the CIS, Africa and South Asia. Before shipment the partner runs a multi-hour GPU stress test on all eight modules, checks the NVSwitch topology, HBM error counters and thermals, completes a burn-in and full memory pass on the host, updates BIOS, BMC and GPU firmware, and replaces fans, power supplies and thermal interface material as required. The partner limited warranty, typically 90 days to 12 months, is confirmed in the quotation.

Supply of used HGX A100 systems goes through InTechnologies' licensed partners, who issue the contract, invoice and partner warranty. InTechnologies coordinates sourcing from decommissioning programmes and verified brokers, consolidation in Dubai Airport Freezone and DDP or DDU delivery; because of the system's weight and power draw, freight and site preparation are planned with the customer. A100 systems remain subject to US export regulations regardless of condition and are supplied only after end-user verification.