NVIDIA DGX H800 640GB AI System (8x H800 SXM5)

A complete, turnkey system. Unlike the bare HGX baseboards in this catalog, the DGX H800 is the entire finished machine: eight H800 SXM5 GPUs, dual 112-core Xeon Platinum processors, 2TB of system memory, NVMe storage, 4x NVSwitch, and NVIDIA's software stack, integrated, tested, and ready to power on. There is nothing to integrate yourself.

The H800 is NVIDIA's export-compliant version of the H100, the same Hopper silicon with interconnect bandwidth reduced to meet US export rules, built for markets like China. The DGX H800 is, in effect, the DGX H100 built with H800 GPUs: the same system design, CPUs, memory, and storage, with export-compliant GPUs instead of full H100s. It is the China-market counterpart to the DGX H100.

It matters mainly for large-scale distributed training across many systems and GPUs, where the speed of cross-GPU communication determines how well a job scales; there the DGX H800 is slower than a full DGX H100. For single-system training, fine-tuning, and inference, where interconnect is not the binding constraint, it performs in the DGX H100's class. The honest question is whether your training is communication-bound at large scale, and MillionMiner will work through it.

32 petaFLOPS of FP8 across its eight H800 SXM5 GPUs, with 640GB of total GPU memory interconnected by 4x NVSwitch. That is frontier-class Hopper-generation AI compute for training and serving large models. The dual 112-core Xeon Platinum 8480C processors and 2TB of system memory feed the GPUs.

Everything needed to run: dual 112-core Intel Xeon Platinum 8480C processors, 2TB of system memory, NVMe storage (2x 1.92TB M.2 for the OS and 8x 3.84TB U.2 for data), 4x NVSwitch, and NVIDIA's software stack. It is a complete, self-contained AI system weighing about 130kg. MillionMiner confirms the exact software and networking configuration in the quote.

Mainly destination and availability. The H800 is the export-compliant Hopper part for markets where the full H100 exceeds US export thresholds, so in those regions the DGX H800 is the compliant path to a turnkey Hopper DGX. Where a full-H100 system is permitted and the workload is large-scale interconnect-bound training, the DGX H100 may perform better. MillionMiner confirms which is appropriate and available for your destination in the quote.

The DGX H800 draws up to 10.2kW and requires data center power delivery and cooling appropriate to a system of that density and weight (about 130kg). This is data center infrastructure, not an office machine. MillionMiner confirms the power and cooling requirements for your site in the quote, and offers hosting in its own data centers if you lack suitable facilities.

Yes. Eight Hopper-class H800 GPUs delivering 32 petaFLOPS of FP8 with 640GB of memory is top-tier AI training and inference infrastructure. Unlike the workstation and display cards elsewhere in this catalog, there is no overstatement to qualify, only the two scopes to note: it is a complete system, and it is the export-compliant H800 configuration.

The H800 is itself an export-compliance product, and Hopper-class systems are subject to US export controls that depend on destination. MillionMiner confirms destination eligibility and the compliance position before sale. As flagship hardware, condition and provenance matter, and each system is verified before shipment with its condition stated clearly in the quote.

Submit your workload, large-model training, fine-tuning, or inference, and scale through the quote form. A MillionMiner specialist confirms the configuration, the software entitlement, power and cooling requirements, destination eligibility under the export rules for Hopper-class systems, and whether the DGX H800 or a full-H100 path best fits your destination and needs, then delivery. Every system is tested before shipment and shipped worldwide DDP with duties handled, with hosting in MillionMiner's own data centers available.