NVIDIA DGX H100 640GB AI System (8x H100 SXM5)

A complete, turnkey system. Unlike the bare HGX baseboards in this catalog, the DGX H100 is the entire finished machine: eight H100 SXM5 GPUs, dual 56-core Xeon processors, 2TB of system memory, NVMe storage, ConnectX-7 networking, and NVIDIA's software stack, integrated, tested, and ready to power on and deploy. There is nothing to integrate yourself.

It is the key choice for acquiring an 8x H100 system. HGX is NVIDIA's GPU baseboard that OEMs (ASUS, Supermicro, Lenovo) build their own servers around; DGX is NVIDIA's own complete system on that same GPU technology. They use the same H100 GPUs, so raw GPU performance is comparable. The difference is integration and support: the DGX is NVIDIA's validated reference platform with its own software and support program; an OEM HGX server is that vendor's integration at a different price and support model. Choose the DGX for NVIDIA's reference design and support; an OEM HGX server for a different price or integration point.

32 petaFLOPS of FP8 across its eight H100 SXM5 GPUs, with 640GB of total GPU memory fully interconnected by NVLink and NVSwitch. That is frontier-class AI compute, the configuration used to train and serve large models. The dual 56-core Xeons and 2TB of system memory feed the GPUs, and ConnectX-7 networking allows scaling across multiple DGX systems for larger clusters.

Everything needed to run: dual 56-core 4th Gen Intel Xeon processors, 2TB of system memory, NVMe storage (2x 1.9TB M.2 for the OS and 8x 3.84TB U.2 for data), ConnectX-7 networking with OSFP ports, NVIDIA's AI Enterprise software stack pre-loaded, support for the major enterprise Linux distributions, and a 3-year warranty. It is a complete, self-contained AI system.

It ships with NVIDIA's AI Enterprise software stack and supports the major enterprise Linux distributions (Ubuntu, Red Hat, and others). The DGX platform is designed so the software, drivers, frameworks, and management are validated together with the hardware as a single supported stack, part of what distinguishes a DGX from a self-assembled system. MillionMiner confirms the exact software entitlement in the quote.

The DGX H100 draws approximately 10.2kW under load and requires data center power delivery and cooling appropriate to a system of that density. 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 do not have suitable facilities.

Yes, unreservedly. Eight H100 SXM5 GPUs delivering 32 petaFLOPS of FP8 is top-tier AI training and inference infrastructure, and the DGX is the system NVIDIA itself positions as its AI infrastructure reference standard. Unlike the workstation and display cards elsewhere in this catalog, there is no overstatement to qualify, the DGX is exactly what it claims to be.

Through its ConnectX-7 networking and OSFP ports, multiple DGX H100 systems interconnect into larger clusters (NVIDIA's DGX BasePOD and SuperPOD reference architectures are built this way) for training that exceeds a single system. For multi-system deployments, MillionMiner advises on the networking and cluster configuration in the quote.

Confirmed for the specific unit in the quote. As flagship Hopper-class hardware, condition and provenance matter, and MillionMiner verifies each system and states its condition clearly, with the 3-year warranty terms confirmed. The DGX H100 is also export-regulated, so MillionMiner confirms destination eligibility and the compliance position before sale.

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, export eligibility for Hopper-class systems, and whether the DGX or an OEM HGX H100 server best fits your needs and budget, then delivery. Every system is tested before shipment and shipped worldwide DDP with duties handled, with hosting in MillionMiner's own data centers available.