Exeton Quasar 640X 8x A100 AI Server (640GB NVLink)

Exeton is a specialist AI server integrator rather than a tier-one marque like Supermicro or NVIDIA. The components inside are standard: NVIDIA A100 GPUs, Intel Xeon Platinum processors, enterprise SSDs. What the integrator contributes is the configuration, and the spec sheet shows experienced judgment throughout. Every unit MillionMiner supplies is sourced and tested before shipment, carries three years of replacement parts coverage, and is supported by MillionMiner directly, so the brand on the chassis does not change who stands behind your hardware.

Fixed, and that is the point of this product. The Quasar 640X is the complete, decided build for buyers who want one quote and one known-good machine. If you need a different CPU platform, memory capacity, or storage layout, the Supermicro AS-4124GO-NART+ in this catalog is the configurable 8x A100 path, and MillionMiner quotes both.

Both deliver eight A100 80GB GPUs with NVLink and 640GB of combined GPU memory. The DGX adds NVIDIA's full software stack, Base Command, NGC certified containers, and single-vendor support, at appliance economics. The Quasar delivers the same class of GPU compute in a complete fixed build at integrator economics, with you running standard open tooling on top. Certainty-led enterprise buyers tend toward the DGX; value-led teams comfortable with the standard CUDA ecosystem tend here.

The Supermicro is a configurable platform where you specify CPUs, memory, storage, and networking yourself, on an AMD EPYC host. The Quasar is a fixed Intel Xeon build where those decisions are already made, including 1TB of memory and a 200G InfiniBand NIC in the base spec. Choose the Supermicro to tailor the host; choose the Quasar to skip the configuration exercise entirely.

Because the working rule for an 8-GPU training node is system memory at or above total GPU memory. Data loading, preprocessing, and augmentation for eight A100s need host-side headroom, and an undersized host starves the accelerators you actually paid for. With 640GB of GPU memory in the node, 1TB of ECC system memory clears the rule with margin, which is exactly why this configuration was chosen.

Separation of duties. Two 960GB enterprise SSDs carry the operating system, and a dedicated 7.68TB U.2 NVMe drive carries datasets and checkpoints. Training jobs read datasets and write checkpoints constantly; keeping that I/O off the boot volume prevents contention and keeps both the OS and the training pipeline responsive. It is the layout an experienced operator would have specified anyway.

Yes, and the base build anticipates it. The included 200G InfiniBand NIC provides the compute fabric port for GPU-to-GPU communication across nodes, while the dual 10 Gigabit Ethernet handles management traffic. Scaling from one node to several is an interconnect and switch exercise rather than a hardware retrofit, and MillionMiner advises on fabric design during configuration.

Full fine-tuning of models in the 30B to 70B class with parallelism across the eight GPUs, training of smaller models at very large batch sizes, and simultaneous serving of multiple 70B-class inference workloads. Using MIG, the node partitions into up to 56 hardware-isolated GPU instances, each suited to serving models in the 7B to 13B class, which makes one server a complete multi-tenant inference platform.

This is a data center machine. Four 3000W high-efficiency power supplies define the envelope, and an eight-GPU A100 node draws several kilowatts under sustained training load with high-volume front-to-back airflow required. MillionMiner confirms the exact power and cooling plan for your site during configuration, and hosting in MillionMiner's own facilities is available if you prefer not to provision infrastructure.

Submit your workload and deployment details through the quote form, and a MillionMiner specialist replies with a complete quote covering the system, three-year replacement parts coverage, and delivery. Every unit is tested before shipment and delivered worldwide DDP with duties and customs handled. Rack integration guidance and hosted deployment are both available.