ASUS XA NB3I-E12 NVIDIA HGX B300 NVL8 GPU Server

Every HGX B300 NVL8 system carries the same NVIDIA heart, so the honest way to present one is to state the shared complex once, then spend the page on what the manufacturer actually decided. This is ASUS's build, and its decisions are an Intel host and an operations-first chassis. The shared complex, stated once. Eight NVIDIA Blackwell Ultra B300 SXM GPUs, 288GB of HBM3e each, pool 2,304GB across the node, the largest single-node memory in this catalog. Fifth-generation NVLink joins every GPU at 1.8 TB/s through NVSwitch in full mesh. NVIDIA rates the generation's dense FP4 at 1.5 times the HGX B200, an uplift aimed at the serving economics of models that reason before answering, where chain-of-thought traces and long contexts multiply KV cache until memory capacity, not compute, sets batch size and margin. A 2.3TB node holds frontier weights and those caches together. The baseboard also carries its own networking: eight ConnectX-8 InfiniBand ports at 800G, one per GPU, double the prior generation's rate, which is why systems of this class no longer need a dozen PCIe slots full of adapters. The fabric topology GPUDirect RDMA wants is built in at the factory. The Intel host case. Dual Intel Xeon 6700P series processors anchor this build, the Xeon 6 performance-core platform reaching 86 cores per socket at 350W. Two things make the Intel choice substantive rather than tribal. First, AMX, the matrix extensions present across Xeon 6 P-cores, accelerates exactly the host-side work an 8-GPU node generates: tokenization, preprocessing, embedding work, and the CPU-resident inference stages of mixed pipelines. Second, fleet reality: organizations that standardized on Intel carry Xeon-tuned tooling, per-core licensing positions, validated images, and operational playbooks, and for them the Intel B300 node is the one that drops into the estate without a parallel platform to certify. Eight DDR5 memory channels per processor, with support for high-bandwidth MRDIMM modules on this platform generation, keep the host memory system in proportion to the GPU complex it feeds; the as-built memory and storage configuration is specified on your quote. The ASUS engineering case. The chassis is modular with minimal internal cabling, and at this power class that phrase carries real weight. Cables obstruct airflow in a machine whose job is moving kilowatts of heat; every connector is a failure point in a node expected to run at sustained load for years; and service time is production time when eight Blackwell Ultra GPUs sit idle waiting on a fan swap. ASUS's modular assembly reduces all three, which is the kind of advantage that never appears in a benchmark and appears constantly in an operations log. Power is a 5+5 redundant arrangement of 80 PLUS Titanium supplies, the efficiency tier where conversion losses at this scale stop being a rounding error. The decision between Blackwell Ultra siblings. This catalog carries two HGX B300 NVL8 systems, and the GPU performance between them is identical by construction. The Supermicro AS-8126GS-NB3RT runs the AMD EPYC 9005 host with up to 192 cores per socket and up to 6TB of DDR5, the pick for AMD-standardized fleets and core-density-hungry preprocessing. This ASUS build runs the Intel Xeon 6 host with AMX and the Intel estate advantages, in a chassis engineered for serviceability. MillionMiner quotes both and recommends on your existing infrastructure, not on preference. The decision against the rest of the ladder. Teams working inside 1.4TB per node are usually better served by the Gigabyte HGX B200, and teams inside the Hopper envelope by the Lenovo or ASUS H200 systems; the B300 earns its premium where reasoning-scale KV cache, frontier model size, or 800G fabric standardization decide the economics. Above it sit the GB200 and GB300 rack-scale systems, with their 72-GPU NVLink domains and liquid-cooled facility commitments, for the minority whose parallelism genuinely outgrows eight GPUs in one domain. Blackwell-class accelerators are subject to US export controls, and MillionMiner confirms destination eligibility as part of every quote. Each system is configured to order, tested, and shipped worldwide DDP with duties and customs handled. Deployment planning and hosting in MillionMiner's own data centers are available for teams that prefer not to provision rack power at this scale on-site.