Supermicro AS-8126GS-NB3RT-01-G2 8U Gold Series HGX B300 NVL8 GPU SuperServer

Experience The HGX Platform with Eight NVIDIA B300 GPUs for Large-Scale AI

The Supermicro AS-8126GS-NB3RT-01-G2 is an 8U rackmount AI workstation built around NVIDIA HGX B300 8-GPU architecture. With eight NVIDIA B300 SXM GPUs delivering 288GB GPU memory, the platform is aimed at demanding AI workloads including large-scale inference serving, model development and other GPU-accelerated applications that require substantial compute and memory resources.

Dual AMD EPYC 9575F Processors

At the heart of the system are 2x AMD EPYC™ 9575F processors, each providing 128 cores at 3.3GHz. The dual-socket configuration delivers 256 CPU cores and provides considerable host-side compute for data preparation, orchestration, virtualisation and CPU-intensive AI workloads. The platform supports AMD EPYC 9005 Series processors, with the underlying architecture supporting configurations of up to 384 cores and 768 threads.

3TB of High-Speed DDR5 ECC

The workstation is configured with 3TB of DDR5-6400 ECC RDIMM memory, giving large AI models and datasets ample system memory for demanding workloads. The motherboard provides 24 DIMM slots, while the platform supports up to 6TB of ECC DDR5 RDIMM memory at 6400MT/s or 4800MT/s in a 1DPC configuration, allowing the system to be specified for even larger memory requirements.

NVIDIA NVLink and NVSwitch GPU Fabric

The eight B300 GPUs communicate through NVIDIA NVLink with NVSwitch, creating a high-bandwidth GPU-to-GPU fabric for multi-GPU applications. This is particularly important for AI workloads that distribute models or datasets across multiple accelerators, reducing the communication bottlenecks that can arise when GPUs rely solely on conventional host-level connectivity.

PCIe 5.0 CPU-to-GPU Connectivity

The HGX B300 architecture is paired with PCIe 5.0 x16 CPU-to-GPU interconnects, providing high-speed communication between the AMD EPYC processors and GPU subsystem. This supports efficient movement of data between the host CPUs and accelerators for workloads where CPU-side processing and GPU compute operate closely together.

3TB-Class NVMe Storage Configuration

Storage is built around eight 7.6TB E1.S NVMe SSDs, providing 60.8TB of raw SSD capacity for high-speed datasets, model repositories and application data. The system also includes 2x 1.9TB M.2 NVMe SSDs, adding a further 3.8TB of raw capacity for operating system, application or auxiliary storage requirements. In total, the specified configuration offers 64.6TB of raw NVMe storage capacity.

800GbE ConnectX-8 Networking

For high-throughput networking, the workstation features 8-port 800GbE/XDR800 connectivity through ConnectX-8, with eight OSFP 800GbE ports on the front panel. This gives the system substantial network bandwidth for distributed AI infrastructure, high-speed data movement and connectivity to other compute and storage nodes. This front-accessible arrangement is suited to dense AI deployments where rapid network connectivity is a core part of the system architecture.

Additional 200GbE Connectivity

The networking configuration also includes 2x dual-port 200GbE/NDR200 connectivity, complementing the eight 800GbE interfaces. Alongside the high-speed OSFP ports, the platform provides 2x RJ45 10GbE connections for conventional Ethernet networking and system management requirements.

Six 6600W Titanium Power Supplies

Power delivery is built around six 6600W Titanium-level power supplies, arranged in a 3 + 3 redundant configuration. The redundancy architecture provides resilience across the power supply subsystem, while the Titanium-level units are rated at 96% efficiency, helping reduce power losses in a workstation with a substantial GPU and CPU power requirement.

Heavy-Duty Cooling for AI Workloads

Thermal management is handled by 12 heavy-duty system fans, with fan-speed control and monitoring built into the H14DSG-OM platform. PWM fan connectors, tachometer monitoring and chassis temperature monitoring provide the control and telemetry required to manage airflow across the dense compute hardware.

Comprehensive Server Management

Management support includes SuperCloud Composer (SCC), Supermicro Server Manager (SSM), Super Diagnostics Offline (SDO), Supermicro Thin-Agent Service (TAS), SuperServer Automation Assistant (SAA) and SFT-DCMS-SINGLE. These tools provide administrators with multiple options for system monitoring, diagnostics, automation and infrastructure management.

Flexible I/O for Instant Deployment

Despite its focus on high-speed compute and networking, the system retains practical connectivity for administration and local servicing. The front panel provides 1x USB 3.0 Type-A and 1x USB 2.0 Type-A, while display connectivity includes 1x VGA and 1x mini-DP. The platform also includes AMI 64MB UEFI BIOS for system firmware and configuration.

• 8U rackmount AI workstation with dual AMD EPYC 9575F processors, 256 combined CPU cores and 3TB DDR5-6400 ECC memory (12DPC Each).
• NVIDIA HGX B300 8-GPU platform with 8x B300 SXM GPUs and 288GB HBM3e memory.
• NVIDIA NVLink and NVSwitch GPU-to-GPU connectivity for multi-GPU workloads.
• 64.6TB total raw NVMe storage from 8x 7.6TB E1.S SSDs and 2x 1.9TB M.2 SSDs.
• ConnectX-8 networking with 8x 800GbE OSFP ports, plus 2x 200GbE QSFP112 ports and 2x 10GbE connectivity.
• PCIe 5.0 x16 CPU-to-GPU connectivity for high-speed host-to-accelerator communication.
• 3+3 power redundancy with 6x 6600W Titanium-level PSUs rated at 96% efficiency.
• 12 heavy-duty fans with PWM control, tachometer monitoring and thermal management.
• USB 3.0, USB 2.0, VGA and mini-DP I/O connectivity.
• 10 total 2.5" hot-swap bays with eight supporting NVMe and two SATA.
• Supermicro management and diagnostic support through SuperCloud Composer, Server Manager and Super Diagnostics Offline