NVIDIA B200 SXM6 192GB

The NVIDIA B200 SXM6 192GB is a flagship data center GPU built on NVIDIA's Blackwell architecture, representing the most advanced compute accelerator in NVIDIA's current product family. Featuring 192GB of HBM3e memory across a unified memory pool, the B200 delivers up to 20 petaFLOPS of FP4 AI inference performance and 9 petaFLOPS of FP8 training throughput, enabling organizations to train and deploy the largest frontier AI models at scale. The SXM6 form factor connects via NVIDIA's NVLink 5.0 interface, supporting up to 1.8TB/s of GPU-to-GPU bandwidth in multi-GPU configurations, making it purpose-built for tightly coupled HPC and generative AI clusters.

Supermicro's implementation of the B200 SXM6 192GB (part number 900-21901-0000-000) integrates directly into Supermicro's next-generation MGX and SYS-based server platforms optimized for Blackwell-class compute density. The GPU's second-generation Transformer Engine, combined with NVLink Switch technology, allows seamless scaling from individual nodes to large-scale GPU fabrics, supporting collective operations at near-native bandwidth. The 192GB HBM3e memory subsystem provides approximately 8TB/s of memory bandwidth, eliminating memory bottlenecks for models with hundreds of billions of parameters.

Designed for enterprise data centers, cloud service providers, and national research laboratories, the B200 SXM6 supports a comprehensive software stack including CUDA 12.x, cuDNN, TensorRT-LLM, and NVIDIA NIM microservices. Its confidential computing capabilities and hardware-level security features make it suitable for regulated industries requiring data sovereignty and workload isolation. The accelerator is available through Omnixon Global, serving enterprise buyers across UAE, GCC, EMEA, and APAC regions.

Ideal for

• Large-scale generative AI model training for LLMs with hundreds of billions to trillions of parameters, leveraging FP8 mixed-precision and the Transformer Engine for maximum throughput
• High-performance AI inference serving for production deployment of frontier models, utilizing FP4 quantization to maximize tokens-per-second per GPU
• Scientific computing and HPC simulation workloads in climate modeling, drug discovery, and computational fluid dynamics requiring extreme double-precision floating-point performance
• Retrieval-Augmented Generation (RAG) pipelines and vector database acceleration for enterprise AI applications requiring large in-GPU memory capacity
• Sovereign AI and national AI infrastructure deployments where data residency and confidential computing hardware controls are mandatory requirements
• Multi-node deep learning research clusters utilizing NVLink 5.0 and NVSwitch for all-to-all collective operations across GPU fabrics at supercomputer scale

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.