Lenovo ThinkSystem SR680a V3 with NVIDIA GB200 NVL72 Rack-Scale System

The Lenovo ThinkSystem SR680a V3 with NVIDIA GB200 NVL72 is a rack-scale AI supercomputing system engineered for the most demanding large-scale generative AI training, inference, and HPC workloads. Built around NVIDIA's GB200 NVL72 architecture, the system integrates 36 Grace CPU modules and 72 Blackwell B200 GPU dies within a single NVLink-interconnected rack domain, delivering unified high-bandwidth memory and compute fabric at unprecedented scale. The GB200 NVL72 configuration treats the entire rack as a single logical GPU, connected via fifth-generation NVLink with an aggregate NVLink bandwidth of 130 TB/s across the system, enabling model parallelism strategies that were previously impractical within a single rack enclosure.

The system is purpose-built for trillion-parameter foundation model training and real-time inference, leveraging NVIDIA Blackwell GPU architecture with support for FP4, FP8, FP16, BF16, TF32, and FP64 precision formats. Each B200 GPU die delivers up to 20 petaflops of AI compute at FP4 precision, and the 72-GPU NVL72 configuration aggregates this to rack-scale throughput suited for next-generation large language models, multimodal AI, scientific simulation, and drug discovery pipelines. Liquid cooling is integral to the design, with Lenovo's direct liquid cooling infrastructure supporting the extreme thermal demands of the Grace Blackwell Superchip modules throughout the rack.

Lenovo integrates the GB200 NVL72 into the ThinkSystem SR680a V3 platform with enterprise-grade systems management via Lenovo XClarity Administrator, alongside support for the NVIDIA AI Enterprise software stack, CUDA, cuDNN, and the NVIDIA NeMo framework. The system is designed for deployment in hyperscale and enterprise AI data centers requiring maximum GPU memory bandwidth, NVLink fabric scale-out, and dense compute per rack unit, making it a flagship platform for organizations building sovereign AI infrastructure across EMEA, GCC, UAE, and APAC regions.

Ideal for

• Large-scale foundation model and trillion-parameter LLM training requiring rack-scale NVLink-connected GPU memory pools
• High-throughput generative AI inference serving for real-time applications including conversational AI and multimodal workloads
• Computational drug discovery and molecular dynamics simulation leveraging FP64 and mixed-precision Blackwell compute
• Sovereign AI infrastructure build-out for national AI initiatives, government agencies, and hyperscale cloud operators in EMEA and GCC
• Digital twin simulation and large-scale scientific HPC workloads requiring dense GPU interconnect and high-bandwidth memory
• Enterprise AI platform consolidation enabling simultaneous multi-tenant LLM fine-tuning and inference on shared rack-scale GPU fabric

Technical specifications

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