Lenovo ThinkSystem SR675 V3 with AMD Instinct MI300X 192GB

The Lenovo ThinkSystem SR675 V3 (7D9QCTO4WW) is a purpose-built 2U rack server engineered to support up to eight AMD Instinct MI300X 192GB accelerators, delivering one of the highest aggregate GPU memory capacities available in a production rack server platform. The AMD Instinct MI300X is an Accelerated Processing Unit (APU) built on AMD's CDNA 3 architecture, integrating 192GB of HBM3 memory per accelerator with 5.3TB/s of peak memory bandwidth. With eight accelerators installed, the system provides up to 1.5TB of unified GPU-accessible HBM3 memory, making it exceptionally well-suited for large language model (LLM) inference and training workloads that demand massive in-memory model capacity without model sharding across nodes.

The SR675 V3 platform is powered by dual AMD EPYC 9004 series (Genoa) processors, providing a high-bandwidth CPU-to-GPU interconnect fabric via PCIe 5.0 and supporting AMD Infinity Fabric links between accelerators. The server is designed to handle the substantial thermal and power demands of fully populated MI300X configurations, incorporating Lenovo's Neptune liquid-cooling-ready architecture alongside high-efficiency direct air cooling options. The chassis supports high-density NVMe storage, dual-port HDR100/NDR200 InfiniBand or high-speed Ethernet network adapters, and Lenovo XClarity Controller (XCC) for enterprise-grade out-of-band management and telemetry.

Positioned within Lenovo's ThinkSystem portfolio for AI infrastructure, the SR675 V3 with MI300X is qualified for deployment in hyperscale data centers, national research computing facilities, and enterprise AI factories requiring ROCm-compatible software stacks. The system supports AMD ROCm open software platform, enabling compatibility with popular AI frameworks including PyTorch and TensorFlow, and is validated for large-scale generative AI, high-performance computing (HPC), and scientific simulation workloads across industries including financial services, healthcare, energy, and government sectors.

Ideal for

• Large language model (LLM) training and inference — the 1.5TB aggregate HBM3 memory across eight MI300X accelerators enables deployment of 70B+ parameter models entirely in GPU memory without tensor parallelism across nodes
• Generative AI and foundation model development — supports multi-modal AI model training pipelines requiring extreme memory bandwidth and high FP8/FP16/BF16 compute throughput
• High-performance scientific computing and simulation — CDNA 3 architecture delivers high double-precision (FP64) compute performance for computational fluid dynamics, molecular dynamics, and climate modeling
• Enterprise AI inference serving — dense memory capacity and high bandwidth allow multiple large models to be co-resident, reducing latency for real-time inferencing in financial analytics and healthcare diagnostics
• RAG (Retrieval-Augmented Generation) and vector database acceleration — large unified HBM3 memory pool supports in-memory vector indexes and embedding pipelines at scale
• Sovereign AI and national research infrastructure — open ROCm software stack and EPYC platform provide a supply-chain-diversified, auditable GPU compute platform for government and public-sector HPC facilities

Technical specifications

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