Supermicro has announced new AI systems for large-scale generative AI featuring NVIDIA’s latest data centre products. This includes the NVIDIA GB200 Grace Blackwell Superchip, and the NVIDIA B200 and B100 Tensor Core GPUs. Supermicro is updating its current NVIDIA HGX H100/H200 8-GPU systems to accommodate the NVIDIA HGX B100 8-GPU seamlessly and to support the B200, thus ensuring quicker delivery times. Furthermore, Supermicro will expand its extensive NVIDIA MGX systems range with new models featuring the NVIDIA GB200, such as the NVIDIA GB200 NVL72, offering a comprehensive rack-level solution with 72 NVIDIA Blackwell GPUs. Additionally, Supermicro is introducing new systems including a 4U NVIDIA HGX B200 8-GPU liquid-cooled system.
Charles Liang, president and CEO of Supermicro, stated: “Our emphasis on building block architecture and rack-scale Total IT for AI has facilitated the design of next-generation systems tailored to the advanced requirements of NVIDIA Blackwell architecture-based GPUs. Our new offerings, such as the 4U liquid-cooled NVIDIA HGX B200 8-GPU based system and our fully integrated direct-to-chip liquid-cooled racks with NVIDIA GB200 NVL72, are built on the proven HGX and MGX system architecture by Supermicro and NVIDIA, optimising for the new capabilities of NVIDIA Blackwell GPUs. Supermicro possesses the expertise to incorporate 1kW GPUs into a broad spectrum of air-cooled and liquid-cooled systems, as well as the rack scale production capacity of 5,000 racks/month and anticipates being the first to market in deploying full rack clusters featuring NVIDIA Blackwell GPUs.”
Supermicro’s direct-to-chip liquid cooling technology will cater to the increased thermal design power (TDP) of the latest GPUs, fully unleashing the potential of NVIDIA Blackwell GPUs. Supermicro’s HGX and MGX Systems with NVIDIA Blackwell will serve as the foundational elements for the future of AI infrastructure, promising unprecedented performance for multi-trillion parameter AI training and real-time AI inference.
An extensive selection of GPU-optimised Supermicro systems will be prepared for the NVIDIA Blackwell B200 and B100 Tensor Core GPUs and validated for the latest NVIDIA AI Enterprise software, which includes support for NVIDIA NIM inference microservices. The Supermicro lineup includes:
- NVIDIA HGX B100 8-GPU and HGX B200 8-GPU systems
- 5U/4U PCIe GPU system with up to 10 GPUs
- SuperBlade with up to 20 B100 GPUs for 8U enclosures and up to 10 B100 GPUs in 6U enclosures
- 2U Hyper with up to 3 B100 GPUs
- Supermicro 2U x86 MGX systems with up to 4 B100 GPUs
For training massive foundational AI models, Supermicro is prepared to be the first on the market to release NVIDIA HGX B200 8-GPU and HGX B100 8-GPU systems. These systems, featuring 8 NVIDIA Blackwell GPUs connected via a high-speed fifth-generation NVIDIA NVLink interconnect at 1.8TB/s, double the previous generation’s performance. With 1.5TB total high-bandwidth memory, they will offer 3X faster training results for LLMs, such as the GPT-MoE-1.8T model, compared to the NVIDIA Hopper architecture generation. These systems also feature advanced networking to scale to clusters, supporting both NVIDIA Quantum-2 InfiniBand and NVIDIA Spectrum-X Ethernet options with a 1:1 GPU-to-NIC ratio.
Kaustubh Sanghani, vice president of GPU product management at NVIDIA, remarked: “Supermicro continues to introduce an impressive range of accelerated computing platform servers, finely tuned for AI training and inference, catering to any current market need. We collaborate closely with Supermicro to deliver the most optimised solutions to customers.”
For the most demanding LLM inference workloads, Supermicro is unveiling several new MGX systems built with the NVIDIA GB200 Grace Blackwell Superchip, which merges an NVIDIA Grace CPU with two NVIDIA Blackwell GPUs. Supermicro’s NVIDIA MGX with GB200 systems will achieve a substantial performance leap for AI inference, with up to 30x speed-ups compared to the NVIDIA HGX H100. Together with NVIDIA, Supermicro has developed a rack-scale solution with the NVIDIA GB200 NVL72, linking 36 Grace CPUs and 72 Blackwell GPUs in a single rack. All 72 GPUs are interconnected with fifth-generation NVIDIA NVLink for GPU-to-GPU communication at 1.8TB/s. Additionally, for inference workloads, Supermicro is announcing the ARS-221GL-NHIR, a 2U server based on the GH200 line of products, featuring two GH200 servers connected via a 900Gb/s high-speed interconnect. Visit the Supermicro Booth at GTC to learn more.
Supermicro systems will also support the forthcoming NVIDIA Quantum-X800 InfiniBand platform, including the NVIDIA Quantum-X800 QM3400 switch and the SuperNIC800, and the NVIDIA Spectrum-X800 Ethernet platform, comprising the NVIDIA Spectrum-X800 SN5600 switch and the SuperNIC800. Optimised for the NVIDIA Blackwell architecture, the NVIDIA Quantum-X800 and Spectrum-X800 will deliver the highest level of networking performance for AI infrastructures.
There’s plenty of other editorial on our sister site, Electronic Specifier! Or you can always join in the conversation by commenting below or visiting our LinkedIn page.
