Dell PowerEdge XE9712: NVIDIA GB200 NVL72-based AI GPU cluster for LLM training, inference

Dell has just unleashed its new PowerEdge XE9712 with NVIDIA GB200 NVL72 AI servers, with 30x faster real-time LLM performance over the H100 AI GPU.

2

Dell Technologies' new AI Factory with NVIDIA sees the GB200 NVL72 AI server cabinet with 30x faster real-time LLM performance, lighting-fast connectivity with 72 x B200 AI GPUs connected and acting as one with NVLink technology. Dell points out that the liquid-cooled system maximizes your datacenter power utilization, while rapid deployment will see your AI cluster at-scale, with a "white glove experience" adds Dell.

We have 25x more efficiency than Hopper H100, 8K for LLM training with the highest performance delta at 8K+ GPU clusters, and 30x faster real-time trillion-parameter LLM inference compared to the H100 AI GPU.

Arthur Lewis, president, Infrastructure Solutions Group, Dell Technologies said: "Today's data centers can't keep up with the demands of AI, requiring high density compute and liquid cooling innovations with modular, flexible and efficient designs. These new systems deliver the performance needed for organizations to remain competitive in the fast-evolving AI landscape".

• Read more: NVIDIA GB200 NVL72 AI server: 'highest-power-consuming server in HISTORY'
• Read more: NVIDIA 'halting developing' of GB200 NVL36x2 AI servers
• Read more: NVIDIA, Foxconn to build Taiwan's fastest supercomputer: with GB200 NVL72
• Read more: Supermicro confirms NVIDIA B200 AI GPU delay: offers liquid-cooled H200 AI GPUs instead

Dell explains that part of the Dell AI Factory with NVIDIA includes its new Dell PowerEdge XE9712 offers "high-performance, dense acceleration for LLM training and real-time inferencing of large-scale AI deployments. Designed for industry-leading GPU density with NVIDIA GB200 NVL72".

Dell continues: "This platform connects up to 36 NVIDIA Grace CPUs with 72 NVIDIA Blackwell GPUs in a rack-scale design. The 72 GPU NVLink domain acts as a single GPU for up to 30x faster real-time trillion-parameter LLM inferencing. The liquid-cooled NVIDIA GB200 NVL72 is up to 25x more efficient than the air-cooled NVIDIA H100-powered systems".

The company adds that "building on the success of the XE9680 with 8-way HGX GPUs, the XE9712 offers faster LLM performance with 72 GPUs acting as one in a single rack. It will be deployed as super PODs at scale, complete with full networking between racks, supported by Dell's turnkey rack-scale deployment services, supply chain, and logistics".

• 25x more efficient than the H100
• 30x faster real-time trillion parameter LLM inference vs. the H100
• Liquid-cooled architecture ensures efficient heat management, enabling higher performance and faster data processing