CoreWeave shows live demo of NVIDIA GB200 NVL72 rack under load
CoreWeave has presented one of the first live demonstrations of the NVIDIA GB200 NVL72 by a cloud provider, testing the 72-GPU rack's performance, cooling and energy use. The demo follows photos of the hardware the company shared the week before.
- CoreWeave ran one of the first live cloud-provider demos of the NVIDIA GB200 NVL72, a rack of 72 GPUs.
- Tests included the NCCL Allreduce Test, the GPUBlaze Test and a Megatron training run on CKS and SUNK.
- Dashboards showed GPU, memory, power and coolant temperature data in real time.
- The racks use NVIDIA Quantum-2 InfiniBand at 400Gb/s per GPU, with Quantum-X800 at 800Gb/s planned.
- Switch is CoreWeave's data center partner for the project.
Benchmarks and a training run
According to CoreWeave, the demo began with the NCCL Allreduce Test, a standardized benchmark meant to show the high-speed NVIDIA NVLink interconnectivity across the rack's 72 GPUs. The company says the test confirms that all GPUs communicate seamlessly, which matters for distributed workloads behind large foundational models.
Next came the GPUBlaze Test, in which the GPUs run complex matrix multiplication workloads. CoreWeave describes these as mimicking operations used in AI training, scientific simulations and advanced data processing.
To validate the system further, the company ran a Megatron training job using CoreWeave Kubernetes Service (CKS) and Slurm on Kubernetes (SUNK). According to CoreWeave, the job simulated the computational and memory demands of large foundational models. The company says CKS and SUNK are fully optimized for the GB200 NVL72. CKS is described as a hypervisor-free Kubernetes solution tailored for AI workloads, while SUNK extends it to burst and batch job scheduling.
Observability, cooling and power
The training run also made use of CoreWeave Observability Services. Through dashboards, the team tracked GPU utilization, memory and VRAM allocation, power consumption and temperature, which CoreWeave says gave real-time visibility into how the hardware handled the workload.
The rack's Cooling Distribution Unit (CDU) was a second focus. CoreWeave says it dynamically adjusts cooling output in response to GPU activity to keep the hardware at optimal temperatures without compromising performance. The CDU dashboard showed the return coolant temperature rising in sync with computational load, while a separate power dashboard gave a live view of how energy demand scales with performance. CoreWeave argues that this real-time monitoring highlights the system's efficiency and offers control over energy management.
Networking and what comes next
The GB200 NVL72 systems are supported by NVIDIA Quantum-2 InfiniBand networking, which CoreWeave says delivers 400Gb/s bandwidth per GPU through a rail-optimized topology. With NVIDIA Quantum-2's SHARP In-Network Computing technology, collective communication is offloaded, which the company says results in ultra-low latency and faster training.
CoreWeave also said it plans to deploy the NVIDIA Quantum-X800 800Gb/s InfiniBand platform in the near future, doubling the bandwidth to each GPU. The company thanked Switch, its data center partner, for its facilities and its role in the project.
What it means for customers
CoreWeave positions the GB200 NVL72, together with its product portfolio and engineering expertise, as a complete offering for AI development, scientific research and large data workloads. The source gives no pricing, availability dates or measured performance figures, and the demo results are presented by CoreWeave itself. Interested customers are directed to contact the company or attend an upcoming webinar.
