NVIDIA Launches GB300 NVL72 with Enhanced AI Scheduling Features
NVIDIA's GB300 NVL72, featuring 72 Blackwell Ultra GPUs and new NVLink Domain-Aware Placement Groups, boosts AI performance significantly. This development allows for optimized GPU workloads, essential for training large language models and other complex AI tasks.
Theia Market Signal Identification - AI Assisted
The GB300 NVL72 AI platform incorporates 72 Blackwell Ultra GPUs and 36 Grace CPUs, achieving up to 1.1 exaFLOPS of FP4 compute. The introduction of NVLink Domain-Aware Placement Groups enhances scheduling efficiency by colocating processes within a single NVLink domain, reducing inter-node communication overhead.
Testing demonstrated a 1.13x speed increase in iteration rates for Vision Language Action workloads. NVIDIA's advancements align with a growing demand for AI supercomputers, as shown by Microsoft's deployment of over 4,600 GB300 GPUs in 2025.
Kog, a French startup, is also focusing on optimizing conventional GPUs to accelerate AI inference, achieving 3,000 tokens per second with a small model. Their approach potentially offers cost-effective solutions for AI applications without requiring new hardware investments.
Comments