NVIDIA's next-generation AI GPU 'Blackwell Ultra GB300' specifications have been leaked - 288GB HBM3e memory for 50% performance improvement.

Details of NVIDIA's next-generation AI server GPU 'Blackwell Ultra GB300', scheduled for release in the second half of 2025, have been revealed. Achieving significant performance improvements over the current GB200, increasing AI computing power by about 1.5 times while reaching an expected power consumption of 1,400W.

The B300 chip, which is the core of the GB300, adopts multiple innovative technologies that significantly improve AI workload processing capabilities. The most notable evolution is the expansion of memory capacity, increasing the HBM3e memory from 192GB to 288GB, a 50% enhancement. To achieve this, the memory stack structure has been expanded from the conventional 8 layers to 12 layers, enabling higher density memory implementation. In addition, the computing board adopts LPCAMM memory...

In terms of performance, particularly in FP4 operations optimized for AI inference processing, it has achieved approximately 50% performance improvement compared to the current GB200. This significant performance improvement is the result of a fundamental review and optimization of the chip architecture.
The network functionality has been significantly enhanced, and the evolution from ConnectX 7 to ConnectX 8 has improved data transfer capabilities. Additionally, the optical modules have been upgraded from 800G to 1.6T, doubling the bandwidth. This enhanced network functionality enables more efficient data transfers between multiple GPUs during the training of large-scale AI models.
On the architectural side, the adoption of socket configuration is being considered, which can lead to improved productivity and maintainability. However, this design change is expected to increase power supply and cooling requirements, thereby significantly affecting the overall system design. These technological innovations demonstrate a comprehensive approach to meeting the demands of rapidly evolving AI workloads.

The GB300 may coincide with the introduction of NVIDIA's next-generation architecture, known as 'Rubin.' Rubin is planned to adopt TSMC's 3nm process node and HBM4 memory, and is expected to be a significant turning point in semiconductor technology evolution.

Blackwell Ultra GB300 is facing unprecedented challenges in power management and cooling as a trade-off for its innovative performance improvements. Each B300 chip consumes an extremely high power of 1,400W, requiring a distinct power management system that sets it apart from traditional GPUs designed for data centers.
To ensure stable power supply and transient fault protection, NVIDIA has adopted a hierarchical power backup system. At its core are more than 300 supercapacitors equipped in each GB300 NVL72 cabinet. The production cost of each supercapacitor is estimated at $20 to $25, leading to a substantial cost for the power management system per cabinet. Additionally, the battery backup units (BBU) offered as an option incur a manufacturing cost of about $300 per module, amounting to a high cost of approximately $1,500 for the whole system.

The cooling system has been completely revamped with the adoption of a new water cooling plate design and enhanced liquid cooling system with high-performance quick disconnects (UQD). These improvements ensure the cooling capacity to handle the unprecedented heat generation of 1,400W. Particularly noteworthy is the transition of the overall system design to liquid cooling, impacting the selection of manufacturing partners. For example, Foxconn has been chosen as a key supplier for the GB300 due to its extensive experience in producing liquid cooling systems.
These innovations in power management and cooling systems represent significant technological milestones in the new era of high-performance computing. However, they also suggest the need to fundamentally reassess existing norms related to data center design and operation. The transition from conventional air-cooled data centers to liquid cooling systems necessitates a wide-ranging transformation from facility design to operational procedures. This transformation holds the potential to become a new standard for next-generation AI infrastructure.

NVIDIA's new product promises astonishing performance improvements, while also highlighting serious challenges. The 1,400W power consumption will place a significant burden on the electrical infrastructure of datacenters. Moreover, the adoption of over 300 supercapacitors raises new challenges in terms of reliability and cost.
However, what is more intriguing is the future direction indicated by this product. The increase in demand triggered by the AI boom is forcing the semiconductor industry towards a new paradigm shift. The current necessity to pursue performance improvements despite the increasing power consumption implies that the industry is approaching a turning point.