Nvidia GB300 Blackwell Ultra for AI at GTC, Will Be Fully Liquid-Cooled | Types of computer hardware | Gpu vs cpu architecture | Cpu hardware list and their functions | Turtles AI

 NVIDIA will introduce GB300 “Blackwell Ultra” AI servers at GTC 2025, marking a significant transition to fully liquid-cooled AI clusters. These servers, with a TDP of up to 1400W, will offer superior performance through four times the cooling efficiency of previous models.

Key points:

• Integral liquid cooling: GB300 servers eliminate air-cooled components, adopting an all-liquid system to handle the increased heat generated.
• Increased TDP: With a TDP reaching 1400W, the GB300s represent a significant increase over their predecessors, requiring advanced cooling solutions.
• Memory Capacity Expansion: Using HBM3E technology with 12-Hi stacks, memory increases from 192 GB to 288 GB, improving performance in complex AI applications.
• Market impact: The adoption of liquid cooling systems could increase the cost of GB300 servers, affecting customers’ purchasing strategies and NVIDIA’s revenue.

NVIDIA is set to unveil its next-generation GB300 AI servers, known as “Blackwell Ultra,” at GTC 2025, marking a step toward fully liquid-cooled AI clusters. The technology evolution eliminates traditional air-cooled components, adopting an all-liquid cooling system to handle the significant heat increase generated by the new servers’ high performance. According to Taiwan Economic Daily, the Taiwanese supply chain is seeing growing demand for liquid cooling components, as NVIDIA plans to ramp up production of the Blackwell Ultra to meet market interest that has exceeded that of the original Blackwell series. The GB300 specifications list a TDP of up to 1400W, a substantial increase over previous models, making advanced cooling solutions essential to ensure optimal performance and reliability. Additionally, the updated architecture promises a roughly 1.4x increase in floating point performance (FP4) over the previous generation, and memory capacity expanded from 192GB to 288GB, thanks to the implementation of HBM3E technology with 12-Hi stacks. However, the integration of liquid cooling systems could lead to an increase in the cost of GB300 servers. Considering that the current GB200 NVL72 servers are priced at around $3 million, it is expected that the flagship configurations of the GB300 series will exceed this cost, influencing customer purchasing strategies and NVIDIA’s revenue. Despite previous challenges related to the Blackwell architecture’s rate of return, NVIDIA’s strong presence in the market suggests that widespread adoption of the new servers is likely. In addition to the GB300 series, NVIDIA plans to provide updates on the development of the Vera Rubin lineup at GTC 2025, although an immediate market launch is not expected.

This event represents an opportunity for NVIDIA to consolidate its position in the AI ​​and high-performance computing sector, offering advanced solutions that respond to the growing needs of the market.