At the upcoming GTC 2025 conference, NVIDIA is poised to introduce its groundbreaking GB300 AI servers, marking a significant shift toward fully liquid-cooled AI cluster solutions.
Unveiling the Blackwell Ultra GB300 AI Servers: Enhanced Cooling and Performance
This year’s GTC will see NVIDIA reveal its highly anticipated “Blackwell Ultra”series, a notable refinement over the existing Blackwell architecture. Although specifics about the GB300 AI servers remain limited, reports indicate a dramatic enhancement in their cooling mechanisms, promising to quadruple the efficiency over the current GB200 clusters. This upgrade is essential as the performance outputs of the Blackwell Ultra design are expected to generate substantial heat.
According to Taiwan Economic Daily, the GB300 lineup will be entirely liquid-cooled, moving away from air-cooled components. This transition has led to an escalation in demand for liquid-cooling systems in the Taiwanese supply chain, with NVIDIA anticipated to expedite Blackwell Ultra production due to heightened consumer interest compared to the original Blackwell series. The GB300 servers are anticipated to exhibit significantly improved heat dissipation capabilities, necessitating top-rated cooling solutions.
Liquid cooling implementation is likely to elevate the costs of the GB300 AI servers significantly. With the existing GB200 NVL72 servers priced around $3 million, it’s expected that the premium configurations of the GB300 series will surpass this cost. This move not only promises increased revenue for NVIDIA but also raises questions regarding market acceptance, especially in light of previous yield rate challenges associated with the Blackwell architecture.
Rumors suggest that the Blackwell Ultra “B300″series will prioritize power efficiency, featuring a total design power (TDP) of up to 1400W—a substantial increase compared to earlier models. Thanks to architectural improvements, users can expect around 1.4 times the floating-point performance (FP4) of the preceding generation, alongside enhanced memory capabilities that will rise from 192 GB to 288 GB, utilizing advanced 12-Hi stacks of HBM3E technology.
In addition to the GB300 series, NVIDIA is also set to unveil its Vera Rubin lineup at this year’s conference, providing insights into its development, although a market launch is not expected in the immediate future.
Related Articles:
NVIDIA & AMD May Be Required to Prioritize GPU Sales to U.S. Customers Before International Exports
6:49
Qualcomm CEO: Intel’s Chip Production Still Lacking, Yet Potential for Team Blue to Supply Chips Remains
18:40
HWiNFO Updates: Support for Nova Lake-S and Upcoming AMD Platform, Possibly Indicating AMD 900-Series for Zen 6
18:37
Leave a Reply