Growing demand for high-performance computing means that data center operators require increasingly efficient cooling systems for AI servers. According to TrendForce, the emergence of NVIDIA Blackwell by the end of 2024 will lead to the fact that the level of penetration of liquid liquids into the data center can increase to 10%. Blackwell solutions will account for about 83% of the company’s advanced products.
Individual (G)B200 accelerators will consume approximately 1000 Watts. HGX platforms will continue to combine up to eight accelerators, and NVL racks – 36 or 72 at once. Of course, all this will contribute to the growth of the supply chain for life support systems for AI servers, since traditional air cooling systems may simply not cope. Specifically, the GB200 NVL36 and NVL72 can consume up to 70 kW and 140 kW respectively. TrendFirce is confident that the NVL36 will use a combination of liquid and air cooling, but the NVL72 cannot do without a liquid cooling system.
Image source: NVIDIA
The agency identifies the key elements included in the GB200 rack LPS supply chain: water blocks, distribution units (CDUs), manifolds, quick disconnects (QDs), and rear door heat exchangers (RDHx). The main CDU supplier for NVIDIA AI solutions today is Vertiv, with Chicony, Auras, Delta and CoolIT being tested. According to rumors, NVIDIA has already encountered leaks due to low-quality components.
In 2025, deliveries of GB200 NVL36 should reach 60 thousand racks (a total of 2.1–2.2 million accelerators). At the same time, NVIDIA will begin offering HGX, GB200 Rack and MGX configurations to cloud providers and corporate clients, the supply ratio is expected to be 5:4:1. The GB200 Rack rack-mount options are designed primarily for hyperscalers.
Image source: TrendForce
TrendForce predicts that NVIDIA could introduce NVL36 as early as late 2024 and quickly enter the market, while NVL72 will not appear until 2025. However, the cloud giants will probably try to avoid being tied to a single supplier and may want to deploy HGX/MGX options on x86 chips or even use ASICs of their own design.
By the way, NVIDIA is looking for a materials chemist to develop submersible liquid cooling systems for high-performance accelerators. The specialist will be required to test fluids and materials for compatibility, cooling efficiency, corrosion resistance and environmental friendliness of materials to ensure optimal performance of new NVIDIA data center products.