- Cooling costs rise as Nvidia increases power limits for future rack generations
- Compute trays dominate costs due to increasing demands for cold plates and thermal density.
- Savings on switch trays cannot offset the growing demand for high-end GPU trays.
Maintenance cost NvidiaCool's high-end rackmount systems continue to grow as each generation moves deeper into extreme power levels.
Morgan Stanley report obtained @Jukanlosreve reports that the liquid cooling hardware inside the GB300 NVL72 costs $49,860, which is about enough to buy a new one Tesla Model Y.
The report also estimates that the liquid cooling system required for the new Vera Rubin NVL144 configuration will cost around $55,710, an increase of 17%.
Cooling economics at tray level
This platform depends on the hotter Ruby GPUs up to 1800W per device with next generation NVSwitch 6.0 components.
The cost of cooling this system depends on the individual compute trays, and each compute tray will require larger capacity cooling plates.
The cost per compute tray is expected to increase 18% to approximately $2,660, and since the Vera Rubin NVL144 system has 18 trays, total compute-side cooling costs will reach approximately $47,880.
The increase comes at the expense of higher-capacity cold plates, the cost of which increases to $400 per unit. processors and GPUs are pushing thermal limits.
Switch tray cooling, meanwhile, appears to be less of a burden, dropping to $870 per tray and $7,830 per rack.
However this The reduction is overshadowed by a much larger leap in computing as the cost trajectory follows a pattern: cooling requirements for the GB300 NVL72 increased by 20% when moving from the GB200 NVL72.
Likewise, upgrading from the GB300 NVL72 to the Vera Rubin NVL144 adds another 17%. Power levels explain this trend.
Each Blackwell Ultra data center The GPU consumes 1400W, the Grace CPU consumes 300W, and the memory consumes 200W per socket.
As workloads grow, the value of precision cooling grows just as quickly, but future systems will make this even worse. Nvidia plans to move to Rubin Ultra GPUs, which can have TDPs of up to 3,600 W per case, and meeting this requirement could lead to new types of cold plates or more aggressive cooling methods.
Nvidia is also preparing a liquid-cooled NVL576 “Kyber” system that will include 144 GPU stacks and will offer better performance than the Vera Rubin NVL144, but will require even more cooling costs.
Although the final dollar amount has not been confirmed, high-power plates capable of rejecting 3.6 kW of heat would clearly exceed the current $400 per unit.
This is a signal that future data center installations will face even higher heat costs.
By using Toms Hardware
Follow TechRadar on Google News. And add us as your preferred source to get our expert news, reviews and opinions in your feeds. Be sure to click the “Subscribe” button!
And of course you can also Follow TechRadar on TikTok for news, reviews, unboxing videos and get regular updates from us on whatsapp too much.
