Server Cooling Design

Liquid Cooling .......

Text

Chilldyne can quickly develop a liquid cooled solution for any server. So far, our highest density prototype was a 4 kW 1U server, but that is not even close to the limit of liquid cooling technology. We have cold plate designs for all the major chipmakers, and we can build prototypes of new models quickly in house. We have the know-how to run whatever programs are needed to exercise your server to its limits.

For a typical server conversion, we take a baseline measurement of the server air cooled and measure the component temperatures using an IR camera during a LINPACK run. Then we instrument the hottest components that don’t have internal temperature sensors and measure the temperatures during air cooling. Then we turn on the water cooling, slow down the fans (which normally slow down with a cooler CPU temperature) and make sure that all the temperatures are the same or lower than with air cooling. Any components which are running over 70 C are investigated to make sure that they are running within their factory specifications. If necessary, the local air flow velocity can be measured using a hot wire anemometer to verify that the local convection environment is within specifications.

color_header_bar_flip

Here are some typical IR photos of a dual socket server before and after liquid cooling.

Dual Xeon Skylake

IR Photo Fan Cooled

IR Photo Liquided Cooled

If any of the air-cooled parts run too hot with the lower fan speed, then we turn up the fans to keep them cooler. If necessary we can add cold plates to memory, SSDs or communication chips. Finally, we test the server for 24 hrs in a heated thermal test chamber to make sure that it works properly at the highest expected air and water temperatures.

For servers with multiple cold plates, the Chilldyne solution is to use a simple manifold with barbs. No internal dripless connectors are required.

color_header_bar_flip

Chilldyne has deployed 3400 liquid cooled GPUs starting in 2014.

GPU Server with 8 GPU's

With some GPU cards such at the Nvidia A100, liquid cooling results in 20% higher clock speeds, offering a performance benefit that is 5x more valuable than the cost of the liquid cooling. Our 8 GPU server needs no internal quick disconnects, simple barbs on a PVC manifold are all that is needed for a negative pressure system. Even if someone forgets to connect the tubing on one of the GPU cards, there are no leaks. This reduces the risk of hardware damage to a negligible level. We can build GPU cold plates quickly from drawings or using a sample of the GPU card. The cooling can extend to the memory and voltage regulators. With our internal model shop, we can build and test prototypes in a few weeks.