How to Get Your Data Center Ready for AI? Part One: Advanced Cooling

Industry insiders familiar with the natural progression of the modern data center will appreciate that the current artificial intelligence (AI) trend has accelerated rather than altered the evolution of this bedrock of the information technology (IT) world. As technology progresses, it is natural for server processors to grow more powerful and consume more power, and so the predominant method of cooling them with air conditioning was always going to run into a wall. The advent of AI spurred the adoption of state-of-the-art CPUs and GPUs characterized by unprecedented thermal design power (TDP). As a result, the inevitable day when air cooling will no longer be enough may come sooner instead of later.
 
Already, the most advanced AI chips on the market are barreling past the thermal limits of air cooling. At the time of writing, the average power density of an air-cooled server rack is well below 20 kilowatts (kW). A single NVIDIA H100 accelerator has a maximum TDP of up to 700 watts, while the next generation of enterprise GPUs, like NVIDIA B100 and B200, may have a maximum TDP of 1,000 watts. If we use GIGABYTE’s G593-ZD1 AI Server as an example, a 5U server can house eight GPUs. In other words, even a moderately populated rack (which offers 42 to 48 rack units of space) will generate more heat than what’s generally considered cost-effective for air cooling. Because of this, it should come as no surprise that NVIDIA’s rack-scale AI supercomputer, the GB200 NVL72, was designed from the get-go to be deployed with liquid cooling.

Learn More:
《Deep-dive into GIGABYTE’s Advanced Cooling Solutions》
《Visit GIGABYTE’s Artificial Intelligence Solution Page》
《Browse GIGABYTE’s AI Training and Inference Servers》

If the AI wave is the “push” that’s motivating data center operators to evaluate new thermal management tools, the “pull” must be the benefits of utilizing such tools, which can be summarized as upgrades to “performance”, “sustainability”, and “cost-efficiency”.

● Performance
The most sophisticated chip in the world won’t be able to unleash its full potential if it keeps getting too hot. When processors run at peak capacity, the heat that they generate must be quickly dispelled to avoid overheating and throttling. Effective cooling can ensure maximum performance as well as stable operations.
 
● Sustainability
Cranking up the AC for better cooling is a double-edged sword because it also consumes more energy, which hampers the data center’s power usage effectiveness (PUE) while ballooning the carbon footprint. The strength of advanced cooling is that it can dissipate more heat with less energy, allowing organizations to ramp up productivity while lowering emissions and moving toward ESG targets.

● Cost-efficiency
Procuring new cooling equipment may take a chunk out of the IT budget, but the smaller power bill will help decrease operating expenses (OpEx) in the long run. Improved operational stability also means less downtime and fewer components to repair and replace, which reduces total cost of ownership (TCO). Greater cost-efficiency is a competitive edge that no organization can afford to pass up on.

The three benefits listed above guarantee that advanced cooling will become a mainstay of data centers and server rooms around the world, regardless of whether AI continues to be the driving force behind its adoption. In the next sections, we will discuss the three primary methods of cooling: liquid cooling, immersion cooling, and air cooling—and how GIGABYTE can help you incorporate them into your IT infrastructure.

There is no doubt that liquid cooling, which pipes liquid coolant into servers through cooling loops and absorbs the heat from key components through cold plates, is gaining traction faster than ever thanks to the requirement of AI processors. As already stated, many AI supercomputers are designed with liquid cooling in mind. As data center operators, including major cloud service providers (CSPs), upgrade their infrastructure to adopt liquid cooling, it may only be a matter of time before this becomes a new standard in the industry.

Adding DLC to your data center may be easier than you think. You may start with the “liquid-to-air” option, which doesn’t require a significant overhaul of the traditional air cooling infrastructure. Hot coolant coming out of the servers are chilled with coolant distribution units (CDUs) that expel the heat into the hot aisle, in a manner that’s identical to air-cooled server racks. The “liquid-to-liquid” variant can dissipate more heat, but it requires the facility to have a built-in liquid-based cooling loop that draws water from an on-site source. A rear door heat exchanger (RDHx) can also be added to the back of the server rack for greater energy efficiency.

For data center operators who want to plan ahead and future-proof their infrastructure, immersion cooling is the logical next step after liquid cooling. By submerging the servers directly in a bath of nonconductive coolant, heat from the components can be dispelled through a CDU (which is the case for single-phase immersion cooling) or through the natural vaporization of the coolant inside the tank (two-phase immersion cooling). Immersion cooling can lower PUE to 1.02, which means only 2% more power than what’s used for computing is needed for cooling. However, the barrier to entry is that many aspects of the data center, from the physical infrastructure to safety permits, must be specially prepared for immersion cooling.

If immersion cooling is on your IT roadmap, there are two excellent reasons why you should choose GIGABYTE as your partner. One, GIGABYTE has a comprehensive product line which encompasses everything from immersion-ready servers to immersion tanks of both the two-phase and single-phase variety, available for EIA or OCP servers. Accessories such as the IT lift for lifting the servers vertically out of the tanks and IT dry racks for resting the servers on during maintenance, and verified coolant products for submerging the servers in, are also available from GIGABYTE and our partners.

While it is absolutely recommended that you consider liquid or immersion cooling to prepare for AI workloads, it is perfectly understandable that the process may be complicated, and so air-cooled servers will still be your primary source of compute for the foreseeable future. Fortunately, GIGABYTE’s air-cooled solutions are designed with top performance and optimal cooling in mind.

GIGABYTE’s air-cooled servers feature a proprietary airflow-friendly hardware design that provides the most optimized ventilation, as evaluated by simulation software. Powerful heat sinks are combined with specially designed air ducts to enhance heat dissipation. An automatic fan speed control system linked to sensors strategically placed inside the chassis adjusts fan speed according to the temperature of key components, allowing for unrivaled thermal control while remaining cost-effective and energy-efficient. RDHx can also be installed on air-cooled server racks to improve the data center’s performance and PUE.