Cabinet Technology is Bridging the Efficiency of Air Cooling with the Performance of Liquid Cooling for HPC and AI Workloads

Print Friendly, PDF & Email

Self-contained Dynamic Density-Controlled Cabinets Allow for More Efficient and Scalable Operation of High-Density Servers

By Chris Orlando, Co-Founder and CSO, DDC

[SPONSORED GUEST ARTICLE]   Servers running high-performance computing (HPC) and artificial intelligence (AI) workloads are forcing data center operators to accommodate higher-density machines in smaller spaces while dealing with the complexity of managing widely varying workloads in a single facility or data center hall. These new deployments output more heat, increasing pressure for operators to find better cooling methods to ensure safe, fast, and reliable management of rapidly changing servers and storage platforms.

In the past, when rack power requirements rarely exceeded 20 kilowatts (kW), data centers could rely on traditional in-room air-based cooling to maintain safe operating temperatures. However, HPC and AI deployments in the United States are now running up to 50-70kW in a single rack, and traditional air-cooling methods can’t keep up.

Data center operators need to modernize and future-proof their facilities, so they don’t get left behind as AI and HPC deployments expand. While the industry is exploring several liquid cooling solutions, operators must find proven solutions that work for every customer and do not require massive retooling of the data center and the computing hardware being deployed.  In addition, data centers must scale from small multi-rack deployments to massive multi-megawatt installations without impacting existing operations or stranding data center resources. Self-contained, surgically controlled cabinet enclosures, which seamlessly bridge the gap between liquid and air cooling, are an innovative option for data center operators today.

Current Cooling Methods Aren’t Designed for the Data Centers of the Future

Cooling a data center is expensive. McKinsey analysts estimate that cooling accounts for 40 percent or more of a data center’s energy consumption. In the face of rising energy costs and hotter server hardware, the ability to effectively and efficiently cool high-density servers will play a vital role in the future success of nearly all data center operations.

The pressure rises even more as chip manufacturing companies like Intel, Nvidia, and AMD find new ways to pack more power into higher-density and hotter chips. The old way of cooling – pumping cooled air into a large room – cannot support these new increased densities, and while this approach is a common practice, it is not practical or efficient at supporting AI or HPC deployments.  As such, the industry is looking for alternative ways of designing data center cooling to deal with hardware density deployments in a single facility.

Liquid cooling is becoming increasingly popular as liquid is fundamentally more efficient at transferring heat than air. However, this cooling method poses its own set of risks and challenges. The primary challenges operators face are that (1) few, if any, standards exist for liquid cooling approaches, (2) many data center sites are not designed to support liquid deployments at the server level, and (3) data center workloads often vary greatly from application to application, so deciding how much air cooling vs liquid cooling creates opportunities for costly errors when planning for data center resources.  In addition, the inherent risk of bringing water into the hardware rack remains a genuine concern for many.

Rear-door heat exchangers have been used to help increase data center density, but the units’ cost, coupled with their overall density limitations and inability to manage temperature and airflow at the front or supply side of the server hardware, puts them at a significant disadvantage.  Further, these solutions often require that you maintain your existing air-cooling systems in the data center, which often results in higher long-term operating costs for the facility.  Today, new computing requirements are beginning to define how many CFM of airflow is required per kilowatt of power each server consumes to guarantee the right performance for these critical and expensive applications.  Rear door solutions, by nature, can’t address challenges at the front of the server, where surgical airflow and temperature management continue to become table stakes for HPC and AI deployments.

Bridging the Gap Between Air and Liquid

Despite rising interest in liquid cooling for AI and HPC environments, the demand for high-efficiency air cooling continues to grow, and the need to support lower and mixed-density workloads continues to evolve. Forecasts vary, but many of the highest-density future workloads will still have an overall air-cooled requirement of 10 – 40% of the total kW required.  With rack densities for AI and HPC being planned for 300+kW in a single cabinet, traditional air-cooled data centers may face 30-100kW of air-cooled demand.   And since the infrastructure needed to support both cooling methods is incredibly different, it can be challenging to retrofit a data center today that will also support workload needs into the future.

Chris Orlando, DDC

A solution that embraces the scalability of air cooling with the effectiveness of liquid cooling is dynamic density-controlled cabinet enclosures. Each cabinet has a dedicated cooling and heat exchange system at the rack level for a scalable bridge between air and liquid cooling. The cabinets use high-velocity fans in the front and the rear that push cold air to the supply side of the servers while pulling hot air from the back of the cabinet enclosure. This approach is called Dynamic Density Control (DDC) because the system automatically adjusts CFM levels and water fly requirements in real-time, based on the kW load of the cabinet itself as workloads change.   By focusing on the supply side of the server, and enclosing the cabinet into a two-part, cold and hot plenum, we remove the need to attempt to cool an entire data center floor wastefully.   By helping the servers move more air across the machine’s body, the solution improves server hardware performance and supports much higher rack-level density support.

The system uses completely enclosed cabinets with liquid heat exchangers outside the IT hardware area. No liquid is ever introduced into the cabinet environment itself, and the system is generally designed as a closed-loop, so no net new water is ever needed, nearly eliminating water waste and improving the system’s overall sustainability scores. Fire suppression at the cabinet level also protects increasingly expensive equipment and isolates potential incidents into a single cabinet environment. Advanced software monitoring and management at the rack level gives an unprecedented look at what’s happening in every rack at every moment of the day. The system moves water resources dynamically to wherever the heat load happens at any given time.

A Better Way to Manage and Scale Data Centers 

Cabinet enclosures act as self-contained building blocks so operators can add cabinets and better manage capital investments. A data center can have 6kW cabinets right next to 20kW cabinets and 85kW cabinets—all without sacrificing efficiency or stranding potential resources when under or overestimating high-density demand. Over the last decade, the DDC cabinet platform has been successfully deployed worldwide and within some of the largest colocation providers in the country to support highly specialized HPC-AI services.

Dynamic density-controlled cabinets are building a bridge into the future that allows customers to go beyond the kilowatt rating of a traditional airflow rack and scale operations many years into the future.

Heading to ISC High-Performance Conference

If you are heading to ISC in Hamburg, Germany, DDC Cabinet Technology will speak on this topic. Be sure to attend:

DDC- Bridging the Gap Between Air and Liquid Cooling
May 15 | 12:40 – 1:00 pm
Hall H – Ground Floor, Booth L01

DDC’s Chris Orlando, Co-Founder and CSO, will discuss how to manage cooling requirements driven by HPC, AI, and similar computationally intense workloads. In addition, stop by their booth, Stand A42, to view their data center solutions in action. Join DDC for happy hour on Tuesday, May 14, from 3 – 5 pm, and see how to future-proof your data center for HPC and AI workloads.

Not attending ISC? Visit www.ddc-cabtech.com to learn more.

Speak Your Mind

*