Data center rack densities are rapidly increasing to support high-density AI workloads and high-performance computing (HPC). As a result, the data center power architecture is changing, with 800 VDC (direct current) emerging as the new standard for rack power to support these workloads. Alongside this shift, the cooling implications of 800 VDC architecture are becoming a critical consideration. As these changes take hold, they could place new demands on data center thermal management, requiring operators to adapt to high-density rack cooling.
Initial deployments of 800 VDC solutions do not necessarily impose a drastic shift in cooling infrastructure, however. The larger cooling architectural shift was already prompted a few GPU generations and a few hundred kW per rack ago. Data centers have already implemented or are planning to implement liquid cooling systems to handle heat increases. These systems are designed to scale with the growth of high-density computing (including 800 VDC racks) and can also cool the power infrastructure as it evolves.
Why 800 VDC matters for data centers
800 VDC addresses the challenge of powering racks approaching 400 kW and beyond—well past the limits of traditional AC and 48 VDC designs. As part of a broader shift toward 800 VDC, power is delivered to the rack in DC, moving AC-to-DC conversion out of the IT rack. In the near term, this is enabled by deploying 800 VDC power racks, or “sidecars,” without requiring major changes to upstream infrastructure in existing environments, including hyperscale data center power systems. In cooling terms, this is similar to how Liquid-to-Air CDUs were initially used to deploy liquid-cooled IT in air-cooled data centers.
How 800 VDC will impact the whitespace
Early adoption of 800 VDC will focus on the latest and highest-performing GPU racks, pushing the boundaries of power density. While heat generated by these racks can be supported by current liquid cooling approaches, the magnitude of heat rejected into the TCS loop will increase as row sizes increase. Further, the heat capture rate into the liquid loops in these racks will increase with more components within the rack being liquid-cooled. This includes 50 VDC in-rack busbars and power supplies, which will also move to liquid cooling to gain density.
In initial deployments, the power sidecars in the whitespace will need to be considered from a cooling perspective. First deployments will be air-cooled, and with high power ratings, they can reject over 20 kW of heat to the air, comparable to a higher-density air-cooled rack. Since the neighboring GPU rack will likely have limited active air cooling, most whitespace designs should be compatible with 800 VDC rack plus sidecar deployments from an air-cooling perspective.
How 800 VDC will transform data center infrastructure
Just as facility level liquid-cooling infrastructure was targeted to replace air-to-liquid CDUs, data center infrastructure is being pressured to provide 800 VDC directly and replace power sidecars. Centralizing the supply of 800 VDC can better enable the scale out of large clusters, but it does pose implications for cooling infrastructure.
When centralized 800 VDC power conversion is provided by power converters with low-voltage AC input, such as DC UPS or AC/DC Converters, AC power can be supplied to mechanical equipment (CDUs and CRAHs) from the same powertrain that serves 800 VDC to GPU racks. This is the easiest initial approach and would not require changes to cooling equipment.
As data center scale increases, there are drivers to convert directly from medium-voltage (MV) to 800 VDC, either via a solid-state transformer (SST) or a transformer rectifier unit (TRU). While this equipment may be air-cooled, like traditional electrical room equipment, pressure on power density will eventually bring about liquid cooled power infrastructure. In some cases, localized or smaller CDUs will be needed to manage thermal conditions between servers and power systems.
Converting directly from MV to 800 VDC for GPU powertrains has implications for how thermal and mechanical equipment can be powered. AC input power can be maintained if cooling equipment is powered from alternate powertrains from the GPUs. This is very likely for chillers and other large, centralized equipment that, for large AI factories, is likely already on dedicated powertrains. Further, networking and support equipment may necessitate maintaining some AC powertrains in the whitespace that could power smaller cooling equipment such as CDUs and CRAHs. Powering cooling equipment from the same powertrain as the 800 VDC IT can be facilitated with a dedicated inverter, retaining AC equipment, or deploying 800 VDC compatible CRAHs or CDUs as part of a larger shift to 800 VDC.
Preparing for the 800 VDC future
The transition to 800 VDC represents a significant step forward in how data centers power high-density AI workloads—but it doesn’t require a complete reinvention of cooling. In many ways, the industry has already laid the groundwork. The shift to liquid and hybrid architectures for AI and high-density compute has created the flexibility needed to support both today’s GPU-driven environments and the next phase of 800 VDC solutions powered by sidecar. Centralizing 800 VDC into the facility infrastructure can be more impactful to cooling system design, but by aligning power and cooling strategies early, operators can scale efficiently and capture the full benefits of AI. Learn more by exploring Navigating Liquid Cooling Architectures for Data Centers with AI workloads to assess and select the best approach for your AI data center liquid cooling strategy.
Add a comment