Current Trends in Power Usage

Modern data centers and even consumer PC hardware, such as high-performance CPUs and GPUs, are becoming increasingly power-hungry, driven by the surge in AI, machine learning (ML), and big data workloads and the strive for the most FPS or performance.

An article by Goldman Sachs illustrated the impact that high-performance computing, especially AI chatbots like ChatGPT, will soon contribute to record-breaking energy consumption numbers. This is likely due to the increased data usage outweighing the efficiency gains seen year in and year out. This shift has resulted in a sharp increase in energy consumption, marking data centers as one of the largest consumers of electricity worldwide.

There’s a shift with the escalating energy demands with silicon fidelity having diminishing returns and Moore’s Law starting to level. With that, let’s explore some pivotal changes in limiting your usage to create more energy-efficient and sustainable data centers or desktops.

The Growing Energy Demand

The rapid development of high-performance CPUs and GPUs has enabled unparalleled computing power but at a significant energy cost. AI and ML workloads, in particular, require substantial power to process vast amounts of data, exacerbating energy demands. Dedicated GPUs for the data center like the NVIDIA B200 and H200 use over 700W per GPU housing 8x GPUs in a single system.

On the consumer side, the flagship GPUs follow this trend with the NVIDIA RTX 4090 reaching upwards of 450W TDP. The processors are also a culprit of increased power draw, with the most notable being the Intel Core i9 14900K pulling an eye-watering 300W.

The rising energy costs around the world, as well as the environmental concerns, are becoming critical issues for data center operators. More and more companies are stressing the value of being more green and reducing their power draw. New use cases like generative AI and training large language models pushing hardware to the max, on full tilt 24/7. Because of these highly dense, highly computationally reliant workloads, our data centers are sucking down so much power.

Reducing power draw or taking a step towards efficiency can lower operational costs and mitigate the environmental footprint for contributing to global sustainability goals. However, the landscape shows that when energy usage is feasible and available, the only way to “reduce” energy consumption is to increase computing efficiency per watt.

Strategies for Regulating Energy Consumption

Hardware Optimization

There are two big ways how you can reduce energy usage through hardware optimization: choosing the right hardware or finetuning existing hardware.

Intel and AMD both have released efficiency-focused data center processors this generation: Intel Xeon 6 with E-cores and AMD EPYC 9005 with Zen 5c cores. NVIDIA is also pushing toward more efficiency by offering their NVIDIA Grace CPU, a purpose-built low-power CPU for the data center.

• Intel Xeon 6 with E-cores features dense processors with increased core counts while utilizing a core architecture suitable for many small and lightweight tasks. Individual cores occupy a single task in parallel, each drawing less power per core compared to their previous generation processors. Intel markets their Xeon 6 with E-core to encourage data centers to deploy these dense and efficient processors to consolidate legacy hardware for better efficiency and overall less TDP per socket.
• AMD EPYC 9005 with Zen 5c cores is AMD's version of extremely dense yet highly efficient processors. These processors feature a full-fat Zen 5 core but feature less cache and lower clock speeds. By increasing the density of a single socket, fewer systems need to be deployed and less power is needed to run hundreds of calculations simultaneously. AMD also has their EPYC 8004 which are efficient edge servers also perfect for head nodes and lightweight workloads and fewer cores.
• NVIDIA Grace released in the NVIDIA data center lineup featuring an ARM-based CPU with LRDDR5 memory which both are low power-based solutions. While NVIDIA's GPUs still consume tons of power, by pairing them with highly efficient and tailored processors, you can limit the total power per system.

Furthermore, there are methods in which processors and hardware to be undervolted while still performing at a very effective level. For example, the AMD Ryzen 7000X Processors initially released at 105W could be undervolted to about 65W and still deliver comparable performance. Just like overclocking can squeeze out extra performance numbers, underclocking may deliver convincing, almost equal performance in some instances while saving on power, reducing heat, and lowering noise levels.

Cooling Design

Efficient data center design plays a pivotal role in minimizing energy consumption. Thoughtful layout and infrastructure planning can enhance airflow, reducing the need for excessive cooling. Strategic placement of equipment, coupled with advanced cooling systems, ensures optimal thermal management. Some data centers are pivoting to new cooling methodologies such as liquid cooling or even submersion cooling where hardware is suspended in a liquid medium for increased passive cooling performance.

Modern data centers are also incorporating modular designs that allow for scalable and efficient operations when their computing infrastructure grows. This approach enables operators to align capacity with demand, ensuring energy is used only where it is needed, further driving efficiency and cost-effectiveness while also being vigilant of possible upgrades and future-proofing.

While cooling design will not reduce the total energy consumption, it can increase the power consumption efficiency as data centers, AI assistants, and reliance on generative AI continue to grow in the coming decade.

Conclusion

Data centers must address the challenges posed by rising energy demands by adopting energy-efficient practices and integrating innovative solutions. If your data center compute is a couple of generations old, it might be time to make a significant upgrade. Those systems that run slow and sluggish will only run longer, while a newer more energy-efficient can complete jobs faster, consolidate workloads into a single system, and increase overall efficiency and maximize the energy you are bound to use. Consider the new Intel Xeon 6 with E-core platforms as well as the AMD EPYC 9005 systems for better performance per watt or lower TDP per core performance!

It is important to realize that the growth in energy consumption and demand will only rise. Data center operators and even home users alike just need to find a way to value the energy they have into more efficient hardware that can do more per watt than before.

If you have any questions about how you can upgrade your computing to the latest and greatest platforms, feel free to contact us today or visit one of our pages and custom configure a system!