- US - English
- China - 简体中文
- India - English
- Japan - 日本語
- Malaysia - English
- Singapore - English
- Taiwan, China - 繁體中文
With the rapid emergence of AI technology, data centers face an ongoing challenge — how to maximize compute performance while lowering power. Electricity consumption from U.S. data centers and AI could triple by 2028, driving enormous growth in our nation’s energy demand. In 2023, U.S. data centers consumed an estimated 176 terawatt-hours (TWh) of electricity. Projections estimate that, by 2028, that number could rise to 580 TWh, which would represent 12% of total electricity use in the U.S..1 and 3.3 times more energy use in just half a decade.
Driven by the expansion of AI and other data-intensive applications, this expected surge underscores the importance of advanced hardware technologies to support increasing energy needs of data center infrastructures in both the U.S. and worldwide.2 Through the development and adoption of innovative, low-power (LP) memory architectures like Micron® LPDDR5X, data centers can deliver substantial performance gains without the energy penalty of traditional DDR5 memory.
Why LP memory?
Micron® LPDDR5X is engineered to deliver high-speed performance while consuming much less energy. Unlike traditional memory technologies like DDR5, LP memory operates at lower voltages, which improves both power and energy efficiency through:
- Reducing power consumption
- Lowering heat generation
- Optimizing circuit designs focused on energy savings
For AI-driven data centers, achieving gains in power and energy efficiency is an ongoing challenge. Consider Llama 3 70B running inference in a large-scale customer support environment. A single GPU manages a complex dance of AI interactions, simultaneously handling thousands of intricate customer queries in real time. The use of LP memory transforms this intensive computational workload into a markedly more energy-efficient process.
Figure 1: Normalized inference throughput Llama 3 70B
Inference performance
Our results revealed key performance gains when we tested LPDDR5X memory (on the NVIDIA GH200 Grace Hopper Superchip with NVLink) to a traditional DDR5 (on an x86 system with a PCIe-connected Hopper GPU). When we tested inference performance with Meta Llama 3 70B, the LP memory system delivered:
- 5 times higher inference throughput
- Nearly 80% better latency
- 73% less energy consumption
Figure 2: Normalized latency Llama 3 70B
In a time of increasing computational demands from AI applications along with environmental consciousness, low-power memory is more than just a technology upgrade — it's a strategic imperative for modern data centers. In practice, LP memory technologies benefit data center economics by simultaneously reducing power use and lowering operational costs. Reduced power needs translate directly into lower cooling requirements and electricity expenses. For data center operators, these improvements mean smaller utility bills and a significantly reduced carbon footprint. Moreover, the power and performance gains extend beyond operational efficiency. With higher throughput and better latency, users can enjoy a more seamless experience characterized by improved response times.
In a time of increasing computational demands from AI applications along with environmental consciousness, low-power memory is more than just a technology upgrade — it's a strategic imperative for modern data centers. In practice, LP memory technologies benefit data center economics by simultaneously reducing power use and lowering operational costs. Reduced power needs translate directly into lower cooling requirements and electricity expenses. For data center operators, these improvements mean smaller utility bills and a significantly reduced carbon footprint. Moreover, the power and performance gains extend beyond operational efficiency. With higher throughput and better latency, users can enjoy a more seamless experience characterized by improved response times.
Figure 3: Energy efficiency for LLM inference
The future is energy efficient
As AI marches on, continuously pushing the boundaries of compute and memory in data centers, advanced memory technologies like LPDDR5X are emerging as enablers of sustainable computing by allowing data centers to operate more efficiently. Speeding up the performance of AI tasks like inference while reducing power requirements will allow us to do more with less. The future of AI can be energy efficient as LP memory proves to us that we can push the boundaries of AI performance while simultaneously reducing our carbon footprint, ultimately leading to a more sustainable path forward for AI.
Learn more
- Download our comprehensive technical brief, The role of low-power (LP) memory in data center workloads, to learn how low power memory is transforming data center performance and energy use in next-generation AI infrastructure.
- For more information on low power memory technologies, check out our LPDDR5X product page.
1. U.S. Department of Energy. (2024). DOE releases new report evaluating increase in electricity demand from data centers. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
2. International Energy Agency. (2024). Electricity 2024: Executive summary. https://www.iea.org/reports/electricity-2024/executive-summary
