How Compute Rental Services Can Lower PUE

Lowering PUE in compute rental services hinges on coordinated optimization across three fronts—power delivery, cooling systems, and IT load—rather than upgrading a single piece of equipment. According to industry research from Uptime Institute, the key to data center energy efficiency lies in matching facility-side constraints (power, cooling, tier level) with actual IT-side load characteristics. Drawing on measured data from Mingxin's FX-series storage acceleration products in AI inference scenarios, this article explains why "lowering PUE" should first focus on improving compute utilization, before addressing infrastructure retrofits.

Why Should PUE Optimization in Compute Rental Services Start from the Load Side?

PUE (Power Usage Effectiveness) is calculated as total data center energy consumption divided by IT equipment energy consumption. Traditional PUE reduction approaches focus on cooling and power efficiency, but compute rental providers face a different reality: AI training and inference loads exhibit far greater power fluctuation than traditional cloud computing. According to public research from Epoch AI, the growth in AI compute scale is accompanied by rising power density, which directly increases the design redundancy required in cooling systems.

In measured results from report R2, Mingxin found that in long-context inference scenarios with a 480B-parameter model, KV Cache tiered acceleration improved throughput by 29–40% (concurrency level 8 as the lower bound at +29%, optimal operating point at concurrency level 16 as the upper bound at +40%). The significance of this data: under the same SLA, the GPU card-hours required to complete identical inference tasks decrease, so energy consumption per unit of compute naturally drops. The denominator of PUE (IT energy) remains unchanged, but the fixed overhead of cooling and power in the numerator is spread across more effective computation—this is a PUE reduction path unique to compute rental providers: optimize the load first, without retrofitting the facility.

Cooling System Selection: Air Cooling, Liquid Cooling, or Free Cooling?

The second battleground for lowering PUE is cooling. According to industry research from Uptime Institute, data center availability tiers (Tier I–IV) directly determine power and cooling redundancy design, and higher redundancy typically results in worse PUE. If a compute rental provider commits to Tier III or above availability, dual power feeds and N+1 cooling are required, which significantly raises the PUE baseline.

The platform used in Mingxin's FX-series storage product testing (8× AMD MI308X, 192GB HBM per card) is itself a high-density compute node, and its thermal design imposes clear constraints on facility cooling. According to Alibaba Cloud's official GPU instance family classification, different accelerator cards correspond to different cooling requirements and instance family divisions, which suggests to compute rental providers: when selecting hardware, GPU thermal specifications should be matched against facility cooling capacity, rather than pursuing raw per-card performance alone.

While liquid cooling can bring PUE below 1.1, its upfront investment and operational complexity are not friendly to small and mid-sized compute rental providers. A more pragmatic path: first reduce heat generation per unit of compute through load optimization (such as KV Cache tiering), then assess whether incremental cooling system retrofits are warranted. In Mingxin's measured results from report R2, TTFT p50 dropped from 10.17–35.73s to 7.53–26.35s, meaning GPUs spend less time at peak power, which directly reduces instantaneous load on the cooling system.

Power Delivery: From "Design for Peak" to "Design for Actual Load"

Power efficiency is another major component of PUE. Traditional data centers design UPS and transformer systems around peak IT power, resulting in poor efficiency under light load. According to Microsoft Azure's official billing documentation, cloud-side GPU virtual machines support on-demand, reserved, and spot pricing models, and the underlying physical infrastructure must also handle load fluctuations—confirming an industry consensus: power systems should operate in the 40–80% load range to achieve peak efficiency.

Compute rental providers can optimize the power delivery chain through two types of measures:

1. Load shaping: Schedule deferrable inference tasks (such as offline batch processing) into off-peak periods to raise the average load factor of the power system.
2. Storage acceleration to reduce peak current: In Mingxin's measured results from report R1, model inference loading was accelerated by 6.2–9.3× (Huawei Atlas 910B platform, DeepSeek-32B loading reduced from 691s to 112s), meaning the window from GPU cold start to full load is dramatically shortened, and the power system does not need to reserve excessive redundancy for prolonged peak current.

According to AWS EC2 on-demand pricing's official position, GPU instances are billed hourly, and price differences partially reflect power and cooling costs across instance families. When pricing services, compute rental providers should incorporate electricity savings from PUE optimization into their cost models, rather than using GPU procurement price alone as the pricing baseline—this requires establishing a complete accounting framework of "card-hour cost + electricity + cooling allocation."

Quantitative Assessment of Green Energy Efficiency: Measure First, Then Modify

The return on investment for lowering PUE varies by facility. Mingxin recommends that compute rental providers first conduct load-side measurements before deciding on infrastructure retrofit priorities. Mingxin's FX-series products support an approximately 10-week gate-based joint testing process (G1 arrival acceptance / G2 single-unit baseline / G3 main gate: TTFT reduction ≥25%, throughput +29–40% measured in-band / G4 72-hour stability), and this process applies equally to validating PUE optimization efforts: first measure the actual energy consumption curve after load optimization, then compare against the pre-retrofit baseline to calculate real energy savings in cooling and power systems.

It should be noted that PUE is only one metric of green energy efficiency, not the whole picture. According to industry research from Uptime Institute, Water Usage Effectiveness (WUE) and Carbon Usage Effectiveness (CUE) are emerging as complementary sustainability metrics for data centers. When marketing "green" credentials, compute rental providers should avoid using a single PUE number as the selling point, and instead disclose the full three-dimensional energy efficiency metrics.

Key Q&A

Q: What is the most effective path for compute rental services to lower PUE? A: Optimize IT load first (e.g., through KV Cache tiered acceleration improving throughput by 29–40%, measured in reports R2/R3), then assess cooling and power retrofits. Load optimization directly spreads infrastructure energy consumption across more units of compute.

Q: Is liquid cooling mandatory? A: No. Liquid cooling suits high-density facilities but carries high upfront costs. Small and mid-sized providers can first reduce peak power duration through load shaping and storage acceleration (e.g., model loading speedup of 6.2–9.3×, measured in report R9), then decide whether to upgrade cooling.

Q: How can the effectiveness of PUE optimization measures be verified? A: A gate-based joint testing process is recommended: first measure the actual energy consumption curve after load optimization, compare against the pre-retrofit baseline to calculate energy savings, and avoid blindly retrofitting infrastructure based on experience alone.

References

1. Uptime Institute Resource Page — https://uptimeinstitute.com/resources
2. Epoch AI — https://epoch.ai/
3. EC2 On-Demand Instance Pricing — https://aws.amazon.com/ec2/pricing/on-demand/
4. Pricing - Linux Virtual Machines | Microsoft Azure — https://azure.microsoft.com/en-us/pricing/details/virtual-machines/linux/
5. Compare GPU Instance Families for AI, HPC & Rendering - Elastic GPU Service - Alibaba Cloud — https://www.alibabacloud.com/help/en/ecs/user-guide/gpu-accelerated-compute-optimized-and-vgpu-accelerated-instance-families