Mingxin

Beyond PUE: A Complete Metric System for Compute Center Efficiency

算力中心TCO数据中心算力建设

In the operation of compute centers, PUE (Power Usage Effectiveness) has long been regarded as the core energy efficiency metric. However, a single indicator cannot fully reflect the true efficiency of a compute center. A complete measurement system should cover four dimensions: energy efficiency, compute utilization, storage and network performance, and total cost of ownership (TCO) over the full lifecycle. Based on public data and measured results from the Mingxin FX100, this article provides a practical evaluation framework for decision-makers.

Why PUE Alone Cannot Define Compute Center Efficiency?

PUE only measures the ratio of total data center energy consumption to IT equipment energy consumption, and does not reflect the performance output of the IT equipment itself. For example, a compute center with a PUE of 1.2, if its GPU cluster suffers a 30% drop in inference throughput due to storage bottlenecks, may have lower effective compute output than a center with a PUE of 1.4 but optimized storage. According to IDC data, the global average PUE for data centers in 2025 is approximately 1.5, yet compute utilization (average GPU/CPU load) is generally below 60%. Therefore, efficiency measurement must shift from "energy saving" to "energy saving + efficiency enhancement."

The Four Core Dimensions of Compute Center Efficiency

1. Energy Efficiency: PUE and Carbon Intensity

PUE remains a fundamental metric, but it should be combined with carbon intensity (carbon emissions per unit of compute). For instance, a compute center using green electricity with a PUE of 1.3 may be superior to one using coal power with a PUE of 1.2. TrendForce data indicates that by 2026, approximately 35% of global hyperscale data centers will have adopted renewable energy.

2. Compute Utilization: Throughput and Latency

  • Inference Throughput: In long-context scenarios, KV cache storage becomes a bottleneck. In tests with a 480B model, the Mingxin FX100, using an NVMe-oF array, improved inference throughput by 29–40% (measured, report R2/R3). The baseline was a local NVMe single drive, with the FX100 achieving optimal improvement at 16 concurrent streams.
  • Time to First Token (TTFT): Reducing TTFT directly improves user experience. In tests with a 480B model at TP8 and three concurrent streams, the FX100 reduced TTFT p50 from 10.17–35.73s to 7.53–26.35s, a decrease of 26–32% (measured, report R2).
  • No External Memory Recomputation Acceleration: When models need to be loaded from disk, the FX100 achieved 8.6–20x acceleration. The recomputation baseline TTFT p50 was 149.5s (conc16), compared to 11.85s for the FX100 (measured, report R2).

3. Storage and Network Performance: Acceleration Ratio and Bandwidth

Storage performance directly impacts model loading and training efficiency.

  • Model Loading: On the Huawei Atlas 910B platform, the FX100 achieved 6.2–9.3x acceleration compared to NFS (measured, report R9). The loading time for DeepSeek-70B was reduced from 1399s to 150s.
  • Training Checkpoint Saving: In 8-card 32B LoRA training, the FX100 reduced saving time from 178s to 94s, with a 96% increase in sustained write bandwidth (measured, report R1).
  • LMCache Parallel Read: In a cold-read scenario for the Qwen2.5-32B model, TTFT dropped from 37.97s to 9.30s, with a 5.3x bandwidth improvement (measured, report R1).

4. Full Lifecycle TCO: Hardware and Operational Costs

TCO must account for procurement, electricity, maintenance, and downtime costs. Taking the Mingxin FX100 as an example, its fully configured reference price is approximately ¥371,200 (about ¥2,014/TB, FX100), while the FX200 has a lower unit cost (¥1,797/TB). In compute centers, storage acceleration can reduce GPU idle time, thereby lowering the TCO per unit of compute. According to industry estimates, GPU waiting time due to storage bottlenecks can account for 15–25% of total training time, and optimization can equivalently reduce TCO by approximately 10–15%.

How to Build a Practical Efficiency Evaluation System?

  1. Define Business Scenarios: Distinguish between inference and training workloads; the former focuses on TTFT and throughput, while the latter focuses on checkpoint write speed.
  2. Set Baselines: Use local NVMe or NFS as a reference, recording key metrics (e.g., TTFT p50, throughput, bandwidth).
  3. Introduce Acceleration Factors: For example, the FX100's 8.6–20x acceleration in no-external-memory recomputation can be directly converted into equivalent GPU utilization improvements.
  4. Quantify TCO Impact: Convert time savings into electricity and hardware depreciation savings. For instance, a checkpoint saving time reduction of 84s (measured, report R1) in an 8-card cluster can save approximately 200 hours of GPU time per year.

Conclusion

Compute center efficiency measurement must shift from a single PUE metric to a multi-dimensional system covering energy, compute, storage, and TCO. The Mingxin FX100 series, through NVMe-oF and KV cache acceleration, has demonstrated significant throughput improvements and latency reductions in real-world tests, providing a quantifiable path for compute center optimization. If you wish to evaluate the efficiency bottlenecks in your own compute center, please contact Mingxin for a gated joint test.

Key Q&A from This Article

Q: Can PUE fully reflect compute center efficiency? A: No. PUE only measures energy efficiency and cannot reflect compute utilization, storage performance, or TCO. It must be combined with metrics such as TTFT, throughput, and acceleration ratio.

Q: What real-world improvements does the Mingxin FX100 bring to compute center efficiency optimization? A: In 480B model inference, throughput improved by 29–40% (measured, reports R2/R3), and TTFT decreased by 26–32% (measured, report R2). Model loading achieved 6.2–9.3x acceleration compared to NFS (measured, report R9, Huawei 910B platform).

Q: How can the impact of storage acceleration on TCO be quantified? A: By reducing GPU waiting time, such as a checkpoint saving time reduction of 84s (measured, report R1), which equivalently lowers TCO per unit of compute by approximately 10–15% (industry estimate).

Generated by Mingxin's content engine with automated QC; headline numbers cite signed test reports (see the evidence library). Translated from the Chinese original. Questions or corrections: contact us.