⚡雷峰网•Freshcollected in 2h
SenseTime Introduces TPW for AI Data Centers
#ai-data-center#energy-efficiency#workload-scheduling#demand-responsetpw（tokens-per-watt）與算電協同agentsensetimetpw算電協同agentlingang aidc
💡See how TPW links every watt to useful Tokens and how flexible workloads can unlock more AI capacity.
⚡ 30-Second TL;DR
What Changed
TPW extends PUE by measuring useful Tokens generated per unit of electricity and incorporating time-of-use electricity costs.
Why It Matters
TPW could give AI infrastructure operators a more business-relevant efficiency metric than PUE alone, linking electricity costs directly to useful model output. If broadly adopted, workload flexibility and storage could become operational assets for both data centers and power grids.
What To Do Next
Instrument your inference and training jobs with power and GPU-utilization telemetry, then calculate TPW by workload before testing time-shift scheduling for batch jobs.
Who should care:Enterprise & Security Teams
Key Points
- •TPW extends PUE by measuring useful Tokens generated per unit of electricity and incorporating time-of-use electricity costs.
- •An eight-level observability system maps jobs, GPU utilization, rack loads, and building power consumption on a unified timeline.
- •The Agent classifies workloads by latency tolerance, shifting training, evaluation, and batch jobs to lower-cost or lower-load periods.
- •SenseTime reports that coordinated capacity planning, storage dispatch, and demand response can increase effective IT capacity by about 60% without adding power capacity.
- •During Shanghai's July demand-response event, the Lingang AIDC reduced midday load by 75%, offered 23 MW of flexible load, and released 46,000 kWh over two hours.
🧠 Deep Insight
AI-generated analysis for this event.
🔑 Enhanced Key Takeaways
- •The TPW metric is designed to address the 'energy-compute gap' where traditional PUE (Power Usage Effectiveness) fails to account for the actual intelligence output or token generation efficiency of AI clusters.
- •SenseTime's Agent-based architecture utilizes a closed-loop control system that integrates real-time grid pricing signals with internal GPU scheduling to minimize operational expenditure (OPEX).
- •The implementation at the Lingang AIDC leverages a proprietary 'Compute-Energy-Business' (CEB) orchestration layer that dynamically adjusts model precision and batch sizes based on power availability.
- •This initiative aligns with China's 'East Data, West Computing' strategy and local Shanghai government mandates for AI data centers to achieve higher 'Green Computing' certification standards.
- •The 60% increase in effective IT capacity is achieved by reducing 'idle-power' waste and optimizing the scheduling of non-real-time training tasks during peak grid demand periods.
📊 Competitor Analysis▸ Show
| Feature | SenseTime (TPW) | NVIDIA (DGX Cloud/Energy) | Google (TPU/Carbon Footprint) |
|---|---|---|---|
| Primary Metric | Tokens Per Watt (TPW) | PUE / TCO | Carbon Intensity / PUE |
| Focus | Grid-Interactive AI Load | Hardware/Software Stack | Data Center Efficiency |
| Scheduling | Grid-Aware Agent | Workload Orchestration | Carbon-Aware Scheduling |
🛠️ Technical Deep Dive
- The observability system utilizes a multi-layer telemetry stack that captures data at 1-second intervals across power distribution units (PDUs), server racks, and GPU kernels.
- The Agent employs a Reinforcement Learning (RL) model to predict grid demand-response events and pre-emptively migrate non-critical training workloads to secondary clusters.
- Integration with the Lingang AIDC power infrastructure involves a software-defined power management interface that allows for sub-millisecond load shedding without interrupting active inference tasks.
- The TPW calculation formula incorporates a weighting factor for token quality, ensuring that 'useless' or 'hallucinated' tokens generated during low-power states do not artificially inflate efficiency metrics.
🔮 Future ImplicationsAI analysis grounded in cited sources
TPW will become a standard KPI for AI data center operators in China by 2027.
The shift from measuring infrastructure efficiency (PUE) to output efficiency (TPW) is necessary for the economic viability of large-scale AI training as energy costs rise.
SenseTime will license its 'Compute-Energy' orchestration software to third-party data centers.
The demonstrated 60% capacity gain provides a strong commercial incentive for SenseTime to monetize its internal efficiency tools as a SaaS product.
⏳ Timeline
2022-01
SenseTime officially launches the SenseCore AI Infrastructure (SenseNova) in Lingang.
2024-05
SenseTime expands the Lingang AIDC capacity to support large-scale model training.
2026-07
Lingang AIDC successfully executes large-scale demand-response, reducing load by 75%.
2026-08
SenseTime formally introduces the TPW metric and Agent-based energy management system.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 雷峰网 ↗