數據中心如何提升每瓦 AI 效能？

As AI models become larger and more demanding, data centers must deliver more computing power without letting electricity and cooling costs rise at the same rate. AI performance per watt measures how much useful AI computing performance a system delivers for the power it consumes. Improving it helps operators build more energy-efficient AI infrastructure while managing operating costs.

1. Choose AI Processors Designed for the Workload
CPUs, GPUs, AI accelerators, and custom ASICs have different strengths. For suitable workloads, custom AI chips and specialized accelerators can perform specific operations efficiently, potentially reducing unnecessary computation and power use.

What to evaluate:
AI training and inference performance
Power consumption under real workloads
Efficiency at different utilization levels
Compatibility with existing software and infrastructure

Note: The most efficient processor depends on the workload, system design, and software optimization, not simply the chip category.

2. Optimize Memory Bandwidth and Data Movement
AI processors need fast access to model parameters, training data, and intermediate results. If data cannot reach the processor quickly enough, compute resources may sit idle while still consuming power. Technologies such as High Bandwidth Memory (HBM) and efficient memory interfaces can reduce these bottlenecks in supported designs.

3. Use Advanced Chip Packaging and Integration
Approaches such as 2.5D packaging and 3D integration can support higher bandwidth, shorter data paths, and more efficient system integration, helping balance performance, power, and thermal requirements for large AI models.

4. Improve AI Networking and Interconnect Efficiency
Large-scale AI workloads often run across multiple accelerators and servers, and moving data between them can consume significant energy. Efficiency can improve through:

High-speed, low-latency interconnects and networking chips
Optimized communication between processors and memory
Efficient network topologies for distributed AI workloads
Careful workload placement to reduce unnecessary data transfers

5. Reduce Cooling and Infrastructure Power Overhead
Data center power includes more than the processors: cooling, power conversion, memory, storage, and networking all contribute. Operators can improve overall efficiency by:

Matching cooling capacity to actual heat output
Monitoring server inlet temperatures and airflow
Improving power distribution and conversion efficiency
Consolidating underused workloads where appropriate
Tracking Power Usage Effectiveness (PUE)

Note: PUE measures total facility energy relative to IT equipment energy. It complements, but does not replace, AI performance-per-watt measurements.

6. Measure Performance Under Real AI Workloads
Peak chip specifications don’t always reflect production efficiency. Benchmark representative AI models and track:

• AI performance per watt: useful AI output relative to power consumed
• Inference latency: time required to produce a result
• Throughput: requests, tokens, or tasks processed over time
• Power consumption at utilization: energy use under realistic conditions
• Total cost of ownership (TCO): hardware, electricity, cooling, integration, and maintenance

How Does MediaTek Support More Efficient AI Data Centers?

MediaTek’s data center work focuses on custom ASIC and XPU designs that look at efficiency across the whole system, from processor architecture to memory and connectivity. Real-world results depend on the workload, software, and operating environment.

Final Takeaway
Improving AI performance per watt takes more than one efficient chip. It comes from matching processors to the workload, keeping data moving quickly through memory and interconnects, managing cooling and power overhead, and benchmarking real workloads. The best measure is useful AI output for each watt and each dollar spent.