Designing energy-efficient AI chips: Why power must be an early consideration

AI’s demand for compute is rapidly outpacing current power infrastructure. According to Goldman Sachs Global Institute, upcoming server designs will push this even further, requiring enough electricity to power over 1,000 homes in a space the size of a filing cabinet.

As workloads continue to scale, energy efficiency is now as critical as raw performance. For engineers developing AI silicon, the central challenge is no longer just about accelerating models, but maximizing performance for every watt consumed.

A shift in design philosophy

The escalation of AI workloads is forcing a paradigm shift in chip development. Energy optimization must be addressed from the earliest design phases, influencing decisions throughout concept, architecture, and production. Considering thermal behavior, memory traffic, architectural tradeoffs, and workload characteristics as part of a single power-aware design flow enables the development of systems that scale efficiently without breaching data center or edge-device energy limits.

Traditionally, design teams have primarily focused on timing and performance, only addressing energy consumption at the end of the process. Today, that strategy is outdated.

Synopsys customer surveys across numerous design projects show that addressing power at the architectural stage can yield 30-50% savings, whereas waiting until implementation typically achieves only marginal improvements. Early exploration enables decisions about architecture, memory hierarchy, and workload mapping before they become fixed, allowing trade-offs that balance throughput, area, and efficiency.

Architecture analysis as a power tool

Before RTL is finalized, a comprehensive power analysis flow helps reveal where energy is being spent and what trade-offs exist between voltage, frequency, and performance. Architectural modeling enables rapid evaluation of techniques—such as dynamic voltage and frequency scaling (DVFS), power gating to shut down inactive circuits, and optimizing data flow within the network-on-chip (NoC)—and supports smarter, more energy-efficient design choices.

Transaction-level simulation allows teams to measure expected workloads and predict the impact of configuration changes. This early insight informs hardware-software partitioning, interface sizing, and memory placement, all critical factors in the chip’s overall efficiency.

Data movement: The hidden power sink

Computation isn’t the only factor driving energy use. In many AI chips, data movement consumes more power than the arithmetic itself. Each transfer between memory hierarchies or across chiplets adds significant overhead. This is the essence of the so-called memory wall: compute capability has outpaced memory bandwidth.

To close that gap, designers can reduce unnecessary transfers by introducing compute-in-memory or analog approaches, choosing high-bandwidth memory (HBM) interfaces, or adopting sparse algorithms that minimize data flow. The earlier the data paths are analyzed, the greater the potential savings, because late-stage fixes rarely recover wasted energy caused by poor partitioning.

The growing thermal challenge

As designs move toward multi-die and chiplet architectures, thermal density has become a first-order constraint. Packing several dies into one package creates concentrated heat zones that are difficult to manage later in the flow. Effective thermal planning, therefore, starts with system partitioning: examining how compute blocks are distributed and how heat will flow through the stack or interposer.

By modeling various configurations early, before layout or floor planning, engineers can avoid thermally stressed regions and plan for cooling strategies that support consistent performance under load.

Optimizing the real workload

Unlike traditional semiconductors, AI chips are rarely general-purpose. Whether a device runs edge inference, data center training, or specialized analytics, its efficiency depends on how closely the hardware matches the target workload. Simulation, emulation, and prototyping before tapeout make it possible to test representative use cases and fine-tune hardware parameters accordingly.

Profiling multiple operating modes, from idle to sustained training, exposes inefficiencies that might otherwise remain hidden until silicon returns from the fab. And it helps ensure the design can maintain high utilization and consistent energy performance across all conditions.

Extending efficiency beyond tapeout

Energy monitoring and management must persist even after chips are manufactured. Variability, aging, and environmental factors can shift operating characteristics over time. Integrating on-chip telemetry and control using silicon lifecycle management (SLM) solutions allows engineers to track power behavior in the field and apply adjustments to sustain optimal performance per watt throughout the product’s lifecycle.

The next breakthroughs in AI hardware will come not just from faster chips, but from smarter engineering that treats power as a foundational design dimension, not an afterthought. For today’s AI hardware, efficiency is performance.

Godwin Maben is a Synopsys Fellow.

Special Section: AI Design

• The AI design world in 2026: What you need to know
• AI workloads demand smarter SoC interconnect design
• AI’s insatiable appetite for memory
• The AI-tuned DRAM solutions for edge AI workloads
• Designing edge AI for industrial applications
• Round pegs, square holes: Why GPGPUs are an architectural mismatch for modern LLMs
• Bridging the gap: Being an AI developer in a firmware world
• Why Power Delivery Is Becoming the Limiting Factor for AI
• Silicon coupled with open development platforms drives context-aware edge AI

The post Designing energy-efficient AI chips: Why power must be an early consideration appeared first on EDN.