Abstract: Modern large-scale computing systems employ heterogeneous architectures consisting of multicore CPUs, general-purpose many-core GPUs, and programmable FPGAs. A primary challenge in utilizing these architectures is power consumption. Voltage reduction is one of the most efficient methods to reduce power consumption of a chip. With the galloping adoption of hardware accelerators in large datacenters, a comprehensive evaluation of safe voltage reduction levels for each chip can be employed for efficient reduction of total power. We present a survey of recent studies in voltage margins reduction at the system level for modern CPUs, GPUs, and FPGAs. The pessimistic voltage guardbands inserted by silicon vendors can be exploited in all devices for significant power savings. On average, voltage reduction can reach 12% in multicore CPUs, 20% in manycore GPUs, and 39% in FPGAs.
The Problem: The "Conservative Tax" of Hardware Guardbands
Modern chips are designed with a "one-size-fits-all" safety margin. Due to manufacturing variations (static) and fluctuations in temperature or workload (dynamic), vendors set the operating voltage ($V_{dd}$) significantly higher than what is strictly required for functional correctness. This extra "guardband" ensures reliability even in worst-case scenarios but imposes a high cost in power consumption.
As shown in the figure above, the nominal voltage is the sum of the best-case voltage level and a guardband. By reducing this margin, we can reclaim energy efficiency, though it requires careful management of the resulting timing margins to avoid system failure.
Methodology
The analysis employs a combination of physical experimentation and simulation across various platforms:
- Direct Measurement: Researchers performed "undervolting" on real hardware, including ARMv8 and x86 CPUs, NVIDIA GPUs, and Xilinx FPGAs, to identify $V_{min}$—the lowest voltage for correct operation.
- Modeling & Simulation: GPU analysis utilized GPUVolt (a voltage noise simulator) and GPUWattch to pinpoint the root causes of voltage fluctuations.
- Fault Characterization: For FPGAs, the study mapped three regions: the Guardband (safe), the Critical region (bit-flips occur, but the chip runs), and the Crash region.
Cross-Architecture Insights
CPUs: The Multi-core Stabilization Effect
In single-core CPU runs, $V_{min}$ varies significantly by program. However, in multi-core executions, the massive voltage droops (noise) caused by simultaneous core activity make $V_{min}$ more uniform and workload-independent. This suggests that multi-core environments may actually simplify the application of undervolting strategies.
GPUs: The Register File Noise Source
Analysis revealed that the Register File (RF) is the dominant source of voltage noise in GPUs. The massive register files required to support thousands of threads create significant power spikes during high access rates, which necessitates a larger voltage margin.
FPGAs: Temperature and Reliability Trade-offs
Counter-intuitively, increasing environmental temperature was found to reduce the fault rate in undervolted FPGA BRAMs. Furthermore, fault-tolerant applications like Deep Neural Networks (DNNs) can operate in the "Critical" region, tolerating minor bit-flips for massive power gains (up to 90%) with minimal accuracy loss (<3%).
Key Results
The consolidated findings demonstrate impressive quantitative gains across all tested architectures. By pushing past the nominal guardbands, significant reductions in $V_{dd}$ were achieved across various benchmarks.
- Voltage Reduction: Average reductions reached 12% in multicore CPUs, 20% in GPUs, and 39% in FPGAs.
- Power Savings: CPUs saw up to 25.2% savings (multi-core ARMv8), while GPUs improved energy efficiency by 25%. FPGAs achieved the most dramatic results, with up to 90% power reduction in BRAM-heavy workloads.
- Performance: These savings were realized with minimal performance penalties, typically ranging from 2.5% to 3.2% for CPUs.
Conclusion and Future Work
Exploiting hardware margins offers a path to significant energy efficiency in large-scale computing. However, pushing hardware to its limits risks "timing emergencies." Future work points toward "asymmetric resilience"—such as offloading GPU error recovery to the CPU—and more sophisticated hardware controllers capable of reacting to voltage droops in nanoseconds. As we move toward more parallel architectures, manycore-centric modeling will be essential to manage unique noise profiles and maximize efficiency.
Summary updated