TL;DR
This paper analyzes the dynamic power consumption of a newly developed Full Posit Processing Unit (FPPU). The authors present a method to model this power consumption with high accuracy and demonstrate that the overhead for power monitoring hardware is relatively small, making it a viable option for energy-efficient computing.
What Problem Does This Paper Address?
The Posit number format, introduced as an alternative to the widely used IEEE 754 floating-point standard, offers potential advantages in terms of accuracy and efficiency, particularly for applications like deep learning. However, understanding the practical energy implications of hardware implementations of Posit arithmetic is crucial for their adoption. This paper addresses the need to characterize and model the dynamic power consumption of a full Posit Processing Unit (FPPU), providing insights into its energy efficiency and the feasibility of integrating power monitoring capabilities.
Methods
The authors employ a three-stage power identification flow, building upon existing methodologies [21].
- Synthesized Netlist and Simulation: The FPPU design, synthesized for FPGAs, is simulated using various benchmarks. These benchmarks are designed to stress different parts of the architecture and include both active operations and idle states.
- Data Extraction: During simulation, two types of files are generated:
- SAIF (Switching Activity Interchange Format): Captures all switching activity information.
-
VCD (Value Change Dump): Records the values of signals over time. From these, two key metrics for switching activity are extracted:
-
Hamming Weight Count (HWC): For data signals, representing the number of bits that change state.
- Single Toggle Count (STC): For control signals, indicating the number of times a signal changes state, regardless of its bit width. This distinction is made because control signal toggles are more indicative of operational changes than the number of bits changing.
- Model Identification: A linear regression model is used to correlate the extracted switching activity (HWC and STC) with the measured power consumption. This results in an equation where the estimated power at any given time is a weighted sum of the switching activities of classified signals. The model can be constrained by available resources and the depth of the design hierarchy for identification.
Datasets
The paper does not explicitly mention or introduce new datasets. Instead, it uses "benchmarks" which are described as the "basic arithmetic operations" (ADD, SUB, MUL, DIV) along with "no-op periods" to represent operative and idle states. These benchmarks are applied to different configurations of the FPPU.
Key Insights and Ideas
- Posit Power Modeling is Feasible: The study demonstrates that dynamic power consumption of a full Posit Processing Unit can be accurately modeled using switching activity metrics.
- Small Area Overhead for Power Monitoring: Integrating a power monitoring unit alongside the FPPU incurs a relatively small area cost, ranging from approximately 5% for 32-bit FPPU to around 30% for 8-bit FPPU. This suggests that energy-aware design can be implemented without significant hardware penalties.
- Posit Configuration Impacts Power and Area: Different Posit configurations (varying
nbitsandesbits) exhibit distinct power consumption and resource utilization characteristics. For instance, the 8-bit Posit configuration shows a higher area degradation relative to its power consumption. - Accuracy vs. Efficiency Trade-off: The paper hints at a trade-off between accuracy and efficiency, noting that certain configurations (like
esbits=2) are more efficient for deep neural network applications.
Key Results
- Model Accuracy: The dynamic power consumption of the FPPU can be modeled with an acceptable approximation error, ranging from 2.84% for the 32-bit FPPU to 7.32% for the 8-bit FPPU.
- Area Cost of Power Monitoring: The additional hardware for power monitoring adds an area cost from ~5% (32-bit FPPU) to ~30% (8-bit FPPU) of the total unit area.
- FPGA Resource Utilization: Table 1 and Table 3 provide detailed breakdowns of FPGA resource utilization (LUTs and FFs) and synthesis frequency for various Posit configurations across different FPGA families (Artix-7, Kintex-7, Spartan-7). For example, on an Artix7-100T, a P32E2 configuration uses 4049 LUTs and 520 FFs at 25 MHz, while a P8E0 configuration uses 718 LUTs and 238 FFs at 50 MHz.
- Error Metrics: Table 4 shows the RMSE (Root Mean Square Error) in mW and normalized RMSE (%) for different FPPU configurations, with normalized errors generally below 7.32%.
Limitations and Future Work
- FPPU 8 Degradation: The paper notes that the 8-bit Posit configuration (FPPU 8) exhibits higher error and area degradation compared to other configurations. This is attributed to its very low power consumption, making the error metric more sensitive to outliers.
- Comparison with Traditional FPUs: The authors plan to compare their FPPU with traditional Floating-Point Units (FPUs) to highlight the pros and cons of the Posit approach in hardware design, especially for intensive computing workloads.
- Integration into Real CPUs: Future work includes integrating the FPPU into a real-case CPU to evaluate its performance on standard CPU benchmarks and during deep neural network training.
- Further Exploration of Power Monitoring: The authors suggest exploring the integration of the power monitoring logic into a real case CPU to evaluate performance while executing standard CPU benchmarks or training deep neural networks.
Summary updated