Font Size: a A A

Voltage Noise Characterizing And Smoothing Design Research For MCM-GPU Architectures

Posted on:2024-05-31Degree:MasterType:Thesis
Country:ChinaCandidate:K Y ChenFull Text:PDF
GTID:2558307064485094Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
As Moore’s Law gradually slows down,it becomes more and more difficult to increase processor computing power by increasing the area of a single chip.At the same time,applications in deep learning,autonomous driving,graph processing and other fields are still increasing the arithmetic power demand for GPU-based highperformance computing systems.In this context,MCM(Multi-Chip-Module)technology is gaining attention as it can break through the limitation of single-chip area to realize the expansion of GPU processor arithmetic power.The MCM-GPU can achieve a significant increase in performance without increasing the area of the single chip.Compared with traditional single-chip GPUs,MCM-GPU has a larger scale,and in addition to the power consumption inside the processor chip,the communication of different chip modules at the package level generates additional power consumption in the off-chip area,thus leading to serious energy efficiency challenges for MCM-GPU.The voltage noise is an important factor affecting the energy efficiency of the processor.Voltage noise can cause voltage instability during the execution of the program.In order to guarantee the normal execution of programs,processor systems are usually designed with voltage guardband to reduce the impact of large voltage noise.Since the frequency of large voltage noise is low,the voltage protection band will cause the energy efficiency of the system to be wasted in most cases.In this paper,we conduct a voltage noise characterization and smoothing design study for MCM-GPU architecture,aiming to effectively reduce the voltage protection band of the processor by smoothing the voltage noise to achieve the improvement of system energy efficiency.First,we model the MCM-GPU architecture to simulate the performance,power,and voltage characteristics of the program as it executes on the MCM-GPU.Among them,the on-chip performance and power simulation is extended with the existing GPU architecture,and the off-chip region simulation increases the communication latency and power consumption between GPMs.For voltage simulation,we design the power delivery network(PDN)of the MCM-GPU architecture accordingly,scaling the electronic component parameters of PDN to capture the voltage information of different regions on the processor according to the thermal design power(TDP)of the processor.Then,we characterize the voltage noise of the MCM-GPU.The maximum voltage noise for different benchmark programs is first analyzed,and the higher voltage noise amplitude is observed in the MCM-GPU than in the monolithic GPU,and the reason for it is argued in detail.Then,in the analysis of the voltage noise distribution in the time dimension,it is found that the distribution frequency of the substantial voltage noise is lower,and the voltage noise amplitude is smaller for most of the time of program execution.In the analysis of voltage noise in spatial dimension,we found that the voltage noise trends of different GPMs are similar but there are obvious cycle differences,as well as non-uniform power distribution among different GPMs and SMs.After that,we investigate some sensitivity parameters of the MCM-GPU architecture to observe their effects on the voltage noise magnitude.Finally,we propose two smoothing techniques,including GPM aware mitigation(GAM)technique and droop magnitude aware smoothing(DMAS),technique by combining the voltage noise characteristics of MCM-GPU and the causes of generation.The two techniques are used to mitigate the voltage noise by the variation of the number of warp within the GPM and the variation of access instructions with different voltage noise magnitudes,respectively.We then perform a detailed evaluation of the two techniques in several dimensions of voltage noise smoothing effectiveness,performance,and energy consumption,as well as a different sensitivity parameter evaluation study.The experimental results show that the GPM-aware smoothing technique and the voltage noise amplitude aware smoothing technique are able to effectively reduce voltage drops,improve voltage stability,and save energy consumption with low performance loss at an average smoothing amplitude of 29.4%and 38.1%,respectively,under the baseline configuration.
Keywords/Search Tags:MCM-GPU, voltage noise, energy efficiency, GPU architecture
PDF Full Text Request
Related items