IRJET- Proposing a RTD-Based Block for On-Chip GPU Caches to Reduce Static Power Leaks

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 11 | Nov 2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 1021
Proposing a RTD-based block for on-chip GPU caches to reduce static
power leaks
Richa Sharma
Netaji Subhas University of Technology, New Delhi, 110078
----------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - GPUs dominate the industry in the parallel processing. The architecture of GPUs supports parallel algorithms and
process images at much higher rates than CPUs. Each generation of GPUs increases the use of on-chip caches and number of
processor cores embedded on-chip. Also, with each generation CMOS technology gets significantly worse in its leakage energy
consumption. This paper addresses the issues such energy leaks can cause in the efficiency of GPU computing. Itfurtherdelvesinto
Cache Power Management for GPUs. In this paper, the integrationofTunnelingdiodetechnology insidetheembedded architecture
of GPUs is proposed with a RTD switching block. Furthermore, the benefits of proposed integrations are discussed.
Key Words: GPU Caches, Cache Power management, StaticLeaks,ResonantTunnelingDiodes,NegativeDifferential Resistance,
RTD switching block
1. INTRODUCTION
Graphical Processing Units are used to process imagessincetheirinherentparallel structurebooststheperformanceofparallel
processing algorithms. In recent years, GPUs are being utilized for many general purpose computing applications of cloud
computing, machine learning, and other cost-effective High-Performance Computing(HPC) clusters.[1] In this era of green
computing, there has been a shift of focus in producing more high performance- energy efficient results than just achieving
highest peak performance. Achieving energy efficiency has become important in the design of all range of processors, such as
battery-driven portable devices, desktop or server processors to supercomputers.[2] The efforts to achieve this dual goal of
performance and energy efficiency, researchers have suggested various architectural modifications across different
components of the processor, such as processor core, caches, DRAM (dynamic random access memory) etc. Techniques like
power grating, DTCMOS have been applied to reduce static power leaks and cache-decay, drowsy cache has been employedto
reduce dynamic power leaks.[3] However, researchersfacethechallengeofscalingthe devices withCMOStechnologysincethe
leakage currents increase significantly with the decrease in the thickness of the gate oxide. The increase of subthreshold
currents is explained by thermodynamics, more specifically by Boltzmanndistribution.[4]Theproblemofpowerconsumption
with CMOS technology gets worse with each generation of processors. Theestimates ofInternational Technology Roadmapfor
Semiconductors (ITRS) indicates that leakage power consumption could become a major threat for the survival of CMOS
technology.[5] The competition in the development of GPUs and harnessing its computing power has driven companies to
increase the number of processing cores on the processors and their number will continue to rise. Furthermore, to bridge the
gap between the speed of the processor and main memory, thecachesoflargersizesarebeingembeddedonthechip.[6]Tables
1 and 2 list the cache memory requirements of current designs for different processors and fraction of energy consumed in
cache power out of total power consumption. In this paper, I focus on static energy leaksofGPUsthatiscontributedbyexisting
CMOS technology and propose to replace it by Tunneling Diode technology. Furthermore, I try to elaborate on the benefits of
this upgrade in cache power management of GPUs. GPUs are yet to realize their full computing potential. Since energy
management is one of the biggest challenges to overcome in the processor industry, our solution provides a new approach to
this problem of excessive energy consumption in processor chips.
Table-1: On-chip cache memory size in the latest generations Processors.
Processor Type Cache Memory Used
Desktop (CPU, GPU) 8 MB
Server Cores 24-32 MB
Mobile Processors 1 MB

Table-2: Power consumption of on-chip caches.
2. CACHE POWER MANAGEMENT IN GPUs
The power consumption of CMOS circuits is mainly classiﬁed in two parts, namely dynamic power (also called active power)
and leakage power (also called static power).
The mathematical modelling equations for both dynamic and leakage Power are given as the following
Here, dynamic and leakage power can also be written as active and static power respectively. Dynamic Power is dependent on
activity factor(number of cache accesses or number of bits accessed per cache access), effective capacitance, frequency of the
operation and voltage supplied. Static Power is dependent upon supply voltage and leakage currents, thus its consumption can
be reduced by decreasing either of these factors. Modern processors incorporate multi-level cache hierarchy with L1, L2, L3
caches. L1 is classified as First Level Caches (FLCs) and L2, L3 etc. are Lower Level Caches(LLCs).
First Level caches have higher dynamic power consumption rates and Lower Level Caches have higher leakage power
consumption rates. Since the design of FLCs and LLCs incorporates latency of caches differently in each of them. Techniques
developed to improve energy efficiency with Cache includesdevelopingahigh-impedancepath,sothatitcanturn“off”thecache
line when they hold data not likely to be reused. The period of non-usage of caches is referred to as “dead time”. The technique
used to predict “dead time” is cache decay which incorporates a counter to keep the count of pre-set cycles since its last usage.
[7]
3. RESONANT TUNNELING DIODE AND INTEGRATION OF RTD SWTICHING BLOCK IN GPU
3.1 Resonant Tunneling Diode(RTD)
Resonant Tunneling Diode(RTD), a variant of Tunneling Diode, invented by Esaki in 1957, is one of the most promising
quantum effect devices that are operational at room temperature. These diodes produce enhanced quantum tunneling effect,
which in turn produces very high-speed currents, which can be fine-tuned. Resonant Tunneling Diodes similar to CMOS
transistor turns “on”, conducts a current under a specific external bias voltage range.
The inherent difference between these two devices is that current in RTDs tunnelsthroughquasi-boundstateswithina double
barrier structure, instead of going through a channel betweendrainandsource.InotherTunnelingDiodeslikeEsakiTunneling
Diodes current tunnels through depletion region. Shown in the Fig. 1 is the IV characteristic curve of RTDs. The figure
represents the dynamic voltage vs current behavior.
The peak current occurs at the resonance at a specific voltage, then the device exhibits Negative Differential Resistance(NDR)
as the current continues to decrease with the increasing voltage and as it reaches a minimum “valley” current at a specific
voltage it starts to rise again. This minimum “valley” current canbeclassifiedasleakagecurrentis significantlylessthaninTDs.
The bistability of RTDs gives them an advantage over TDs, since TD’sproduceveryhigh leakagecurrentswhenthereverse bias
voltage is applied.
Processor Name Percentage of Total Power Consumption
Alpha 21264 16%
StrongARM 30%
Niagra,Niagra-2 L-2 Caches 24%

Figure-1: I-V characteristic of RTDs. Leakage currents are kept to a minimum due to the resonance and
bistability of the device, therefore providing advantage over TDs.
3.2 Applications of Negative Differential Resistance(NDR) in High-Switching Blocks
A huge advantage of RTDs is the ease of their integration with other technologies like complementary metal-oxide-
semiconductor (CMOS) and high electron mobility transistors (HEMTs). We can utilize the Negative Differential Resistanceof
the RTDs and create a switching block that will switch off and on at the application of specific voltages. This switching block
gives two operational points to switch the device on and off as shown in Figure. 2
These RTDs, resistor and CMOS blocks can provide high speed switching method for cache lines. [8] This provides smooth
integration that can work in tandem with cache decay technique. The counter will indicate when the cache line should be
deactivated, sending the high voltage signal to RTD, RTD will enter NDR region and turn “off” Cache line, hence implementing
cache decay. But due to this method, we could easily reactivate and switch back to low-impedance path way for cache line.
Figure-2: NDR property of RTD allows two stable operation points for switching.
The advantage of using NDR to turn on Cache lines lies in its low static power consumption since the leakage currents in RTDs
are minimal and the switching speeds achieved are in picoseconds.
3.2 Execution of Cache Decay by utilizing Negative Differential Resistance(NDR)
Cache line decay technique as proposed by Kaxiras Et al. [7] utilizes a pre-set counter to count a number of inactive cycles of
cache line and cuts-off that particular line thus dramatically saving leakage power. In thisimplementationcut-offsignal will be
sent to switching block a cycle earlier than proposed in Kraxis model.[7] Kraxis model utilizes gated Vdd CMOS transistor as a
switch as proposed by Powell Et al.,2000.[8]
Switching block of RTD will provide a stable point to cut-off power supply to cache line. Implementation of the block as
demonstrated in Figure 3 will provide a switch that will turn off at a certain input voltageandwill resumeoperation,according
to output Voltage.

Negative Differential Resistance property of RTD will execute in a manner of capacitance, and switching speed of block is
dependent upon Load voltage, shown in Figure 3 as Vout. Thus, it provides us with a probe to control switching speed of RTD.
Figure 3: (a) RTD- switching Block with CMOS transistor and Vout as a probe to control switching speed. (b)
Demonstration of RTD-block as capacitance and current source
4. CONCLUSION
The integration of RTD based structure in SRAMs will be beneficial on many accounts. The proposed utilization of NDR for
cache decay technique to reduce power will speed up the operation and therefore not compromise the performance of GPUs.
The leakage current can be maintained at very low levels and hence reducing leakage power substantially. Hence, I have
proposed the integration of RTDs with CMOS to improve the power efficiency of caches in GPUs and discussed the benefits of
utilizing Negative Differential Resistance(NDR) in reducing total power consumption of GPUs.
REFERENCES
[1] OLCF, http://www.olcf.ornl.gov/titan/, 2012
[2] S. Murugesan, “Harnessing green IT: Principles and practices,” IT professional, vol. 10, no. 1, pp. 24–33, 2008
[3] D. J. Frank, "The Limits of CMOS Scaling from a Power-ConstrainedTechnologyOptimizationPerspective," Nanohub, 4Oct
06.
[4] “International technology roadmap for semiconductors
(ITRS),”http://www.itrs.net/Links/2011ITRS/2011Chapters/2011ExecSum.pdf, 2011.
[5] Sparsh Mittal, “A Survey of Architectural Techniques For Improving Cache Power Efficiency”.
[6] Stefanos Kaxiras, ”Cache-line Decay: A Mechanism to Reduce Cache Leakage Power”
[7] J. L. Huber, J. Chen,” An RTD/Transistor Switching Block and Its Possible Application in Binary and Ternary Adders”
[8] Powell, M. D. et al. 2000. Gated-Vdd: A Circuit Technique to Reduce Leakage in Deep Submicron Cache
Memories. In Proc. of the Int'l Symp. On Low Power Electronics and Design'00 (2000).
BIOGRAPHY
Richa Sharma received the Bachelors of
Engineering from Netaji SubhasUniversity
of Technology, New Delhi formerly Netaji
Subhas Institute of Technology. Richa is
working on technology intensive areas of
physics research and creating innovation
in educational methods of these subjects.
Currently she works with Society of
AppliedMicrowaveElectronicsEngineering
& Research.

IRJET- Proposing a RTD-Based Block for On-Chip GPU Caches to Reduce Static Power Leaks

More Related Content

What's hot (19)

Similar to IRJET- Proposing a RTD-Based Block for On-Chip GPU Caches to Reduce Static Power Leaks (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Proposing a RTD-Based Block for On-Chip GPU Caches to Reduce Static Power Leaks