Thermal Management - Design Automation Conference

Report
Hyung-Ock Kim, Jun Seomun, Jaehan Jeon, Chungki Oh, Wook Kim, Kyung-Tae Do,
Jung Yun Choi, Hyo-Sig Won, Kee Sup Kim
Samsung Electronics, Korea
 Temperature, One of Design Keys in Mobile SoC
 Temperature-limited operation is inevitable to prevent human skin burn in mobile devices
 Owing to performance trends and small form factor, temperature is a crucial design criteria in
mobile SoC design
 Thermal Management
 To keep a silicon below temperature limit at the cost of performance sacrificing
 Conventional thermal management achieved by voltage / frequency drop [1-3], i.e. thermal
throttling in Figure 1, which must accompany performance drop
 Besides, it is required to prevent power source shutdown whenever thermal runaway happens
 Since leakage power has strong feedback to temperature, it is an important momentum of
temperature increase in mobile SoC , which is given by [4]
Voltage,
freq. drop
Freq.
Vdd
Thermal throttling
operation
 K leak 2V th
Pleak  K leak 1V dd e
Temp.
Thermal upper limit
Thermal lower limit
Time
Figure 1. Thermal throttling operation.
1
Tj
(1)
 Body Bias Control [5]
Ioff Decrease by 0.4V RBB @ SS
 Figure 2 shows leakage current reduction by
use of reverse body bias (RBB) in nanometer-scale
technologies
 RBB can be utilized to relieve thermal throttling
by weakening leakage-temperature feedback
-65.0%
-60.0%
-55.0%
-50.0%
-45.0%
65nm
45nm
32nm
28nm
Figure 2. Leakage current reduction by RBB of 0.4V.
 Advanced Thermal Design and Management by Body Bias Use
 We propose body bias design and optimization scheme spanning from system-level
design to post silicon tuning to enhance thermal management
 In design stage, thermal-leakage feedback and body bias design cost are formulated so
as to decide body bias use, which is followed by body bias implementation
 In post silicon, body bias use is explored and optimized both to optimize peak
performance and to save total power
 The proposed scheme has been implemented in 32nm HKMG commercial mobile SoC,
Exynos 4 Quad, and it results in 12.3% performance improvement in high speed mode
and 19.1% total power saving
2
 Overall Design and Optimization Flow
Early-stage
design
decision
Body bias
design
Front-end
design
Body bias
implement.
Back-end
design
Post-silicon
optimization
Silicon
Silicon test
Board,
SW stacks
Figure 3. Design flow of advanced thermal management.
 In early-stage design decision, cost and gain of body bias are evaluated in a CPU core and other
digital blocks (named as SoC)
 Once body bias use is determined, body bias circuits including body bias generators (BBG) and
power management unit are integrated into design
 In back-end, implementation and validation of body bias network are exercised
 Post-silicon optimization is a body bias tuning to minimize total power and maximize peak
performance respect to process variation and temperature
3
 Body Bias Design Flow
CPU
 Early-stage design decision in a CPU core
SoC
- Body bias comes at the cost of area: body bias generator, body bias network,
Chip
and power management unit
- The cost of body bias can be estimated at floorplan stage, and then decided whether body bias
is accepted or not
- If thermal runaway by leakage is expected to appear, we must adopt body bias not to loose
performance by leakage current
Body bias area estimation
Thermal runaway estimation [4]
A abb  n  Abbg  Actrl  route _ area  0 . 01 ,
T j  T a  Ptot   ja ,
(2)
 block _ area ( route _ cong   )
route _ area  
0
( route _ cong   )

(3)
 K leak 2V th
Ptot  Pleak  Pdyn  K leak 1V dd e
n comes from body current calculation
(will be covered later)
 comes from design experiences
Tj
 K dyn V dd f
If equation (3) is not converged, thermal
runaway is expected
4
2
 Body Bias Design Flow
 Body current calculation to determine # BBG
- Body current (GIDL and junction leakage) calculator has been developed to calculate body
current, so that we can find proper number of BBG to drive a block
- A proposed calculator utilizes a set of look-up tables which are pre-defined for logic and
memory bit cell by using SPICE simulation
- Figure 4 presents calculation flow of body current and it is compared to silicon measurement in
Figure 5, which shows proposed calculator over estimates body current up to 20%
- # BBG is expressed by max body current / BBG driving limit
Gate counts,
# bit cells
Operating conditions
25.0%
Process corner, Vdd,
temperature, body bias
20.0%
Searching body currents in
operating conditions
Error
Design information
Body current
look-up tables
15.0%
1.0V Vdd
10.0%
1.1V Vdd
5.0%
Calculating total body currents
with design information
0.0%
0.2V
Body currents
0.3V
RBB
Figure 5. Comparison of calculated and measured body
current in SoC silicon.
Figure 4. Body current calculator.
5
 Body-Bias-aware Thermal Control
 Figure 6 shows overall thermal management scheme utilizing body bias
 Thermal management unit (TMU) periodically reads out temperatures from on-chip sensors
 Once a temperature exceeds thermal upper limit (recall Figure 1), interrupt controller asserts
thermal throttling to CPU, and then CPU controls Vdd, frequency, body bias through Vdd / freq /
ABB manager
 Vdd / freq / ABB values for thermal management is defined in post silicon optimization, which
maximize thermal relaxation efficacy and to minimize performance loss
BBG
Temp.
sensors
TMU
Interrupt
controller
CPU
PLL
Vdd / Freq /
ABB
manager
Regulator
Figure 6. Thermal management scheme using Vdd, frequency, and body bias control.
6
 Search for Optimal Thermal Management Point
 Leakage portion is changed by process variation,
DVFS control, and even ambient temperature
 Figure 7 presents inverter path delay increase by RBB,
where 40~50mV Vdd compensation is required for 0.4V RBB
 This can increase engineering cost and optimization
TAT in post silicon
Speed Slow Down
 Thermal management optimization can be achieved by empirical practice because it can exactly
capture temperature changes by real user scenarios and real mobile sets (e.g. smartphone, tablet)
 Using RBB expands search space because RBB efficacy is dependent to leakage portion and Vdd
compensation for slow down increases dynamic power
14.0%
12.0%
10.0%
8.0%
6.0%
4.0%
2.0%
0.0%
1.0VDD
1.2VDD
0.1V
0.2V
0.3V
0.4V
RBB
Figure 7. Inverter path delay increase by RBB.
 Therefore, we will use “simplified” RBB policies:
- Use of RBB is decided for each silicon group (binning group) respect to process variation
- But we will “merge” RBB applying condition for each silicon group if they show similar
characteristics
7
 Exynos 4 Quad as Test Vehicle
 Figure 8 presents block diagram of Exynos 4 Quad, where CPU is Cortex A9-based quad cores
running up to 1.4GHz
- To enhance computation performance, GPU, multimedia processors and interface units are integrated along
with 6.4GB/s dual-channel DRAM interface for wide memory bandwidth
 Body bias is used in quad core CPU to optimize thermal throttling
 Thermal throttling evaluation board is shown in Figure 10
 Power can be measured by external multimeter, and thermal throttling and on-chip temperature
are transferred to PC via RS-232-C connection on the fly
Body bias domain #2 Body bias domain #1
Video
CPU
core
RS-232-C for mode control in PC
Exynos 4 Quad
Audio
File
Image
L2 Cache
DRAM
Controller
Firmware
Pins for power
measurement
6.4GB/s
dual channel
Camera
Display
GPU
Figure 9. Thermal throttling evaluation board.
Figure 8. Exynos 4 Quad block diagram.
8
 Thermal Optimization Practice
 Process variation and body bias control
1.6
1.6
1.5
1.5
1.4
1.3
1.2
Bypass
Breakeven point
1.1
RBB
1
0.9
Total Power [normalized]
Total Power [normalized]
- The use of RBB can be manipulated respect to process variation
- Figure 10 shows total power measurement in various temperatures in high performance mode
- Fast silicon shows steeper total power increase over temperature owing to leakage current
- As is clear, in lower temperatures, RBB use results in power increase owing to voltage
compensation of 50mV; breakeven points are 65°C and 75°C in fast and slow silicon, respectively
- Because typical temperature in high performance mode is over 75°C, RBB can be activated
regardless of process variation
1.4
1.3
1.2
Bypass
1.1
Breakeven point
1
0.9
0.8
0.8
20
40
60
80
100
20
120
40
60
80
100
Junction Temperature [C]
Junction Temperature [C]
(a)
(b)
Figure 10. Total power saving by RBB use in (a) fast silicon and (b) slow silicon
running at max speed.
9
120
RBB
Mode
transition High performance mode
 Thermal throttling improvement and total power saving
Thermal
throttling
Freq.
 Thermal Optimization Practice
- Table 1 shows thermal throttling improvement measurement
Time-before
throttling start
by RBB use in real application setup (running OS)
- Time-before throttling start is improved by up to 171.0%, which means if an application requires
only short time of high performance mode, it may not experience performance loss by throttling
- In real application, normal status is improved by up to 12.3% by using RBB
- Figure 11 shows total power saving measurement by RBB use in 1.0GHz operation and total
power saving is up to 19.1%
- It is clear that RBB efficacy is getting better in high temperature and fast silicon
Table 1. Performance improvement by RBB
in real application setup (running at max speed)
Slow silicon
Fast silicon
Time-before throttling start
improvement [%]
82.0
171.0
Normal status improvement [%]
7.0
12.3
Total power saving [%]
Slow chip
Fast chip
25
20
15
10
5
0
25
75
85
Chip temperature [ºC]
105
Figure 11. Total power reduction by RBB use
in running at 1.0GHz.
10
 Thermal throttling has been used so as to obey thermal limit to prevent
human skin burn while maximizing user experience in high performance
mobile SoC
 We have proposed a new thermal throttling method based-on RBB, which
spanning from system-level design to post-silicon optimization
 In system-level design, cost and efficacy of RBB use are formulated for high
performance CPU
 Body current calculator has been developed for robust design of RBB and thermal
management scheme using RBB is presented
 In post silicon optimization, we have proposed empirical policy to reduce engineering
cost and maximize RBB efficacy to reduce thermal throttling
 Proposed design and optimization have been applied to commercial mobile
SoC, Exynos 4 Quad in 32nm HKMG
 Proposed method improves peak performance by up to 12.3% in fast silicon and it
can save total power up to 19.1% in 1.0GHz operation
 171% improvement of time-before throttling start means proposed methodology can
decrease thermal throttling chance when an application requires short period of high
performance
11
[1] D. Brooks and M. Martonosi, “Dynamic thermal management for high-performance microprocessors,” in Proc. ISCA,
pp. 171‒182, 2001.
[2] K. Skadron and et al, “Temperature-aware micro-architecture: modeling and implementation,” ACM Transaction on
Architecture and Code Optimization, Vol. 1, No. 1, pp. 94‒125, 2004.
[3] A. Naveh and et al, “Power and thermal management in the Intel Core Duo processor,” Intel Technology Journal, Vol.
10, No. 2, pp. 109‒122, 2006.
[4] J. H. Choi, A. Bansal, M. Meterelliyoz, J. Murthy, and K. Roy, “Self-consistent approach to leakage power and
temperature estimation to predict thermal runaway in FinFET circuits,” IEEE Transaction on Computer-AidedDesign, Vol. 26, No. 11, pp. 2059‒2068, 2007.
[5] J. W. Tschanz and et al, “Adaptive body bias for reducing impacts of die-to-die and within-die parameter variations
on microprocessor frequency and leakage,” IEEE Journal of Solid-State Circuits, Vol. 37, No. 11, pp. 1396‒1402,
2002.
[6] D. Markovic, C. C. Wang, L. P. Alarcon, T.-T. Liu, and J. M. Rabaey, “Ultralow-power design in near-threshold region,”
Proc. IEEE, Vol. 98, Issue 2, pp. 237-252, 2010.
[7] Y. Wang and et al, “A 4.0 GHz 291Mb voltage-scalable SRAM design in 32nm high-κ metal-gate CMOS with integrated
power management,” in Proc. ISSCC, pp. 456‒457, 2009.
[8] C.-H. Jan and et al, “A 32nm SoC platform technology with 2nd generation high-k/metal gate transistors optimized for
ultra low power, high performance, and high density product applications,” in Proc. IEDM, pp. 1‒4, 2009.
[9] S. Borkar, T. Karnik, S. Narenda, A. Keshavarzi, and V. De, “Parameter variations and impact on circuits and
microarchitecture,” in Proc. Design Automation Conference, June 2003, pp 338‒342.
12

similar documents