Published August 31, 2015

Why More-Than-Moore Power Management Is Required to Keep Up With Exponential Growth in ICT Data Consumption

Share

Share

Tweet

Share

Share

Email

Print

Significant gains in energy efficiency are required to keep up with the exponential growth in the data consumption of Information and Communications Technology (ICT) systems: end-user devices, networks, and data centers. Moore’s Law scaling (monolithic integration in silicon) is the historical technology driver, but it no longer achieves the required gains. Fortunately, a new power management technology has emerged. It achieves More-than-Moore scaling by integrating different components and materials to increase functional diversity and parallelism.^1,2,3 This technology can improve voltage regulator power density, response time and granularity by an order-of-magnitude to reduce the ICT system energy consumption by 30% or more. This paper explains why a Heterogeneously Integrated Power Stage (HIPS) enables power management scaling to keep up with the rising demands on data centers and network systems.

 

Exponential Growth in Data Consumption

The digital universe – the data we create and copy annually – will grow 10x, from 4.4 zettabytes to 44ZB, from 2013 to 2020.⁴ The forecast for 2020 compared to 2014 expects many more Internet users (3.9 billion versus 2.8 billion), more connected devices (24.4 billion versus 14.2 billion), faster average broadband speeds (42.5Mbps versus 20.3Mbps) and more video streaming (80% versus 67% of traffic).⁵ Most people will use a tablet or smartphone for all online activities by 2018.⁶ Mobile data traffic is increasing 10x from 3.3 exabytes per month in 2014 to 30.5EB/month in 2020.⁷ Many websites (e.g., Facebook) require an order of magnitude more power to build a web page than to deliver it.

ICT Energy Consumption

According to one study,⁸ ICT systems in 2012 consumed 920 terrawatt-hours of power, which is 4.7% of global electricity consumption.⁹ That power requires the equivalent of 300 coal plants¹⁰ and emits 2 trillion pounds of CO2-equivalent greenhouse-gas emissions.¹¹

A second study forecasts that improvements in energy efficiency will slow the growth in ICT electricity consumption from the historical 6.7% per year¹² to 3.8% per year¹¹ due to the following:

• Each new generation of ICT systems and components is more energy efficient. For example, improvements in optical components enable an order-of-magnitude increase in data rates of optical modules within roughly the same power envelope.
• Usage is shifting from energy-intensive TVs and conventional desktop PCs to energy-efficient smartphones, ultrathin notebooks, tablets, and fanless all-in-one desktop PCs. In 2011, the average annual electricity consumption was just 5.5 kilowatt-hours for smartphones and 16kWh for tablets, compared to 219kWh for PCs.¹¹
• Increasingly, networks and data centers will be optimized for energy efficiency rather than capacity. One example of an industry initiative is Greentouch, whose mission is to deliver the architecture, specifications, and roadmap to increase network energy efficiency by 1,000x compared to 2010 levels.¹³

However, a third study¹⁴ estimates that ICT systems’ percentage of global electricity consumption¹⁵ was 9.3% in 2012 and will grow in 2017 to 9.2%, 11.7%, or 15.6% according to three different growth scenarios, as Figure 1 shows. Data center operating costs (OPEX) are beginning to exceed capital expenditures (CAPEX) in the total cost of ownership.¹⁶ As applications and data move into the Cloud, energy consumption is shifting from client devices to networks and data-center infrastructure. This shift highlights the need to increase the energy efficiency of data centers and networks.¹⁷

Figure 1: ICT Electricity Consumption Forecasts

Moore’s Law Scaling

Moore’s Law (as revised by Gordon Moore in 1975) predicts that the number of transistors per chip can double every 24 months. Although the industry is now slipping behind this rapid pace, each new process-technology generation still provides significant gains. State-of-the-art processors used 22–28nm CMOS in 2013 and are advancing to 10–14nm in 2018 and 7–10nm in 2020. The increases in transistor density and processor shipments indicate that the number of processor transistors is scaling pretty well with data consumption. Processors continue to innovate, significantly increasing their compute density, data rates, and performance per watt. For example, 97% of mobile processors will use 64-bit cores in 2018, versus 15% in 2014. In addition, 93% will use 4–8 cores in 2018, versus 43% in 2014. A growing number of these processors support UltraHD (4K) video.¹⁸

However, Moore’s Law has hit some painful limits. Costs keep rising, with new fabs costing upwards of $10 billion. Designing a new high-performance processor using the latest technology costs more than $100 million.¹⁹ In the past, designers lowered the operating voltage to keep power consumption constant while doubling the transistor density (described as Dennard scaling^20,21). But this practice no longer works, because physical laws prevent designers from lowering the operating voltage much further below 1V.² A direct result of this is that the amount of “dark silicon” – the processor cores that must be powered down at any given time to meet the chip’s power budget – is rising with transistor density.^23,24

Energy efficiency is the new fundamental limiter of processor performance.^25,26 Increasing it requires large-scale parallelism with discrete dynamic voltage and frequency scaling (DVFS), near-threshold voltage operation, and proactive fine-grain power and energy management. DVFS^27,28,29,30 significantly reduces power consumption (up to 100x) by dynamically adjusting the operating voltage to its optimal level, which varies according to the software workload and operating temperature.³¹ It gets more effective by increasing the response time of the voltage regulator (VR) and the number of voltage levels.³² To keep up with greater transistor densities, VRs require significant improvements in the following:

• Response time – to quickly change the processor’s operating voltage, which typically varies from 1.0–1.5V for peak performance to 0.3–0.6V for low-power idling³³ from a 12V supply used in data-center and network systems. (Low-power idling is preferred over power gating to minimize latency issues.) Fast response is required to provide energy proportionality — power consumption that scales with workloads.³⁴ Fast response also realizes the benefits of the new adaptive voltage scaling (AVS) standard³⁵ and software-defined power architecture.³⁶
• Power density – to decrease size, because supplying higher current requires additional VRs. The processor’s operating voltage is no longer scaling with transistor density, so the VR-supplied current is increasing. Today, each system board needs 10 to 100 or more VRs to supply 100A to 1,000A or more (e.g., up to 20,000A for a high-end server card). Each rack in a data center or network system has many system boards (Figure 2).
• Granularity – to support the massive parallelism of many small energy-efficient elements (e.g., many heterogeneous processor cores, micro-server cards, small-cell base stations, etc). Benefits include more integration and specialization, reduced thermal loads, and better dynamic resource allocation, thereby providing high efficiency at low to medium loads. Many ICT systems will increasingly spend a large percentage of their time operating in low-power standby modes.³⁷

Figure 2: Many Voltage Regulators per System

Moore’s Law scaling does not help the incumbent VR technologies, because they use many discrete components, such as controllers, MOSFETs, inductors, and capacitors. Due to MOSFETs’ high switching-power losses, today’s 12V input, 10–30A VRs operate at a low (1.0MHz or less) switching frequency. This requires bulky inductors and capacitors, and they consume a large percentage of board space (~40%). They also suffer from a slow response time. MOSFETs are an aging 35-year-old technology and are hitting a performance asymptote. Moore’s Law scaling does not help MOSFETs, either, because they must withstand high voltage on the drain (e.g., 18–20V for 12V input). There are a finite number of well-known techniques for improving MOSFET performance, most of which have already been exercised.

More-than-Moore Scaling

More-than-Moore scaling – the heterogeneous integration of different components, materials, and parallelism – boosts VR response time, power density, and granularity. An example is a Heterogeneously Integrated Power Stage (HIPS), shown in Figure 3, which uses the optimum technology for each function:

• Gallium arsenide (GaAs) for the field-effect transistors (FETs).
• CMOS for drivers, protection, and control, handling the GaAs FETs’ unique requirements.
• 3D packaging using embedded die-in-substrate technology³⁸ to integrate in a 5mm x 5mm x 1mm QFN package the GaAs die, CMOS driver die, and passive components required to minimize parasitics for efficient high switching frequency operation.

 

Figure 3: Cross-section of a 5mm x 5mm x 1mm HIPS Module

A HIPS module is an evolutionary leap over the Driver-MOSFET (DrMOS) integrated power stage module.³⁹ It replaces the MOSFET dies with a GaAs die, reducing packaging parasitics and integrating performance-critical components in a very small package. GaAs FETs have much lower switching-power loss than MOSFETs due to their superior intrinsic material properties: 6x higher electron mobility (8,500 versus 1,400cm²/Vs), 5x lower on-resistance * gate charge, and no body diode (which eliminates reverse recovery loss). GaAs FETs also are more reliable, because they have no gate oxide, higher activation energy, higher bandgap, and the primary failure mechanism (sinking gates) is self-limiting.⁴⁰

A HIPS module can increase the VR switching frequency by 10x or more, which reduces the size of the output capacitors and inductors while reducing the transient response time. HIPS can readily be used in industry-standard synchronous buck converters with complementary components (PWM controllers and inductors) from third-party suppliers. No new architectures, materials, or components are required.

HIPS also increases the power density and granularity (providing more than one output) for board-mounted VRs. HIPS uniquely enables Package-Integrated Voltage Regulators (PIVRs) that integrate many fast, small VRs in the processor package as close as possible to the processor die without increasing its cost or heat dissipation.

HIPS modules enable many fast, small VRs that can reduce the energy consumption of ICT systems by 30% or more through fine-grain power management of multi-core processors,^41,42 energy proportionality and reduced cooling requirements.

Since VR-supplied current scales roughly with processor performance and price, let’s look at the processor market to estimate the HIPS market size. The market for processors, application-specific standard products (ASSPs), and FPGAs in data centers and network systems, the initial target markets for HIPS, is forecast to grow to $21.9 billion in 2018:⁴³

• $10.3 billion for server processors, most of which are used in data centers (Figure 4). Captive processors, such as IBM’s Power and Oracle’s SPARC, are excluded.
• $11.6 billion for processors, ASSPs, and FPGAs used in voice and data networks (Figure 5). Communications semiconductors include components for Ethernet, broadband infrastructure, customer premise equipment, home/access networking, network processors (NPUs), transport (Sonet/SDH, OTN), PCI Express, RapidIO, and network search engines.

Figure 4: $10.3 Billion Market for Processors Used in Data Centers in 2018 (Source: The Linley Group)

Figure 5: $11.6 Billion Market for Processors, ASSPs, and FPGAs Used in Voice and Data Networks in 2018 (Source: The Linley Group)

Assuming it’s approximately 5% of the processor content, the HIPS’ content for these systems is approximately $1.1 billion. HIPS modules also improve energy efficiency and, hence, the performance of ultrathin, fanless all-in-one desktop PCs, notebooks, tablets, and smartphones within a very challenging thermal envelope — a small, thin case that must remain cool enough to touch.

The GaAs industry today makes products for RF and microwave communications. In 2013, the industry produced 29.3 million square inches of GaAs,⁴⁴ which is equivalent to 100,000 GaAs 150mm wafers per month. GaAs industry revenue exceeded $6 billion revenue in 2014.⁴⁵ Hence, the GaAs industry has ample well-proven, high-volume manufacturing capacity for producing HIPS modules.

Summary

ICT energy consumption is significant today and growing, driven by the exponential increase in data consumption. Moore’s Law, the historical technology driver for improving energy efficiency, has hit the power wall. More-than-Moore scaling is the way forward. Heterogeneously Integrated Power Stages, which replace MOSFETs with GaAs FETs, enable power-management scaling to keep up with increasing processor transistor density by significantly improving VR response time, power density, and granularity.

References: 

   ITRS 2.0: Heterogeneous Integration
   Moore’s law: Repeal or Renewal?, McKinsey, December 2013
   More-than-Moore White Paper
   The Digital Universe of Opportunities, IDC, April 2014.
   Cisco’s Visual Networking Index, May 2015
   Gartner Says By 2018, More Than 50 Percent of Users Will Use a Tablet or Smartphone First for All Online Activities
   Ericsson Mobility Report, June 2015
   The Impact of Information Technology on Energy Consumption and Carbon Emissions, June 2015.
This study excludes entertainment systems (TVs) and energy consumed in manufacturing ICT systems.
   In 2012, 19,443 TWh of electricity was consumed per Enerdata Global Energy Statistical Yearbook 2015, and 22,668 TWh of electricity was produced (15% of global energy supply) per IEA 2014 Key World Energy Statistics.
Defining a Standard Metric for Electricity Savings
GeSI SMARTer 2020: The Role of ICT in Driving a Sustainable Future
The 2007-2012 average per Trends in Worldwide ICT Electricity Consumption From 2007 to 2012
Greentouch website
In 2012, ICT systems consumed 1,815 TWh when including entertainment systems and energy consumed in manufacturing ICT systems per Emerging Trends in Electricity Consumption for Consumer ICT, July 2013.
Assuming 2.2% growth in global electricity consumption (the 2012-2013 actual rate per Enerdata)
The Cloud Begins with Coal: Big Data, Big Networks, Big Infrastructure and Big Power, August, 2013
Power Consumption and Energy Efficiency in the Internet
A Guide to Mobile Processors, The Linley Group
Moore’s Law Shows Its Age, Wall Street Journal, April 17, 2015.
The Death of CPU Scaling: From One Core to Many — And Why We’re Still Stuck
Moore’s Law is Dead – (Part 1) What?
Multicore Processors: Challenges, Opportunities, Emerging Trends
The BubbleWrap Many-Core: Popping Cores for Sequential Acceleration
Performance Enhancement under Power Constraints using Heterogeneous CMOS-TFET Multicores
The Future of Microprocessors
The Future of Computing Performance – Game Over or Next Level?
Enabling Improved Power Management in Multicore Processors through Clustered DVFS
Phase-Based Application-Driven Hierarchical Power Management on the Single-chip Cloud Computer
Predicting Performance Impact of DVFS for Realistic Memory Systems
A Hybrid Local-Global Approach for Multi-Core Thermal Management
Technologies for Ultradynamic Voltage Scaling Circuits
Power vs. Performance Management of the CPU
Intel’s Near Threshold Voltage Computing and Applications
Google’s Energy Proportional Power Management for Data Center Applications, DesignCon 2015
Open Standard AVS for ASICs, CPUs, and FPGAs, March 2015
Meet Increasingly High Current and Power Demands in Networking
More Data, Less Energy: Making Network Standby More Efficient in Billions of Connected Devices
Current Development in 3D Packaging With Focus on Embedded Substrate Technologies
Integrated Power Stage Modules Support High Current Processors
Qorvo’s Reliability FAQs
Fine-Grain Power Management in Multicore SoCs using Integrated Voltage Regulators
Low-Cost Per-Core Voltage Domain Support for Power-Constrained High-Performance Processors
The Linley Group
Markets for Semi-Insulating GaAs Epitaxial Substrates: 2013 – 2018
Where Does the RF Compound Semiconductor Market Go from Here??