How to use FPGA to solve the power consumption problem of MPU in handheld devices

The consumer handheld device market is developing in leaps and bounds. The processing power of portable products is increasing, and more and more applications are supported; the speed of product replacement is accelerating, and new products must meet the time-to-market requirements in order to obtain the greatest market opportunities; the shortening of product life cycles requires shortening the development cycle, while placing greater emphasis on reusability and reprogrammability. There is also an interesting trend in the emerging handheld device market, that is, the shipment volume of each device in a series is decreasing, but the customized functions between series devices are increasing, which effectively increases the total shipment volume of the product. In this way, the key challenge becomes how to develop a system that can be widely reused and customized at the same time.

To meet the above challenges, more and more designers are starting to use FPGAs for handheld product development. FPGAs are becoming increasingly powerful and feature-rich, while gate count, area, and frequency are also increasing. FPGA development and turnaround time is much shorter than custom ASICs, and the additional advantage of reprogrammability makes FPGAs an attractive solution in the field of handheld embedded systems. In designs based on ASIC or FPGA, designers must carefully consider certain performance criteria, and the challenges they face are mainly reflected in area, speed, and power consumption.

As with ASICs, vendors also need to address area and speed challenges in FPGA design. As the number of gates continues to increase, FPGAs require larger areas and sizes to accommodate more applications, and design tools need to adopt better algorithms to use area more efficiently. The evolving FPGA technology also brings a series of new challenges to designers, one of which is power utilization, which is an urgent problem to be solved for designing FPGA-based embedded systems for handheld or portable devices.

FPGAs in Embedded Systems

A typical embedded system consists of a processor, memory, standard interfaces including USB, SPI, I2C, and peripherals such as LCD display and audio output. The core of the device is still the processor and processor interfaces, which are connected to various peripherals through on-board connections. The performance of the system mainly depends on the performance of the processor, which usually has a very standard architecture and is not easy to customize.

Sometimes the processor may be busy processing information from a low-speed peripheral. Although the processor utilization rate may reach 100% in this case, it is not doing microprocessor-centric affairs, but working at a particularly low performance level. Regardless of its core frequency, the microprocessor must wait for data from the low-speed clock. This also leads to higher power consumption because the processor utilization rate is 100%. The result will shorten battery life and require a larger heat sink or fan for cooling, which will ultimately affect the reliability of the entire system.

Therefore, FPGAs begin to play an important role in this regard because they can offload many peripheral interaction tasks from the processor. As shown in Figure 1, an embedded distribution system for uncompressed audio and video data streams using a standard Gigabit TCP/IP network. It has a dedicated DSP processor that is connected to the Xilinx FPGA through a standard bus interface, and the FPGA is then connected to various low-speed peripherals.

Figure 1: FPGA architecture for audio/video distribution systems. [page]

As a starter development kit, this FPGA connects to 12-bit PCM audio input and 12-bit PCM audio output through the I2S interface; it also connects to the video encoder and decoder, and communicates with I2C slave devices and RS232 devices; there are few general-purpose I/Os connected to the FPGA. The standard bus connected to the processor operates at a high-speed 66MHz, while the audio peripherals operate at a low speed of 1.182MHz; the UART and I2C serial interfaces operate at 56.6kHz and 100kHz respectively. Since data transmission occurs in multiple clock domains, only the processor can configure the data flow.

In this case, the processor no longer interacts with the low-speed peripherals, but the FPGA reads data from the low-speed PCM ADC audio device and stores the data in the FPGA's internal cache. The processor can periodically read data from this cache, or when there is enough data in the cache, the FPGA sends an interrupt to the processor. In this way, the processor has more time to perform necessary processor-centric work and enters sleep mode when idle.

Power consumption issues

In battery-powered embedded systems, energy conservation is of paramount concern. Power consumption can be broken down into three broad categories: startup power, static power, and dynamic power. Designers have no control over startup power, and it plays a major role in determining power supply selection. Most maximum current values ​​refer to the values ​​reached during this phase. However, static power and dynamic power are two different areas, and with proper planning and following the right guidelines, embedded designers using FPGAs can make significant improvements in power optimization.

Static power consumption refers to the power consumption generated when current still flows through the components when the system is not working. It is generally caused by device bias current and leakage current. Static power consumption also depends on the operating voltage. Reducing the operating voltage can reduce static power consumption, but this strategy is not always in the hands of designers. What designers can do is to define a reasonable architecture in which the least resources need to be used, while using resource sharing as much as possible and using FPGA modules in the most efficient way.

Another technique to reduce static power consumption is to estimate power consumption early in the design cycle, change the topology or use different IP blocks. For example, the Xilinx xPower Estimator tool is very useful at this time, and it can know whether the design meets the power budget very early. Power consumption estimation in the early stage may not be completely accurate, but it is indeed helpful as a guiding tool.

Dynamic power consumption is caused by some behaviors of FPGA gates, such as signal switching. When two gates are temporarily turned on, current flows and capacitance is generated. The speed of signal switching determines the amount of power consumption. Another factor affecting dynamic power consumption is the inherent capacitance formed in the internal structure of the circuit. Dynamic power consumption is a function of clock frequency, the number of gates being switched, and the rate at which these gates are switched. Gate fanout and capacitive loading on the traces will increase dynamic power consumption, and the power consumption value is proportional to the product of capacitance, voltage and frequency squared.

Designers have the most control over this power consumption, and they can use many techniques to achieve the greatest improvement in dynamic power consumption. Reducing the signal switching frequency can reduce power consumption exponentially. As shown in Figure 1, the control logic for the UART, parity checks, or frame overrun errors all occur in the lower speed clock domain. Even if the number of gates is not reduced, power consumption will be reduced. Designers can also reduce dynamic power consumption by reducing the overall operating frequency (if feasible). For example, after completing the feasibility and performance analysis, the designer decided that the above design can operate not only at 133MHz, but also at 66MHz. The DSP supports both speeds, and reducing the voltage will also help reduce power consumption.

Another technique is to reduce the number of active gates in active mode. Sometimes a portion of logic is turned on and configured at power-up, but is not actually required to do anything. For example, the analog audio capture unit is in operation, but the device is not performing any digital SPDIF audio capture. In this case, the general digital SPDIF audio capture circuit will still perform data sampling, bi-phase decoding, etc., thus wasting power unnecessarily. If the entire digital SPDIF audio capture circuit is disabled so that no signal switching occurs in the circuit, the dynamic power consumption will be significantly reduced.

Designers can achieve this by disabling the clock to this part of the circuit. A simple way to do this is to AND the clock signal with the enable signal, as shown in Figure 2. If the enable signal is low, the output of the AND gate will remain low. If the enable signal is high, the AND gate will output the clock signal.

Figure 2: A simple clock gating mechanism.

There are other approaches that can be used. If possible and the topology supports it, the number of signal lines can be reduced by multiplexing the address and data lines. In our example, the output to the video encoder is 16 bits of data, which can be multiplexed into 8 bits and sent out on both edges of the clock (rising and falling). This can also save dynamic power. In addition, choosing a serial interface instead of a parallel interface can also reduce power consumption. Using LVTTL or LVCMOS I/O with lower capacitive loading is also helpful. [page]

Embedded Processors

Embedding the processor into the FPGA is another strategy that handheld device designers can adopt, which can bring many benefits. First, the above challenges brought by custom processors are reduced. Second, the interaction between peripherals and processors occurs inside the FPGA, which can reduce the number of I/Os. Since I/Os consume a lot of power, this move can also achieve a certain degree of energy saving. Xilinx's Virtex-5 version supports PowerPC 440 processors, hard processors, and MicroBlaze soft processors, all of which can be used by designers to create high-end or low-end application systems.

With the invention of 90nm and 65nm semiconductor technology, the size of gates is shrinking, which makes the static power consumption problem more and more prominent. This is a very challenging phenomenon in today's world where people are more and more sensitive to power consumption indicators. As the power consumption problem has gained the attention of many FPGA suppliers, many exciting new technologies have emerged in this field. Low-power design will determine how strong the integration capability of a system is, and the industry is also in urgent need of standardizing design technologies that focus on power consumption.