アプリケーションノート 4304

Interfacing the MAX5881 Direct RF Synthesis DAC to FPGAs

筆者: Brian Dellacroce

要約: This application note discusses techniques for interfacing the MAX5881, a 4.3Gsps cable downstream direct RF synthesis DAC, to field-programmable gate arrays (FPGAs). The focus is on the timing of the MAX5881's high-speed, digital-input data interface to a Xilinx® Virtex™-5 FPGA. However, the techniques outlined here are also applicable to a wide variety of FPGAs and custom ASICs.


この文書は極めて特定的なアプリケーションに向けて書かれたものであり、独自技術に関する情報を含んでいるため、その開示は該当アプリケーションに直接関わっているお客様のみに限定させていただいております。この文書を請求される方は以下のフォームに必要事項を記入してください。適当と判断された場合は2営業日以内にご連絡いたします。

以下の情報はすべて英語で記入してください。

Introduction

Due to the wide bandwidth of the MAX5881 4.3Gsps cable downstream direct RF synthesis DAC, its data interface requires operation at higher frequencies than lower bandwidth DACs. Designing higher bandwidth data interfaces generally requires more attention to ensure robust, error-free operation.

Interface Architecture

Due to the many possible combinations of FPGA features, several interface architectures are possible. One such architecture is shown in Figure 1. This architecture results in a wide operating frequency range and a high tolerance to process, voltage, and thermal (PVT) variations. This application note focuses on the features of this particular architecture. The advantages of this configuration over some other possible configurations are highlighted, along with a detailed timing analysis of the data interface.
Figure 1. MAX5881 to Xilinx Virtex-5 FPGA interface (CLKDIV = 0, DDR data interface configuration).
Figure 1. MAX5881 to Xilinx Virtex-5 FPGA interface (CLKDIV = 0, DDR data interface configuration).

Clocking

The MAX5881 has separate analog and digital clocks. The clock for the digital data interface is output by the MAX5881. Because the source of the data, the FPGA, is not the source of the data clock, a "system synchronous" style of data interface¹ is used. It should be noted that the MAX5881's data clock outputs, DATACLKP/DATACLKN, have very low jitter, as this jitter is derived directly from a normally high-quality analog clock at the CLKP/CLKN inputs.
Using an FPGA DCM² with external clock feedback automatically compensates for the FPGA's inherent output data timing shifts with PVT. The external-clock-feedback path is designed to closely match the delay of the FPGA's internal data path and the MAX5881's DATACLKP/DATACLKN outputs. This ensures that the timing compensation provided by the digital clock manager (DCM) matches the PVT drift of the FPGA's clock and data paths. The length of the clock-feedback signal's PCB traces are designed to be equal to the length of the data traces plus the length of the DATACLK traces; this makes the FPGA's data outputs change coincident with the DATACLK edges. Adding additional delay to the clock-feedback path advances the data in relation to DATACLK. This delay is designed or adjusted to ensure that the data is stable within the MAX5881's data setup to hold time window. This additional delay can be provided by either increasing the clock-feedback signal's trace length or by adding an FPGA ODELAY component to the feedback path.

Data Rate

The MAX5881's CLKDIV pin is set low and the FPGA is configured to output double data rate (DDR) formatted data. In this case, the DATACLKP/DATACLKN signal's frequency is one-fourth that of the MAX5881's input clock on the pin pair, CLKP/CLKN. For the MAX5881's update rate, the CLKP/CLKN frequency is 2.0GHz. This results in a DATACLKP/DATACLKN frequency of 500MHz. This 500MHz clock signal drives the FPGA's serializer (OSERDES) components, which are configured to output data on both the rising and the falling edges of clock. Thus the data period of the FPGA's output data is 1ns (or 1.0GHz). The CLKDIV = 0 configuration avoids creating a frequency-doubled clock to drive the serializers, as Xilinx does not allow a DCM (or PLL, for that matter) to be configured for external clock feedback and a multiplied output clock. The OSERDES components at the FPGA final data-output stage reduce the frequency of the clock on the upstream, or data-input, side of the OSERDES. This reduces the FPGA's placement-and-routing effort to a more practical level. For the MAX5881's update rate, the frequency of this upstream clock is 250MHz.

Interface Timing

A timing analysis of this data interface is provided in spreadsheet format in Table 1. The analysis is presented in two major sections. In the first section, the interface is analyzed for the operating frequency and is named "Data Period Analysis." This section does not consider the exact position of stable data within the data period. However, it serves to establish that it is (or is not) possible to tune the clock-feedback delay to ensure that data is stable within the window bounded by the MAX5881's data setup and data hold times. One operating frequency is analyzed in the spreadsheet, then the analysis is repeated (though not shown) for several other operating frequencies; the results are presented graphically in Figure 2. Also, a timing reference waveform is provided in Figure 3.
Table 1. Data Period Analysis of the Virtex-5 to MAX5881 Interface
Value Unit Specification Description Source Notes
4.300 GHz fDAC MAX5881 output update rate MAX5881 data sheet  
0.5375 GHz fDATACLK = fDAC/8 DDR interface mode MAX5881 data sheet  
930 ps tDATAPERIOD =1/(2 × fDATACLK) Data period MAX5881 data sheet  
50 ps tINFBOFFSET FPGA DCM feedback phase error (peak ±) Xilinx ds202.pdf  
120 ps tPERJITT FPGA DCM jitter (peak ±) Xilinx ds202.pdf  
121 ps dtOUTWC Worst PVT FPGA output-to-output skew (peak-to-peak) ".twr" post-layout static timing report from Xilinx static timing analyzer (worst-case PVT)  
1.4 ps dtPCB Board output-to-output skew Estimate All PCB data and data clock trace lengths matched
462 ps dtTOTAL = (2 × tINFBOFFSET) + (2 × tPERJITT) + dtOUTWC + dtPCB Total FPGA data output timing variance    
1100 ps tSETUP MAX5881 setup time MAX5881 data sheet  
-760 ps tHOLD MAX5881 hold time MAX5881 data sheet  
590 ps tDW = tDATAPERIOD - (tSETUP + tHOLD) MAX5881 data change window   Time in clock period when data does not have to be valid and stable
128 ps tPSL = tDW - dtTOTAL Period timing slack   Extra time between FPGA changing data; data must be stable at DAC
Figure 2. Period margin.
Figure 2. Period margin.
Figure 3. Timing reference waveform.
Figure 3. Timing reference waveform.
The second section of the timing analysis calculates the required clock-feedback delay (Table 2). It also calculates the setup and hold time margins, otherwise known as timing slack, with the calculated-feedback delay.
Table 2. Data Timing Analysis (MAX5881 DELAY Pin = 0)
Value Unit Specification Description Source Notes
0 ps tDOUTNOM FPGA clock-to-data output delay FPGA external clock-feedback configuration Feedback-path length equal to data-path length plus clock-path length; this centers data transitions at the clock edge
-170 ps tSUNOM = tDATAPERIOD - tSETUP - tDOUTNOM Nominal setup margin (without dtTOTAL factor)   Calculate margin without jitter, skew, and phase-error effects first; they are accounted for later
760 ps tHLDNOM = tDOUTNOM - tHOLD Nominal hold margin   Calculate margin without jitter, skew, and phase-error effects first; they are accounted for later
-401 ps tSUABS = tSUNOM - dtTOTAL/2 Absolute setup margin (including dtTOTAL factor)   Include jitter, skew, and phase-error effects
529 ps tHLDABS = tHLDNOM - dtTOTAL/2 Absolute hold margin (including dtTOTAL factor)   Include jitter, skew, and phase-error effects
-529 ps tDADVANCE = -tHLDABS Delay-adjusted FPGA clock-to-data output External clock feedback with feedback-path length equal to data-path length plus clock-path length plus tHLDABS Place data transition immediately after tHOLD
128 ps tSUFINAL = tSUABS - tDADVANCE Delay-adjusted setup margin    
0 ps tHLDFINAL = tHLDABS + tDADVANCE Delay-adjusted hold margin    
Similarly, this section of the analysis is also repeated for various operating frequencies. Due to the volume of data generated, the spreadsheet format data is not shown. Instead, the results are shown in the Figure 4 setup margin and Figure 5 hold margin graphs.
Figure 4. Setup margin.
Figure 4. Setup margin.
Figure 5. Hold margin.
Figure 5. Hold margin.
It should be noted that, in Figures 4 and 5, the clock-feedback delay was calculated for operation across a wide range of operating frequencies (varying fDAC). This was accomplished by placing the leading edge of the data-invalid window where the data starts changing to the next value, which is immediately after the MAX5881's hold time. In other words, the FPGA starts the transition to the next data value as soon as the hold-time specification is met.
For operation at a single specific frequency, the setup and hold time margins can be balanced by increasing the hold-time margin and decreasing the setup time margin. Again, this is accomplished by varying the delay inserted into the clock-feedback path. At low frequencies, the setup and hold time margins are generous, and margin balancing becomes moot.
Figure 6. Margin balancing.
Figure 6. Margin balancing.

Alternative Approaches

Due to the flexible nature of FPGA devices, a multitude of alternative designs can be conceived. It is beyond the scope of this application note to delve into the details of these approaches, so only brief descriptions of a few of them are presented here:
  1. Use the DCM phase offset to place an output data transition instead of increasing the length of the DCM clock-feedback path to add additional delay. This has advantages from a software or firmware delay-adjustment point of view, but it also introduces more timing uncertainty due to the FPGA phase-offset circuitry.
  2. Use the ODELAY element in the feedback path of the DCM. Again, this is advantageous due to flexibility of firmware delay adjustments, but also introduces more timing uncertainty.
  3. Use alternate method #1 or #2 combined with a device-by-device calibration. The MAX5881 characterization data shows a per-part, data-sampling window variation of 151ps vs. 340ps for all parts over temperature. A smaller per-part timing variation may compensate for the additional FPGA timing uncertainty. With this method, some of the part-to-part FPGA timing variation may also be reduced.

Conclusion

Through careful system architecture selection, PCB design, and utilization of specific design features, FPGAs can be successfully interfaced to the MAX5881 with positive data timing margins.

References

  1. System Interface Timing Parameters (PDF).
  2. Virtex-5 FPGA Data Sheet (v4.7, PDF), September 23, 2008.

アプリケーションノート 4304,AN4304, AN 4304, APP4304, Appnote4304, Appnote 4304