I'm trying to design a board with a single SFP+ 10 Gigabit Ethernet transceiver, but I found the power supply noise requirements are specified in an indirect and complex way. If anyone has experience designing 10 Gigabit Ethernet, your answers are greatly appreciated.
In SFF-8431 Specifications for Enhanced Small Form Factor Pluggable Module SFP+, page 17, it says
2.8.2 HOST POWER SUPPLY NOISE OUTPUT The host shall generate an effective weighted integrated spectrum RMS noise less than 25 mV in the frequency range 10 Hz to 10 MHz, according to the methods of D.17.1
2.8.3 MODULE POWER SUPPLY NOISE OUTPUT The module shall generate less than 15 mV RMS noise at point X of Figure 56 in the frequency range 10Hz to 10MHz, according to the methods of D.17.2
Here, the power supply noise requirements are pretty simple: less than 25 mVrms in a 10 MHz measurement bandwidth. It also says the modules themselves should not inject too much noise back into the host - since we're designing a host, it's not our problem.
Later, in page 96, the testing methodology is given.
D.17 POWER SUPPLY TESTING METHODOLOGY The reference power supply filter shown in Figure 56 is provided for module testing, including power supply tolerance testing. This filter will meet the noise filtering requirements in most host systems. Other filtering implementations or local regulation may be used to meet the power noise output requirements described in 2.8.2 and 2.8.3.
For each Vcc, the sum of the equivalent series resistances of the 4.7uH in- ductor, the 22uF capacitor and the damping resistor is 0.5 Ω. This resistance is desirable in actual host filters as well as in the reference filter; however, any voltage drop across a filter network on the host is counted against the host VccT and VccR accuracy specification in Table 8.
It says other filters are allowed, but for the propose of writing a specification, it's assumed that the host power supply uses the reference LC filter here - it makes sense.
Then things get complex. In page 97
D.17.1 HOST POWER SUPPLY NOISE OUTPUT The noise output of a Vcc supply of a host is defined with a resistive load that draws the maximum rated power (1 W or 1.5 W) connected between one Vcc contact and Vee, in place of the SFP+ module. When the noise on VccT is being measured, VccR is left open circuit, and vice versa. 8 Ω is used for a host capable of supporting Power Level II, and 12 Ω otherwise. The AC voltage spectrum is measured at the module side of the SFP+ connector. The noise power spectrum is divided by the truncated response of the reference filter and then integrated from 10 Hz to 10 MHz and converted to a voltage. This function is illustrated in the equation below and Figure 57. The specification limit is given in 2.8.2. The test is performed with all other portions of the host board/system active. Hosts with multiple SFP+ modules shall test ports one at a time, with active SFP+ in all the remaining ports.
\$ H(f) = a \times (log_{10}(f)) ^ 4 + b \times (log_{10}(f))^3 + c \times (log_{10}(f))^2 + d \times log_{10}(f) + e \$
The reference filter response H(f) shown in Figure 57 and the coefficients a, b, c, d, and e for the 5 frequency bands are defined in Table 32.
NOTES - As a lightly loaded power supply might generate more noise than a fully loaded supply, the host implementer may wish to assess the host power supply noise output at less than maximum current draw also. Because a small measured noise signal at high frequencies is multiplied up to give the inferred noise at virtual point X, care should be taken over the noise floor of the spectrum analyzer. Other measurement methods could be used, e.g. a measurement at a point inside the host, with appropriate consideration to any difference between the reference filter and a host’s actual filter.
Question
This arises several questions.
1. What is the intention of the indirect measurement methodology here?
If I'm reading it correctly, to determine the PSU noise, first, your power supply preferably uses the reference LC filter given. Then you take a measurement at the PSU output, at the SFP+ socket. Finally you divide the results by the LC filter's transfer function H(f) to "de-embed" that filter and reconstruct the power supply noise voltage before the reference filter.
Thus, you're measure the noise after the filter, then remove the filter, What is the motivation of the convoluted measurement methodology here?
A friend suggested that it's because all SFP+ modules are connected together, measuring it after the filter allows can to see the coupled noise from other modules. But it's not the case. Each SFP+ modules have two rails, each has its own LC filter, and all ports are also filtered separately (for example, this two-port FPGA reference design clearly shows 4 sets of filters, see page 22, 23).
My best guess is that at the point wbich the power comes out at the VRM, the noise may be acceptable. But later, when it passes through the LC filter, then routed all over the board, more noise can be unintentionally coupled into the rail only after the LC filter. The approach of measuring and de-embed allows one to see these additional noises, hence giving a more accurate measurements.
Can anyone confirm my guess?
2. What to do if I have a Different Filter?
The specification also says that different measurement methods or filter designs are allowed, but...
Other measurement methods could be used, e.g. a measurement at a point inside the host, with appropriate consideration to any difference between the reference filter and a host’s actual filter.
What exactly does it mean by "with appropriate consideration to any difference between the reference filter and a host’s actual filter"? For example, if I'm using a different filter topology for my power supply, do I have to characterize the frequency response of my filter and use my own transfer function in place of the reference filter, or what?
3. Tools, Code, and Scripts
Are there any existing tools, code or scripts for doing the de-embedding in this measurement? I need a reference implementation to ensure what I'm doing is correct.
