Reducing SDRAM Radiation on a Memory Bus

A memory bus is one of the harder interfaces to keep quiet. It is wide, it is fast, it is synchronous, and it runs across a large area of the board with many traces in parallel. The clock and the strobe are periodic, so their harmonics land on specific frequencies, and the data lines carry broadband energy from every transition. The result is a set of emissions that are easy to measure and, fortunately, fairly predictable in origin, which means most of them can be reduced by changes in routing and termination rather than by shielding.

Why the Memory Bus Radiates

Two mechanisms dominate. The first is the loop formed by each signal and its return path; where the reference plane is continuous and close to the trace, that loop is small and radiates little. The second is the loop formed between a signal and its neighbour when the pair is poorly referenced, which is why a wide bus routed far from a plane behaves much worse than the same bus routed close to one.

Edge rate sets the frequency content. A double data rate interface switches on both edges of the clock with rise times measured in hundreds of picoseconds, so its spectrum extends well into the gigahertz range. Reducing the edge rate where the timing budget allows, through driver strength settings or series resistors, lowers the high frequency content without changing the data rate.

Clock and Strobe Routing

The clock and the strobe are the most important nets on the interface, both electrically and for emissions. They should be routed with a solid reference directly beneath them, kept short, and terminated so that reflections do not add to the emitted spectrum. Where the bus uses differential clocks, the pair should be routed with constant spacing and matched length so that the common mode component stays small.

Keeping the clock away from the board edge and away from connectors prevents the field from coupling into whatever lies beyond the board. It is also worth keeping the clock layer away from any plane that carries a noisy switching current, because noise on the reference appears directly as common mode on the clock and radiates from whatever is connected to it.

Termination and Drive Strength

Termination serves two purposes at once. It controls reflections that would otherwise distort the eye, and it limits the current that flows for each transition, which reduces the amplitude of the field produced. On a memory bus the usual arrangement is a terminated scheme at the controller end, sometimes combined with on die termination at the memory device, and the value has to be chosen to satisfy both the timing and the emission requirement.

Drive strength is the other lever. Many controllers allow the output impedance to be programmed, and using the weakest setting that still meets timing reduces the current step and therefore the radiated field. This is a setting that is often left at the default during development and never revisited, even though it is free to change and can make the difference on a marginal design.

Spread Spectrum and Other System Level Measures

Spread spectrum clocking modulates the clock frequency slightly over time so that the energy that would be concentrated at one harmonic is smeared across a band. The peak measured amplitude falls, often by several decibels, without any change to the layout, which is why it is nearly universal on consumer and computing products. It must be enabled and configured, and its deviation and modulation rate have to be compatible with the memory device timing.

Other system level measures are worth considering where layout alone is not enough. Reducing the number of simultaneously switching outputs on a single edge, lowering the supply voltage where the interface allows, and placing the memory devices closer to the controller all reduce the energy involved rather than merely redirecting it. Shielding is a last resort because it adds cost and weight without addressing the source.

Decoupling and the Power Delivery Network

The power delivery network is part of the memory bus, not separate from it. When a wide bus switches, the supply current changes quickly and the impedance of the network converts that change into voltage noise on the reference. That noise appears as common mode on every signal in the bus and radiates from all of them together, which is why a bus can pass a single net test and still fail emissions.

The remedies are a low inductance plane pair dedicated to the memory supply, decoupling placed against each device rather than concentrated at one point, and a plane that is not shared with a noisy switching regulator. Where the memory supply is generated on the board, the regulator should be placed away from the memory area and its switching loop kept small. Our notes on EMC and EMI control describe the wider measures.

Layout Measures That Reduce Emissions

The most effective layout change is to keep the bus close to its reference plane. Routing the data and address lines on a layer adjacent to a solid ground plane contains the field, reduces the loop area and lowers crosstalk at the same time. Where the bus must change layers, the return path should change with it through stitching vias placed close to the signal vias.

The second is to shorten the parallel runs between the clock and the data lines, and between the bus and any other fast signal that shares the area. A bus that runs alongside a high speed serial link for several centimetres will couple with it, and the resulting emissions are often attributed to the memory when the source is the neighbour. Our notes on routing high frequency data buses and microstrip and stripline routing cover the trace geometry.

FAQ

What causes the most radiation from a memory bus? The clock and strobe, because they are periodic and their harmonics concentrate at specific frequencies. The data lines contribute broadband energy, but the periodic signals usually dominate the measured peaks.

Does spread spectrum clocking solve the problem? It reduces the peak amplitude of each harmonic by spreading the energy across a band. It helps significantly on a marginal design but does not remove the underlying coupling, so layout measures are still needed.

How close should the memory bus be to its reference? The dielectric between the trace and the plane should be as thin as the stackup allows, and the plane must be continuous under the whole bus. That single measure does more for emissions than any other layout change.

Measuring and Confirming the Improvement

Before changing anything, establish the baseline. A near field probe swept across the memory area with a spectrum analyser shows which peaks belong to the bus, and correlating those peaks with the clock harmonics identifies the aggressors. Comparing the trace before and after a layout change is the only reliable way to know whether the change helped, because the field pattern is complex and intuition is often wrong.

SDRAM radiation is unusually amenable to this approach, because the bus is periodic and the sources are few. Turning the spread spectrum setting on or off, changing the drive strength, or moving a decoupling capacitor produces a measurable change at the probe, and the improvement usually confirms which mechanism dominated. Once the mechanism is known, the same fix can be applied to other interfaces on the board.

Applying the Same Method to Other Buses

Clock routing discipline transfers directly to any synchronous interface. Keep the clock and the strobe short and referenced, terminate them properly, keep them away from the board edge, and check the drive strength setting rather than leaving it at the default. Those four measures alone account for most of the improvement on a wide bus.

The rest is a matter of containment. Keep the bus close to its reference plane, keep the parallel runs short, and keep the supply network quiet. Together they reduce the energy radiated rather than merely moving it elsewhere, which is the only kind of fix that survives a compliance test at the first attempt.