Yet another SDRAM design :)

Started by Unknown March 2, 2005
Hello ALL,

I am having hard time designing the schematic to interface conventional
SDRAM chip to Virtex-4 FPGA. As a SDRAM we have selected Micron 256Mbit
part with 32 bit wide data bus. Both packages a BGA, Virtex-4 is FF-672
type, SDRAM is FBGA-90. Our chips is planned to be located about 5-10mm
(0.2-0.4 inches) from each other and the longest trace is suppose to be
about 2.0-2.5 inches. We are also planning to have data bus traces to
be as short as possible. First approximation gives about 1.3-1.5 inches
data bus traces. The trace impedance is planned to be about 53-55Ohms.
The interface is planned to run at 125MHz.

The memory we are going to use has conventional 3.3V interface. We are
planning to use DCI for all outputs in our design, i.e. control
signals, address bus, clock and clock enable signals. Doing simulation
we found that DCI does pretty good job for all these lines - there is
no overshot/undershot problems at all on all these lines. The only
problem is the data bus. Than signal originates from Virtex-4 FPGA
everything goes fine, because DCI still active for this direction,
everything becomes bad when signal originates from SDRAM. There is
significant overshot and undershot around every edge of the signal. Its
value goes as high as 4.2V for overshot and to -900mV for undershot.
This extreme condition lasts for 200-300 picoseconds according to our
simulation. For simulation we are using Virtex-4 and SDRAM Ibis models
from manufacturers.

The relatively simple solution is to terminate each line in data bus,
but it requires relatively lots of space on the PCB and can not be done
reliably :( As soon as Micron driver gets terminated properly the line
becomes less than suitable for Virtex-4 driver, something in between is
not very good for both of them.

The funniest thing that I saw several existing and reliably working
designs, using average trace length about 2.5 inches with similar
components working with close frequencies and DOES NOT using any series
termination on data lines with default FPGA drivers. So, what part are
we missing? Should we believe in simulation or it is not as accurate as

With best regards,
Vladimir S. Mirgorodsky