Supercooling FPGAs with Liquid Nitrogen
The Virtex-4 FPGAs we've been using have been having problems with overheating (basically going into thermal runaway until the power supplies shut down). The problem is that synthesis says that the chips can be clocked fairly fast with our cores (200MHz+), but when it comes to actually running the design, the chip becomes too hot and won't actually support the reported clock speed. This is supposedly due to the fact that smaller transistors (90nm now) end up leaking more energy, which increases the thermal problems with the Virtex-4 over the older chips. So, we decided we would try to supercool them to try and fix the problem and see if we could actually clock them as fast as reported.
The result of the experiment was being able to clock our lanman hash cracking designs up to 175MHz (before, we were able to get 125MHz with just a heatsink attached to the card). Clock speeds above that resulted in inconsistent data, which is probably due to the signals traveling much faster than they're anticipated to during synthesis and place & route. Xilinx has options in their tools for specifying operating temperature which might help us run our design consistently at the faster clock speed, but we haven't tried that yet. More details will be posted here if we experiment with this more in the future.
|
