Intel Core i7 overclocked to 5.6 GHz

The 8.58GHz achieved by a liquid nitrogen-cooled Bulldozer, but professional overclocker Hicookie did manage to set a new record for Core i7 3930K by reaching 5.6GHz -- with the help of the aforementioned Dippin' Dots refrigerant. Hicookie unlocked the  overclocker achievement with the help of an X79 motherboard from Gigabyte, which also gets some recognition for being the first such board to reach a multiplier of 57. The super-speedy i7 getting pushed to the brink of destruction, head on after the break.

 

 

 

"Renowned overclocker 'Hicookie' achieved a new high clock speed on the Intel Core i7 3930K processor by cranking the chip past 5.6GHz using a Gigabyte GA-X79-UD3 motherboard, the first mobo in the world to achieve a mulitplier of 57x. There was a bit of a scandal with Gigabyte recently when a YouTube video showed one of its X79 boards going up in smoke. Gigabyte released a BIOS update for several of its X79 boards to prevent such incidents from happening, and there were outcries that the new F7 BIOS would ... [reduce] overclocking performance; Hicookie's achievement should erase those concerns."

Understand why overclocking risks instability

The reason you can overclock without raising the voltage is that there is a voltage "guard band," which is like a safety margin. The ways of reducing that guard band, because it's wasted energy. But that guard band is there for a good reason. Typically the critical paths in the chip (those with the longest propagation delay, which limit the safe clock speed) are a bit faster than the clock period. But that's only true whe the voltage is stable. If the voltage droops, then the propagation delay of those paths will increase, possibly too much, and you get incorrect computation. Voltage droops occur when circuits suddenly start switching a lot, demanding more current, or in other words, the effective impedence of the circuit drops, and by V=IR, for the current being supplied by the voltage regulator at that instant, the voltage inside the chip will drop. The regulator cannot respond instantly, so a guard band is provided so that the maximum droop never brings the instantaneous voltage below a certain margin. If you overclock without raising voltage, then your CPU will work fine most of the time, but certain workloads will cause wide swings in current demand, and if you execute one of those, you may crash your system.

This is why memory tests are worthless for stability testing, because due to cache miss latency, the current demand is relatively low and stable. Prime number generators are also not so great, because their current demand is relatively high and stable. I know that some of the SPEC and PARSEC benchmarks have some wild behavior, like FFT, for instance, or anything that has a lot of barrier synchronization. For the regular user, what's likely going to happen is that you'll get random such events where variation in cache hits and vector computation phases will cause significant spikes in current, and your game will crash.