Bypassing Intel's Overclock Limit Reveals DDR2-667

BatonRogue writes "Slashdot posted a Tomshardware article talking about Intel's 10% overclock limit on their new chipsets not too long ago. The situation has just become even more interesting. AnandTech just posted a roundup of DDR2 memory that sheds some light as to why Intel may have implemented the lock. It seems that on the Abit board they tested, which supposedly bypasses the overclock limit, the first generation of DDR2-533 memory modules had no problems working at 667MHz. Could it be that Intel is keeping DDR2-667 support for yet another revision of their new chipsets even though the memory support is clearly here today?"

Comments

Spikes in the electrical current.

[rice_burners_suck]

My brother, who works on chipset R&D at Intel, says this "artificial" limit is there because of spikes. These are the same spikes that appear in electrical current that are the reason for debouncing all digital inputs to a computer. As it turns out, there are certain spikes that exist at a range above 533 mhz. The chipset will work at the higher speed, but these spikes in the electric current will occasionally cause a bit to go the wrong way, if the clock cycle occurs exactly when a spike is going through the line. For this reason, it is going to take Intel a little longer to clean the signals at the higher clock speed, and therefore they cannot offer 667 at this time.

Incidentally, my brother also says that Intel would be at a great advantage over its competitors if it could offer the 667 mhz, so it is clearly not hiding an existing feature to milk the market. The spikes are the reason.

Re:Spikes in the electrical current.

[spectecjr]

So how exactly do these magical 'spikes' get through a modern, self-regulated switching power supply?

The modern, self-regulated switching power supply creates those spikes in normal operation.

(One of my coworkers once worked on ECGs... they had to synchronize everything so that all data acquisition occurs directly out of phase with the power supply... even the fans in the case had to be timed with the spikes it would generate... and this was a low-noise, custom generated, several hundred thousand dollars worth of R&D supply).

Re:Turn the question around

[ottffssent]

That would be $500 extra revenue for Intel. How come they don't do it?

Because they're making and selling as many $1000 chips as the market will support. Now, they can either sell the $1000 chips for less money and sell more of them, or they can sell a slower chip for $500 and keep the $1000 top-end CPU.

As it happens, the math works out in Intel's favor selling a 3.4GHz chip for your eldest son and a 3.2GHz chip for your right arm and a 3.0Ghz chip for a pretty penny. In the begining this is great. Then for a while the 3.4GHz chips get too easy to make and so they sell some of them as 3.2GHz chips and maybe even some as 3.0GHz chips. But until they can make enough 3.6GHz chips to satisfy current demand for 3.4GHz chips, they can't introduce a new speed grade. When the 3.6GHz parts come out, demand increases for the now-cheaper 3.4Ghz parts, and the now-cheaper 3.2GHz parts and the now-low-end 3.0GHz parts. And the cycle repeats.

Intel and AMD don't sell CPUs to consumers. They don't respond to market pressures except at predetermined times, so when demand for high-end parts is less than supply, they sell the high-end parts for less money, and when demand exceeds supply there's a shortage. These are not spot prices as would be paid by a consumer, these are contract prices as would be paid by someone ordering 20,000 chips a month for the next 3 months: there is no other way to respond to surplus or shortage within the current system. Intel would much rather sell a potentially-high-end P4 as a mid-range P4 and make $100 than force the customer to source mid-range parts from AMD.

Re:Well...

[Jennifer E. Elaan]

It used to matter a lot more back when you could really overclock a processor. My old dual-celeron used to run decently at 550MHz, when it was originally specced at 366. I think ion migration finally started to set in, though, it doesn't take the higher speed properly anymore. Otherwise, it's still running (at 366) to this day.

I'm one of the original overclockers. I had a 486DX36, back when overclocking required replacing the clock module on the motherboard. I'm also an occasional chip designer.

The trick to overclocking is to know what your limit is. Until recently, thermal load was not the limiting factor. The real limiting factor was a condition called "metastability", where a digital transition fails to finish before being latched in the next register (usually due to violations of the setup and hold time restrictions of those registers). The smallest case of metastability can flip a bit.

A larger case can cascade through multiple stages, flipping lots of bits or even pushing the state tables into illegal states. This is where the first real danger lies: a processor that uses one-hot encoding to improve the speed of the controlling state machine can be pushed into illegal states that may cause several circuits to drive the internal busses at once. This leads to large current dissipation, and in some cases it can burn holes in the thin metal layers of the IC.

A less common hazard appears in cases when the CPU is massively overclocked. The CPU in such a case will never exit the metastable state. This causes each clocked circuit on large areas of the chip to dissipate maximum current during those metastable states. This can also lead to high current dissipation, although it is less dangerous than the abovementioned one. It's also worth mentioning that a chip in this state will not function normally.

Generally speaking, you have a wide margin between the onset of metastability and the onset of serious damage. Unfortunately, there is now *another* danger of overclocking processors. High thermal load can cause ion migration. In fact, most processors are now designed to only last 5 years (!) before ion migration renders them useless. (This is also why I personally don't overclock anymore).

It's frightening to notice that mainstream CPU's are less and less overclockable and have higher and higher thermal loads with smaller and smaller featuresizes, though. The manufacturers are simply not leaving as much margin as they once did.