The Temperature Rolloff with Respect to Overclockability

One of the cool realizations that we found on 4790k was there was an interesting temperature rolloff with respect to overclockability. It was fairly distinct, but not perfectly distinct. What we found was that as the junction temperature of the Silicon reached let's say 80 degrees, which is well below where we would normally throttle right, normally we run our parts up to about 100°C Junction temperature and at that point when it reaches 100°C, we start to throttle the performance back to protect the part. We started to see an overclocking rolloff at 80°C, so we would see parts that were cooler below 80 over clock better until they got to 80.

This discovery gave us the ability to find a smaller population of parts from which to do detailed analysis on. The focus there was to really just do speed path debug, which is something that's pretty unique and difficult to do. What you're trying to do is catch the part in a place where it's failing but hasn't failed so badly that it's lost its mind and you can't extract any interesting information out of it. You want to get it at that point where it's just going to kind of croak but not totally croak.

Once we get it there, we can dump out tons of information out of the processor. We can get gain an understanding of what's in the caches, what's not in the caches, what was the last instruction run, what was the state of that instruction, what was in my buffers? You know, am I stuck waiting on a load? Am I stuck waiting on a store? You can get all kinds of interesting State information from these things once you get them to fail in a way that you can capture it.

Trying to get the thing to fail but fail in a way that we could control the failure kind of like a controlled crash. So, we spent quite a bit of effort just getting to that point because we've engineered the platform in such a way that when when it thinks it's in trouble, it's going to try to save itself. And that's a feature but it's also a feature that kind of hurts us when we're trying to do a controlled crash because platform senses that it's in trouble and it pulls up in order to debug things.

We had to figure out oh, we got to tear out our own safety features so that we can crash this thing in such a way that we're not crashing it too hard but we're crashing it hard enough that we could debug it. And we had to get our own features out of the way that were really out there that you are hugely beneficial to overclockers because it allows them to once they've crashed it, get the system back up and running again.

Understanding Temperature Dependency

One of the key realizations we had was that temperature dependency plays a significant role in overclocking. By understanding how different temperatures affect our processor's performance, we can make informed decisions about when to push the limits of our system.

Working with Silicon Debug Teams

To gain a deeper understanding of what's going on inside our processors, we worked closely with silicon debug teams. This collaboration allowed us to understand what part of our silicon is limiting us and identify potential areas for improvement.

Unlocking Performance through Silicon Knobs

After working with the silicon debug team, we discovered that there are actually knobs within the silicon that can be turned to unlock more performance. These knobs can be adjusted to optimize different aspects of the processor's operation, leading to significant gains in speed and efficiency.

Significant Gains in Performance

Through our work on identifying and tuning these silicon knobs, we were able to achieve some remarkable gains in performance. We found a couple of pretty significant knobs that unlocked a good 200 MHz of performance. This demonstrates the potential for overclocking and highlights the importance of working closely with silicon debug teams to unlock maximum performance from your system.

"WEBVTTKind: captionsLanguage: enI I think one of the cool realizations that we found on 4790k was there was an interesting temperature rolloff with respect to overclockability it was fairly distinct I wouldn't say it was perfectly distinct but what we found was is that as the junction temperature of the Silicon reached let's say 80 degrees 80 degrees seat which is well below where we would normally throttle right normally we run our parts up to about 100° C Junction temperature and at that point when it reaches 100° C we start to throttle the performance back to protect the part we started to see an overclocking rolloff at 80° C so we would see parts that were cooler below 80 over clock better until they got to 80 that was a pretty interesting aha moment for us they gave us the ability to find a smaller population of parts from which to do detailed analysis on and and the focus there was to really just do speed path debug which is something that's that's pretty unique and difficult to do because what you're trying to do is is you're trying to catch the part in a place where it's failing but hasn't failed so bad ly that it's lost its mind and you can't extract any interesting information out of it so you want to get it at that point where it's just going to kind of croak but not totally croak and then once we get it there we can dump out tons of information out of the the the processor we can get gain an understanding of what's in the caches what's not in the caches what was the last instruction run what was the state of that instruction what was in my buffers you know am I am I stuck waiting on a load am I stuck waiting on a store you can get all kinds of interesting State information from these things once you get them to fail in a way that you can capture it trying to get the thing to fail but fail in a way that we could control the failure kind of like a controlled crash so we spent quite a bit of effort just getting to that point because we've engineered the platform in such a way that when when it thinks it's in trouble it's going to try to save itself and that's a feature but it's also a feature that kind of hurts us when we're trying to do a controlled crash because platform senses that it's in trouble and it pulls up in order to debug things we we had to like figure out oh we got to tear out our own safety features so that we can crash this thing in such a way that we're not crashing it too hard but we're we're crashing it hard enough that we could debug it and we had to get our own features out of the way that were really out there that you are hugely beneficial to overclockers because it allows them to once they've crashed it get the system back up and running again you know so there was a lot of work from you know taking it from this temperature dependency to getting these these select set of units onto Platforms in the debug lab working with the Silicon debug teams and just understand what part of our silicon is limiting us and then then further taking that over to the test to figure out are there knobs within silicon already that we can use to unlock more of that performance and you know I think we found a couple of pretty significant knobs to unlock a good 200 MHz of performanceI I think one of the cool realizations that we found on 4790k was there was an interesting temperature rolloff with respect to overclockability it was fairly distinct I wouldn't say it was perfectly distinct but what we found was is that as the junction temperature of the Silicon reached let's say 80 degrees 80 degrees seat which is well below where we would normally throttle right normally we run our parts up to about 100° C Junction temperature and at that point when it reaches 100° C we start to throttle the performance back to protect the part we started to see an overclocking rolloff at 80° C so we would see parts that were cooler below 80 over clock better until they got to 80 that was a pretty interesting aha moment for us they gave us the ability to find a smaller population of parts from which to do detailed analysis on and and the focus there was to really just do speed path debug which is something that's that's pretty unique and difficult to do because what you're trying to do is is you're trying to catch the part in a place where it's failing but hasn't failed so bad ly that it's lost its mind and you can't extract any interesting information out of it so you want to get it at that point where it's just going to kind of croak but not totally croak and then once we get it there we can dump out tons of information out of the the the processor we can get gain an understanding of what's in the caches what's not in the caches what was the last instruction run what was the state of that instruction what was in my buffers you know am I am I stuck waiting on a load am I stuck waiting on a store you can get all kinds of interesting State information from these things once you get them to fail in a way that you can capture it trying to get the thing to fail but fail in a way that we could control the failure kind of like a controlled crash so we spent quite a bit of effort just getting to that point because we've engineered the platform in such a way that when when it thinks it's in trouble it's going to try to save itself and that's a feature but it's also a feature that kind of hurts us when we're trying to do a controlled crash because platform senses that it's in trouble and it pulls up in order to debug things we we had to like figure out oh we got to tear out our own safety features so that we can crash this thing in such a way that we're not crashing it too hard but we're we're crashing it hard enough that we could debug it and we had to get our own features out of the way that were really out there that you are hugely beneficial to overclockers because it allows them to once they've crashed it get the system back up and running again you know so there was a lot of work from you know taking it from this temperature dependency to getting these these select set of units onto Platforms in the debug lab working with the Silicon debug teams and just understand what part of our silicon is limiting us and then then further taking that over to the test to figure out are there knobs within silicon already that we can use to unlock more of that performance and you know I think we found a couple of pretty significant knobs to unlock a good 200 MHz of performance\n"