Intel 6900K overclocking. Voltage being misread.

[Carfax83]

I just bought a 6900K and a new motherboard (Asus X99A II with the latest UEFI BIOS), and so far, everything is going fine. I have it set right now to 4.3ghz and it's rock solid stable, but I know it can do more.

However I've noticed a weird discrepancy. The chip seems to be unresponsive to the adaptive voltage setting in the UEFI. I have it set to 1.2v for max turbo voltage, which means it's not supposed to exceed that amount, except for maybe a little bit.

But, this chip routinely hits 1.264v when I monitor the voltage under load in Windows with CPU-Z and AIDA64 Extreme. HWinfo also reported the same voltage.

It maxes out at 1.264v regardless of whether it's at adaptive or just at auto, which makes me suspect the adaptive setting isn't functioning properly. I even put the max turbo at 1.185v, and it still goes up to 1.264v

This is weird, because my previous CPU (a 5930K) worked very well with adaptive voltage, and I know that Broadwell isn't a huge change from Haswell so I don't know what's causing it.

Anyone else with a Broadwell-E CPU ever encountered this before?

 

[IllogicalGlory]

I'm the proud owner of a 4790K that hits 4.7GHz with 0.944V (during the CPU-Z benchmark), or so CPU-Z tells me.

 

[Carfax83]

^^ Very insightful, but perhaps you missed where I said I also tried HWinfo and AIDA64 Extreme.

From my experience, CPU-Z or HWinfo typically correspond well with the voltage settings in the BIOS. This the first time I've ever seen such a large gap, and I had a 5930K before that was overclocked to 4.4ghz using 1.225v and CPU-Z read that correctly.

 

[moonbogg]

I just bought a 6900K and a new motherboard (Asus X99A II with the latest UEFI BIOS), and so far, everything is going fine. I have it set right now to 4.3ghz and it's rock solid stable, but I know it can do more.

However I've noticed a weird discrepancy. The chip seems to be unresponsive to the adaptive voltage setting in the UEFI. I have it set to 1.2v for max turbo voltage, which means it's not supposed to exceed that amount, except for maybe a little bit.

But, this chip routinely hits 1.264v when I monitor the voltage under load in Windows with CPU-Z and AIDA64 Extreme. HWinfo also reported the same voltage.

It maxes out at 1.264v regardless of whether it's at adaptive or just at auto, which makes me suspect the adaptive setting isn't functioning properly. I even put the max turbo at 1.185v, and it still goes up to 1.264v

This is weird, because my previous CPU (a 5930K) worked very well with adaptive voltage, and I know that Broadwell isn't a huge change from Haswell so I don't know what's causing it.

Anyone else with a Broadwell-E CPU ever encountered this before?

I just OC'd my new 6800K last night. I was confused by how the voltage settings worked as well. I couldn't get the voltage to go up and down with clockspeed, but I finally figured it out. Did you try a fixed voltage just to see what happens? Also, I found my best results when using an offset value. Voltage is dynamic+auto+ a negative offset and the max voltage will have a lower ceiling every time I increase the negative offset. In short, I found these settings to work. I couldn't get things to work for me when inputting a specific voltage value. I had to leave it on auto and let it choose whatever it wanted and then I tamed it down with the negative offset. Results are consistent.

 

[FIVR]

Moonbogg got trolled into buying an inferior processor, instead of Zen? LOL. Priceless.


Quality work, I wish I could convince people to buy quad core Atoms over Core M. Would be even more hilarious than getting people to buy 6800k or 6900k over R7.

Trolling is not allowed
Markfw
Anandtech Moderator

 

[Arachnotronic]

I just OC'd my new 6800K last night. I was confused by how the voltage settings worked as well. I couldn't get the voltage to go up and down with clockspeed, but I finally figured it out. Did you try a fixed voltage just to see what happens? Also, I found my best results when using an offset value. Voltage is dynamic+auto+ a negative offset and the max voltage will have a lower ceiling every time I increase the negative offset. In short, I found these settings to work. I couldn't get things to work for me when inputting a specific voltage value. I had to leave it on auto and let it choose whatever it wanted and then I tamed it down with the negative offset. Results are consistent.

when I OC Intel CPUs, I leave it on auto and go from there.

 

[Carfax83]

I just OC'd my new 6800K last night. I was confused by how the voltage settings worked as well. I couldn't get the voltage to go up and down with clockspeed, but I finally figured it out. Did you try a fixed voltage just to see what happens? Also, I found my best results when using an offset value. Voltage is dynamic+auto+ a negative offset and the max voltage will have a lower ceiling every time I increase the negative offset. In short, I found these settings to work. I couldn't get things to work for me when inputting a specific voltage value. I had to leave it on auto and let it choose whatever it wanted and then I tamed it down with the negative offset. Results are consistent.

Yeah I figured out what was causing it. Apparently Intel hardcodes the base VID so that the CPU can draw additional voltage depending on the clock speed. This is why you can overclock with the voltage set to automatic although the CPU draws more power than is necessary most of the time. So basically, this same mechanism is preventing me from lowering the voltage using adaptive voltage, because adaptive voltage works off of the base VID. So while I can't lower the voltage using adaptive, I can INCREASE it and it works properly.

The only way to mitigate it is to use FIXED voltages, which for me is untenable since the processor doesn't enter idle states and is constantly running at whatever voltage you set. It's always better to use adaptive, or positive offsets at the very least.

I thought about using a negative offset, but the thing about that is that it affects all the other voltage parameters, including idle. That's why I never liked using negative offsets, because you could be doing something like surfing the net, and your PC will all of a sudden lock up and crash.

What do you have your 6800K set to? Mine is set to 4.3ghz (the sweet spot), and I think that is as high as I will go. 4.4ghz is probably doable, but the voltage increase required to get there is way too much for just an extra 100mhz.

That said, I am definitely noticing a performance increase in CPU heavy tasks, and even on the desktop due to the extra cores and the bump in IPC. Broadwell-E's increased IPC makes up for the loss in clock speed over Haswell-E.

 

[moonbogg]

Yeah I figured out what was causing it. Apparently Intel hardcodes the base VID so that the CPU can draw additional voltage depending on the clock speed. This is why you can overclock with the voltage set to automatic although the CPU draws more power than is necessary most of the time. So basically, this same mechanism is preventing me from lowering the voltage using adaptive voltage, because adaptive voltage works off of the base VID. So while I can't lower the voltage using adaptive, I can INCREASE it and it works properly.

The only way to mitigate it is to use FIXED voltages, which for me is untenable since the processor doesn't enter idle states and is constantly running at whatever voltage you set. It's always better to use adaptive, or positive offsets at the very least.

I thought about using a negative offset, but the thing about that is that it affects all the other voltage parameters, including idle. That's why I never liked using negative offsets, because you could be doing something like surfing the net, and your PC will all of a sudden lock up and crash.

What do you have your 6800K set to? Mine is set to 4.3ghz (the sweet spot), and I think that is as high as I will go. 4.4ghz is probably doable, but the voltage increase required to get there is way too much for just an extra 100mhz.

That said, I am definitely noticing a performance increase in CPU heavy tasks, and even on the desktop due to the extra cores and the bump in IPC. Broadwell-E's increased IPC makes up for the loss in clock speed over Haswell-E.

That makes sense, because when I had a fixed voltage of 1.25 it was stable at 4.2ghz. Using an offset voltage that equals 1.25 under load was not stable, likely because it was offsetting the lower end voltages as well like you said. So, I had to compromise a little. The offset voltage goes to 1.275 under load and has been stable, but at least its dynamic and most of the chip's life will be spent at idle voltages, under 1v. I can tweak further and certainly will, but that's where I am currently. I went for 4.3 and I can get it stable with good temps, but the voltage was just too high for my liking, crossing my 1.3v comfort zone. Maybe I can get there with less voltage, but honestly I'd rather be at 4.2 with lower voltage and heat, plus more peace of mind. I forgot my negative offset value, but its something like 0.03. Even at 4.2, the chip gets right around 1300 in cinebench whereas my 3930K@4.6 got 1160 despite a 400mhz clock speed advantage. At 4.3 the 6800K got like 1325 and I figured screw it, I'll stick to 4.2 for now. That's not bad for having less heat output with this one. I'll take it. Running 16gb of 3200 cas 16 in quad channel. No issues there. This motherboard has some cool RGB LED's all over it. That also makes it go a little faster, so that's a plus.

 

[Carfax83]

That makes sense, because when I had a fixed voltage of 1.25 it was stable at 4.2ghz. Using an offset voltage that equals 1.25 under load was not stable, likely because it was offsetting the lower end voltages as well like you said. So, I had to compromise a little. The offset voltage goes to 1.275 under load and has been stable, but at least its dynamic and most of the chip's life will be spent at idle voltages, under 1v. I can tweak further and certainly will, but that's where I am currently. I went for 4.3 and I can get it stable with good temps, but the voltage was just too high for my liking, crossing my 1.3v comfort zone. Maybe I can get there with less voltage, but honestly I'd rather be at 4.2 with lower voltage and heat, plus more peace of mind. I forgot my negative offset value, but its something like 0.03. Even at 4.2, the chip gets right around 1300 in cinebench whereas my 3930K@4.6 got 1160 despite a 400mhz clock speed advantage. At 4.3 the 6800K got like 1325 and I figured screw it, I'll stick to 4.2 for now. That's not bad for having less heat output with this one. I'll take it. Running 16gb of 3200 cas 16 in quad channel. No issues there. This motherboard has some cool RGB LED's all over it. That also makes it go a little faster, so that's a plus.

You made the right decision to stick to 4.2ghz. Every chip has a sweet spot, where performance and voltage consumption are at their most reasonable. Going past that provides little gain, but dramatically speeds up the deterioration of the processor. My sweet spot is 4.3ghz, and like I said earlier, I won't be going past that even though I know the chip can do 4.4ghz; likely at 1.3v and above. Still, I would recommend trying out adaptive voltage. It's probably not much higher than your current offset, without having to deal with the possibility of too much vdroop when idling.

And I just ran Cinebench and I got 1784. That's the first time I've ever run it, and I wasn't all that impressed to be honest. It seems like a very old benchmark which doesn't take advantage of the capabilities of modern CPUs. Does it even use AVX?

BTW, do you have AIDA64 installed? If so, can you run the cache and memory benchmark? I'm curious to see how a hexcore Broadwell-E compares to my former 5930K when it comes to cache bandwidth. Here's my best scores on my old 5930K @ 4.4ghz and the uncore at 3.5ghz with a 32GB DDR4-3200 CL14 kit:

Here's another one with the same CPU, but with a slower DDR4 kit and the uncore @ 4ghz:

Now here's my 6900K at 4.3ghz, and the uncore at 3.1ghz. Broadwell-E's memory controller is WAY better than Haswell-E. And this is with my slower DDR4 3200 kit. My faster one is being RMA'd.

 

[moonbogg]

As you can see, I am using the "trial version". Seems kinda slow. What gives? If none of these numbers really effect gaming FPS then I'm not too concerned really.

Also, I'm kinda jelly of that 6900K. Not that I would actually be able to use it, but you know, it would be cool to see 16 threads jamming along in the task manager. Also, it would be cool to give my loop a little more to work on.

 

[Carfax83]

As you can see, I am using the "trial version". Seems kinda slow. What gives? If none of these numbers really effect gaming FPS then I'm not too concerned really.

The benchmark is impacted mostly by the amount of cores, RAM speed, uncore frequency and also the raw speed of the processor which is why yours appears low compared to mine. My Haswell-E system was overclocked to 4.4ghz when I ran the benches. Thanks for posting that though, because I was curious as to whether Intel improved the cache performance for Broadwell-E vs Haswell, but it looks like they didn't. They improved the memory controller and fixed the bug that was affecting write performance. They beefed up the execution units and branch prediction as well, but the cache looks like it wasn't touched.

One thing that is a bit strange though, is that your memory read speed does seem kind of low. I figured it would be in the 60s somewhere, and not the 50s since you are using DDR4 3200 in quad channel mode.

Also, I'm kinda jelly of that 6900K. Not that I would actually be able to use it, but you know, it would be cool to see 16 threads jamming along in the task manager. Also, it would be cool to give my loop a little more to work on.

I don't know if I actually need it to be honest. Whilst I do some encoding from time to time, I'm primarily a gamer. The reason why I got it though is because I figured the hexcore prices will drop significantly once the mainstream hexcore parts come out next year, and since I wanted to wait for the revised version of Skylake-E to arrive, I figured it would be better to sell my Haswell-E system right now than a year and a half from now where it will be worth next to nothing.

One thing that surprises me about the 6900K though, is how cool it runs. Right now it's idling at around roughly 21c, which is the lowest idle temp I've ever seen for an air cooled CPU:

 

[moonbogg]

You're cooling it on air?! Holy crap. That's pretty good.

 

[tamz_msc]

What's your ambient? Idle cannot go below ambient.

 

[Carfax83]

You're cooling it on air?! Holy crap. That's pretty good.

Dat Noctua NH-D15 is a bad mutha

 

[Carfax83]

What's your ambient? Idle cannot go below ambient.

Ambient temp where my PC is, is about 69F. I optimized the airflow in my chassis for maximum cooling ability. Warm air doesn't get a chance to settle anywhere before it's blown out of the system.

 

[tamz_msc]

Bah 69F is 20.56C. I find it difficult to believe that those temperatures are accurate.

 

[Carfax83]

Bah 69F is 20.56C. I find it difficult to believe that those temperatures are accurate.

It's not going to be 100% accurate, but it provides a good estimation. Sticking my hand at the back of the main exhaust where the CPU is, I don't feel any warm air at all. It's all very cool.

 

[tamz_msc]

It's not going to be 100% accurate, but it provides a good estimation. Sticking my hand at the back of the main exhaust where the CPU is, I don't feel any warm air at all. It's all very cool.

Still, that is subjective. DeltaE values shouldn't approach zero. What does the package temperature reported by HWMonitor say?

 

[Carfax83]

Package temperature is much higher, at 35c.

 

[tamz_msc]

Package temperature is much higher, at 35c.

That is more reasonable. Considering heat transfer losses from silicon to solder to IHS, I'd expect the actual core temperatures to be even higher.

 

[lopri]

Try an offset instead of a fixed value?

 

[Carfax83]

Try an offset instead of a fixed value?

As far as I know, an offset is still limited by the base VID. So if the base VID says the processors requires X amount of voltage at Y clock speed, then you can't lower it.

 

[Carfax83]

Figured some might fight this interesting. Here's a comparison between a 6850K and a 5820K at the same clock speed of 4.2ghz. Everything is the same, including the uncore frequency as well. The only thing that's different is that the 5820K has a slightly lower CAS latency on the RAM.

What surprised me is how large the IPC increase is for Broadwell; about 10% faster for Cinebench, and 13% for Pi, both well above the advertised 5% increase for IPC that Broadwell is supposed to have vs Haswell.

I think the explanation is clearly due to the particular enhancements that Broadwell has, like the faster multiplier, divider, vector gather instruction etcetera which lend themselves well for heavy calculations:

5820K:

6850K:

Source