Who's buying Skylake-X? (You may now change your vote)

[ddogg]

Hardware Unboxed managed to cool a Core i9-7900X down a bit by using a gigantic radiator in a custom loop. Prior to using this huge radiator, they were pulling about 420W from the wall at 4.6 GHz - post custom loop (and lower temperatures), they were down to about 370W from the wall at 4.7 GHz. That's somewhere around 40W that the CPU's power consumption actually dropped.

Sounds like a huge issue with leakage. Delidding may actually curb the power consumption significantly due to dropping temperatures. I'm not entirely sure how Intel plans to push 12C-18C parts with this sort of thermal performance. Throwing larger coolers at this issue is relatively ineffective due to the fact that the coolers themselves aren't generally the limiting factor - the poor contact between the die and IHS is.

I was waiting to see something like this as most review sites had AIOs which don't touch the performance of a custom WC. That radiator is not what I would call gigantic but needing a 360 60mm thick rad just to cool that CPU to acceptable levels is crazy. Guess delidding is the only option here if someone wants good temps.

 

[jj109]

40W swing is too much for a 20C decrease. That's over 10% and not even accounting for the increased clock speed! Tom's Hardware measured only a 5% increase from 60C to 100C.

It was pointed out on OCN that the 420W was measured while thermal throttling as well, so the unthrottled power would've been even higher.

Between the oddly low voltages for
high clocks, the high power consumption for low voltage, and this new result... I'm starting to think the fivr output voltage isn't being set/read correctly.

Anyone have insights on how HWInfo reads the applied core voltage?

 

[Timmah!]

This is a disaster. I don't see how they're going to fit the 12-18C models in this platform, are they going to include a 1000W titanium class PSU with these once average joe decides he wants to overclock his shiny new CPU? How are they even going to cool 18 of these toasty cores under the slightest load? The 12-18C die is turned into locked Xeons for a reason.

That rumor of revised 2066 boards with stronger power capability and soldered CPUs incoming
seems more and more plausible by the minute.

Yeah, it would not be surprising at all if the 12-18C die was soldered. Be it intended that way from the start as part of price segmentation (higher priced CPUs getting solder) or as a result of current outcry.

 

[pantsaregood]

You're right, the output voltage settings or readings could be wrong - it would be a bit more consistent with what we know about Intel's 14nm process.

Skylake-X cores are, in themselves, completely fine. The complicating factors seem to stack in a very negative way, though.

Poor IHS contact to the die (the thermal interface material Intel uses is fine, the gap between the die and IHS is just gigantic) is causing temperatures to spiral out of control and presenting a significant cooling bottleneck.

The inability to effectively cool the die is causing some nasty leakage, which is further raising temperatures. It's a negative feedback loop.

Realistically, AIOs and big air coolers like the NH-D15 should be able to easily dissipate 300W of waste heat. They're only ineffective here because heat isn't being transferred to the hestspreader effectively.

This also explains the huge variance we see in temperatures - some 7900Xs are hitting 75C stock while under load - others are breaking 95C and thermal throttling. That's consistent with the variance in the size of the IHS gap from the die.

Finally, soldering 12-18C parts may be necessity rather than reaction to outcry. If not solder, then manufacturing tolerances need to drop significantly. As it looks now, higher core count parts will throttle at stock speeds.

 

[Rifter]

What a disaster. Im so happy i didnt wait and went ryzen sooner rather than later. My original plan was to wait till now and then decide between SLX and ryzen. All this trouble because intel was to cheap to solder them, what a joke.

 

[dwade]

Stock 7820x is still faster and more efficient than a maxed overclock Ryzen. And it runs at cool 60's.

 

[ManyThreads]

Stock 7820x is still faster and more efficient than a maxed overclock Ryzen. And it runs at cool 60's.

This is why I am buying. Even if my OC plans fail, I still have a crazy fast CPU better than an equivalent Ryzen but no doubt at higher cost.

 

[lolfail9001]

The last time I checked, we were talking about 18C CPUs.

Last time i checked we were talking highest end SKU. Intel's highest end desktop SKU was a 3Ghz base clock since Haswell-E.

Well what d'ya know, Computerbase results are out, they test three versions of Prime95, and with version 26.6, which is non-AVX, the results are(using NH-U12S with two fans)

Well, that's another good point about power consumption: thermals. Got me thinking, both AMD for Vega and Intel for SKL-X outright claimed they are to be used with liquid cooling. Similarities are uncanny, aren't they.

If not solder, then manufacturing tolerances need to drop significantly.

Considering how good of result delidding soldered Haswell-E and BDW-E provided, i start to suspect that Intel just uses it's worst assembly lines for packaging HEDT CPUs.

 

[jj109]

You're right, the output voltage settings or readings could be wrong - it would be a bit more consistent with what we know about Intel's 14nm process.
The inability to effectively cool the die is causing some nasty leakage, which is further raising temperatures. It's a negative feedback loop.

Based on Tom's Hardware data, the leakage power increase from 75C to 95C on their 7900X sample was 5 watts. It's not nothing but not really material to the issue.

Finally, soldering 12-18C parts may be necessity rather than reaction to outcry. If not solder, then manufacturing tolerances need to drop significantly. As it looks now, higher core count parts will throttle at stock speeds.

On the contrary, the higher core count parts will have lower heat density from lower all-core boost clocks. It's the smaller dies that really need solder since they have more aggressive boosts.

That's assuming the firmware respects the set power limits (which apparently isn't the case for some brands' X299 BIOS). Without a TDP limit and a broken per-core multiplier, power density would remain the same while the heatsink cold plate gets hotter. Then they'll thermal throttle for a short time until the VRMs explode.

 

[lolfail9001]

That's assuming the firmware respects the set power limits

Unlocked CPUs ignore power limits.

until the VRMs explode.

Oh please, it will take more than that for any quality X299 board VRM to explode.

 

[tamz_msc]

Last time i checked we were talking highest end SKU. Intel's highest end desktop SKU was a 3Ghz base clock since Haswell-E.

Well, that's another good point about power consumption: thermals. Got me thinking, both AMD for Vega and Intel for SKL-X outright claimed they are to be used with liquid cooling. Similarities are uncanny, aren't they.

And this time around the highest end desktop part is a full 18C part from their 2nd largest die that probably has sub-3GHz base clock intended for 165W as a server CPU. Heck, we don't even know what clocks are on the 12C part which is supposedly launching in August.

Oh and what has Vega got to do with this?

 

[VirtualLarry]

Unlocked CPUs ignore power limits.

I'm not so sure that's true. I think that they do, essentially, but that it's just far more likely that BIOSes default the short / long power limits to max threshholds, when an unlocked CPU is installed.

(I've seen comments about installing a "K" CPU into a SFF / ITX rig, and then just setting the TDP power limits in BIOS to keep the CPU under a certain TDP.)

 

[beginner99]

Finally, soldering 12-18C parts may be necessity rather than reaction to outcry. If not solder, then manufacturing tolerances need to drop significantly. As it looks now, higher core count parts will throttle at stock speeds.

Yeah and given that the issue with manufacturing tolerances go all the way back to Ivy Bridge I doubt they will magically solve them now. If 12c+ parts are soldered, Intel will need to fire the guy responsible for this. I mean they will still need 2 ways of applying the IHS, eg 2 production lines. This costs money. When you already have the 2 lines, why not also solder all HEDT parts? At that point the additional cost is pretty small.

I'm not sure I want to support a company with such dubious practices...Actually I'm sure I don't want to.

 

[lolfail9001]

And this time around the highest end desktop part is a full 18C part from their 2nd largest die that probably has sub-3GHz base clock intended for 165W as a server CPU.

Yes, it has 2.7Ghz base clock and 3.4Ghz all core turbo as you mentioned, doesn't it?

But as it is, Intel will set clocks exactly high enough for it to beat TR in AMD's benchmark of choice, that's pretty straight forward.

When you already have the 2 lines, why not also solder all HEDT parts?

That's a curious case and it all comes down to this question: why does Intel use TIM for Xeon Phis? Because they did it even when they soldered every other high end part.

 

[beginner99]

That's a curious case and it all comes down to this question: why does Intel use TIM for Xeon Phis? Because they did it even when they soldered every other high end part.

Maybe thermal load differences between cores and resulting thermal expansion really is a huge issue with process tech getting smaller and smaller? Pure cost reasons doesn't make sense for phi as I suppose it is very low volume. Also given the smaller cores there would be more hot-spots and temp gradients all over the die adding further strain to a similar die size but with fewer big cores.

But even if that is all true, why the f*** can the make billions of tiny, tiny transistors in an extremely complex workflow and then not manage the simple task of applying the right amount of glue and TIM? That is what is beyond me...

 

[coercitiv]

I'm not so sure that's true. I think that they do, essentially, but that it's just far more likely that BIOSes default the short / long power limits to max threshholds, when an unlocked CPU is installed.

(I've seen comments about installing a "K" CPU into a SFF / ITX rig, and then just setting the TDP power limits in BIOS to keep the CPU under a certain TDP.)

Of course they don't "ignore" power limits, they abide power limits set by the mobo manufacturer. It just so happens that Z and X boards tend to come with (understandbly) loose settings in that regard. However, not all of them are like that, my MSI Z170 board did come with a default max current setting that influenced overclocking potential going past 4.2-4,3Ghz, effectively triggering throttling under max load.

And yes, I made tests with a 6600K limited to 35W and even lower through BIOS settings. That's one of the reasons I like Intel's power management when properly implemented and properly exposed through mobo settings, it allows all kinds of opportunities for powerful systems in small form factors.

 

[DrMrLordX]

Results start at around 12m15s. CLU is better by about 5°C than Intel's solder, at least for this particular 6950x. Delidded 7900x is definitely comparable to a soldered/stock 6950x.

Those are the droids we're looking for. Someone needs to set up that test ASAP.

Based on Tom's Hardware data, the leakage power increase from 75C to 95C on their 7900X sample was 5 watts. It's not nothing but not really material to the issue.

Hmm. One wonders what a 7900X would do delidded with direct-die cooling with custom water would do. Some kind of massive rad like a MO-RA3 9x140 Pro. Maybe with a chiller in the loop?

Oh and what has Vega got to do with this?

Nothing. Just somebody else trying in vain to turn this into an AMD thread.

 

[Edrick]

Got my motherboard. Installing Windows now.

 

[IndyColtsFan]

Got my motherboard. Installing Windows now.

I can't wait to hear your report on temps and overclocking. I think the 7820 is the route I may eventually go rather than the 7900.

 

[Edrick]

So far so good.

Stock (4Ghz all cores) and running Prime95 (full FMA3 mode) and my temps are not going over 60C (pulling 144 watts). I am having trouble with the voltage readings as cpuz and aids64 both are giving me incorrect readings.

Got RAM to 3200mhz without issue (no voltage bumps required on the IMC).

Going to try OCing soon, once I get a reliable reading on voltage.

 

[Edrick]

Its kinda cool how it has an offset for AVX and a different offset for AVX-512. They dont need to be tied together like I thought.

 

[.vodka]

This is very concerning, these things could be a potential fire hazard when overclocking, from VRMs with dubious heatsinking on most motherboards out there to boards with a single 8 pin connector risking overloading and PSU cables getting too hot.

What else could go wrong with this platform? As he puts it near the end, "is this meant for spec use only?" It's the HEDT platform FFS!

 

[ZGR]

Looks like it's best to wait on SKY-X. These motherboards have unacceptable cooling. Hopefully when the 12C and 18C gets released more robust X299 boards with better cooled VRMs are released.

On a side note, we need to do more thorough VRM testing on ALL motherboards because OEMs sure aren't (at least in the case of rushed X299).

 

[moonbogg]

This is very concerning, these things could be a potential fire hazard when overclocking, from VRMs with dubious heatsinking on almost every motherboard out there to motherboards with a single 8 pin connector risking overloading and PSU cables getting too hot.

What else could go wrong with this platform? As he puts it near the end, "is this meant for spec use only?" It's the HEDT platform FFS!

This is seriously bad. Very bad. I just don't know what to say about it. A few months ago I was certain Skylake-X was going to be incredible. But now, so many things have gone wrong, and these aren't little things like BIOS issues. These recent revelations actually do indicate a potential fire hazard. This is just crazy.

 

[Edrick]

Having some poor performance issues with NVMe drives on this x299 system. Getting half the speed I was on z170. Still working my way through the BIOS.