Friday, October 19th 2018
Intel Core i9-9900K De-lidded, Soldered TIM Outperformed by Liquid Metal
We kept seeing hints regarding Intel's 9000-series processors running hot, including from their own board partners. As it turned out, the actual results are a mixed bag with some running very hot and most others ending up being power-limited more so than temperature-limited. Our own review sample showed overall better load temperatures relative to the predecessor 8000-series processors thanks to the soldered TIM (sTIM) used here, to give you some context. But that did not stop overclocker extraordinaire Roman "Der8auer" Hartung from de-lidding the processor to see why they were not generally better as expected.
As it turns out, there are a few things involved here. For one, replacing sTIM with Thermal Grizzly Conductonaut (Der8auer has a financial interest in the company, but he does disclose it publicly) alone improves p95 average load temperatures across all eight cores by ~9 °C. This is to be expected given that the liquid metal has a vastly higher thermal conductivity than the various sTIM compositions used in the industry. Of more interest, however, is that both the PCB and the die are thicker with the Core i9-9900K compared to the Core i7-8700K, and lapping the die to reduce thickness by a few microns also does a lot to lower the CPU temperatures relatively. Overall, Intel have still done a good job using sTIM- especially compared to how it was before- but the current state of things means that we have a slightly better stock product with little scope for improvement within easy means to the consumer.
As it turns out, there are a few things involved here. For one, replacing sTIM with Thermal Grizzly Conductonaut (Der8auer has a financial interest in the company, but he does disclose it publicly) alone improves p95 average load temperatures across all eight cores by ~9 °C. This is to be expected given that the liquid metal has a vastly higher thermal conductivity than the various sTIM compositions used in the industry. Of more interest, however, is that both the PCB and the die are thicker with the Core i9-9900K compared to the Core i7-8700K, and lapping the die to reduce thickness by a few microns also does a lot to lower the CPU temperatures relatively. Overall, Intel have still done a good job using sTIM- especially compared to how it was before- but the current state of things means that we have a slightly better stock product with little scope for improvement within easy means to the consumer.
75 Comments on Intel Core i9-9900K De-lidded, Soldered TIM Outperformed by Liquid Metal
Or is it about squeezing the best possible performance out of a CPU / GPU when doing stuff you might actually do ? If you want to show me that Intel doesn't know what they are doing show me the CPU overheating with a core voltage under 1.4 running RoG Real Bench ... then I'll pay attention.
Why doesn't GM put a SUV on the road with the ability to tow 80,000 pound loads ? Because it's a) not cost effective b) serves no useful purpose and c) because what they are doing produces temperatures well below that of any concern doing what the PC is supposed to do. If you are running H264 at 5.0 GHz and 72C, you don't get any bonus points for delidding and doing it at 62C. All it gets you is bragging rights at www.minesbigger.com
We put them in the worst possible conditions so we know more about the probability of catastrophic failure. If it fails in the worst case scenario, there is a chance (however small) it is going to do the same under less stressful circumstances.
Though I have a custom loop with 2x 240 rads and play games with headphones on..
What's the point ? a) Nobody has to die seems like a good one to me and b) the test are done under **real life** conditions ... you know normal traveling speeds and impacts that are likely to occur driving around town. Crashing them head on into the space shuttle at 175 mph while the shuttle is landing might be equivalent to P95 test but both fall outside of real world possibilities. Doing crash tests at impossible impact loads on a TV program called "Let's see what crazy shite we can do next" is not likely to have an impact on auto safety standards.
Front crash test - 35 - 40 mph depending on who doing the tests
Side crash test - 38.5 mph
Rear Crash test = 25 mph
Understand the difference ? RoG Real Bench multitasking test would be the vehicle crash test equivalent of front, side and rear collisions at the same time at about 65 mph. P95 w/ AVX is the equivalent of a having a head-on with the Space Shuttle on your way to Walmart. Not exactly a "real world" scanario. In addition, of what use is a OC that passes 24 hours of P95 running 8 identical task threads when it fails in a multitasking unbalanced load test in 45 minutes, it's happened to me.
2. TV Commercials
Are you really trying to prove your point by claiming realism in TV commercials ? Next you'll tell me that case fans are capable of delivering their advertised specs (50% is typical) and monitors have the response times and HDs can fit the amount of GB (instead of 93% of what is) claimed in their ads. I can tell you this.... I have tried my son's Axe Body Spray and I didn't have 6 supermodels following me around the supermarket like in the commercial. The 10 second stunts represented in TV commercials are not "real life". Try pulling that 80,000 pound trailer up 7,000 feet in elevation at 65 mph in the rockies at 120F and you will have the equivalent of a P95 AVX test.
www.truthinadvertising.org/category/ad-alerts/tv-ads/
3. Worse **possible** condition
Agreed, but let's use all of the words you used .... emphasis being on the middle word. You didn't say "worst condition one can possibly create using loads that no combination of applications will generate and now matter how irrelevant to real word usage". The Prime 96 AVX test does not meet that "possible" qualifier with the exception of pursuing the next mersenne prime .... I'm not likely to live that long let alone my next build and I don't see AVX helping here.
Want to cause a catastrophic failure ? I have run P95 w/o AVX on my current box and stayed stable for 24 hours. It crashed in Real Bench's multi-tasking benchmark in 45 minutes. P95 places an identical load on all cores which is very different than placing varying loads on multiple cores, some using modern instruction sets and some not.
No matter what combination off programs / games we run on my PCs, it's simply not possible to create a similar load scenario to P95 w/ AVX. So the idea that P95 represents a real world **possible** scenario is blatantly false. Even RoG Real Bench pushes the validity of the word "possible" to its limits in real world usage.Just for clarification ... is it 1C core temps or 1C case temps that are the concern ? And what fan speeds are you seeing called for ?
I see gaming CPU core temps from low 50s (5 x 140mm of rad ... 700 watts or so of "max component generated heat") to mid 60s / GPUs from 39C (1200 rpm) to 44C (850 rpm max).... coolant temps max out at 33 - 34C, case temps at 26 - 28C. Water temps are measured in and out of the 2 rads (4 sensors) with 2 more sensors measuring case and ambient air temps. Accuracy is 0/1 C and displayed with a Reven 6 Eyes" in 5-1/4 drive bay. Air turnover is estimated at 2-3 times per second. Fans are inaudible at up to 850 rpm and pr outside of stress testing with Firmark / RoG RB. Presets allowing range of 350 - 850 rpm, but have yest to see them break 650... fans shut off is curve asks for < 350 rpm.
So getting to why I asked, wanted to understand the fan curve where 1C raise in CPU temp makes an audible difference. I have been using a "Inaudible PC" rule of thumb of:
1 fan per 50 - 75 watts of component usage for 120mm
1 fan per 75 - 100 watts of component usage for 140mm
As using the above has not left us with a build that is audible in any real world usage scenario (stress testing excluded) , Id like to see just how far that Rule of Thumb might be relaxed. With you sitting on the edge, your data would be a boon to that understanding
That's why it's called solder.
I'm not calling you out or saying you're wrong or anything, more so trying to expand my own knowledge by asking questions. :toast:
trog
It's a half ass attempt, b/c it's very hard to cool. Otherwise, you'd still get toothpaste. They deserve all the bashing they get. They're STILL trying to pinch pennies with a CPU of this cost.
Flame em back to pentium 4 days and quit buying this garbo. They haven't learned anything.
Basically you have people trying to defend an unstable OC by blaming the stress test tools instead of a poor soldering process,
and its on a product which you are expected to pay $600 for.
CPUs should be and are made to be stable to be run any of its supported workloads indefinitely and that includes AVX.
Funny thing is same can be said about FMA3, people have bashed it of being unrealistic, etc.
The irony is Monster Hunter World originally would crash on SandyBridge CPUs because the game uses FMA3 and the older CPUs doesnt support it.
Its not a big deal to me, each and every one should run their CPU as they see fit. But it is telling. And it also speaks of a lack of knowledge, because AVX is most certainly used even in gaming, and we actually are seeing more of it everywhere. Even if just part of the code. So 'we' blame AVX, we blame bad TIM or shitty application of it, and now even the soldered TIM is apparently done badly 'because 9900K gets hot'.
That is one weird way to view a clear trend of CPUs getting warmer as core counts and base/turbo clocks increase on the same node.
The same reason why we have liquid coolers cooling better than air coolers (even air coolers have liquid within the heatpipes).
edit: Well now i'm not actually that sure i was wrong because it doesn't say that at liquid form it has a higher resistivity, it actually never talks about other physical states, so i might actually be right but just because the metal we're talking about is in another state and not just at a higher temperature. I'm so confused :confused: