[SOLVED] i9-9900k, 5.1Ghz, 0 AVX, did I win the Silicon Lottery?

MonsterMaxx

Distinguished
Jan 23, 2015
110
15
18,615
MB: Asus Maximus XI Hero
Pcsr: I9-9900k
Ram: Corsair 128GB, 3600mhz
GPU: Quadro M6000 24Gb
PS: AX1600i
CPU Cooler: H100i Pro
Drives: 2x M.2 Samsung 970Pro in a stripe
OS: Win10

I found this guy who was OCing my exact setup, he seemed to know what he was doing, so I followed his instructions

First off, the Corsair RAM will NOT post at advertised speeds. XMP is enabled, but I have to slow the ram down to 3333mhz to post and run. Working with Corsair to resolve this (getting nowhere.)

Secondly for my CPU, I did not need nearly 1.32V to reach 5.0Ghz even with 0AVX on all cores. It ran fine at 1.275V.

The H100i was set to 'extreme' and sure enough the pump and fans were at advertised speeds. So good there.

At 5.0Ghz it had quite a bit of headroom for temperature so I cranked it up to 5.1Ghz w/ 0AVX. It booted fine, but no 5.1Ghz, stuck at 5.0. Back into bios, I tried the Extreme Tweaker setting to enabled and now it booted up at 5.1Ghz. This worked, Ran for a few minutes under Cinebench and crashed. Bumped the voltage a couple times and ended up being able to complete Cinebench at 5.1Ghz, 0AVX, 1.32V, 90°C with a Cinbench score of 12925. (IDK if this is good or not, but remember, the ram is running slow so once I get this resolved it should be even faster).

I tried 5.2Ghz with 1AVX. Booted to windows and ran, but cinebench would not complete.

However, I noticed the first two cores were 10-15°F colder so I cranked them up to 52x. Cinebench completes and AIDA64 reports no throttling. But, subsequent testing in my (often single core CAD program) showed this to be slightly slower than all cores at 5.1. Also, I never saw any reporting that a core was actually up at 5.2. IDK.

It's at 5.1Ghz w/ 0AVX now. stable


In my real world CAD benchmark (a timed script of tasks) it's a little bit faster than where I started today.

Only discrepancy I found with the video is that I still do not have the CPU locked into 5.1Ghz all the time. When unloaded it will drop to 3.6Ghz. Also CPU-Z reports .015V less than I set it to in Bios.

Did I win the Silicone Lottery with this CPU?

3-24-21-Apollo-BMOC-5.1-AVX-0-1.32V-HOT-cinbench-score.jpg

3-24-21-Apollo-BMOC-5.1-AVX-0-1.32V.jpg
 
Solution
MonsterMaxx,

Understanding how workloads differ in stress test utilities can be a minefield of confusion.

It's important to emphasize that CPU Utilization is not actual CPU workload. Most users, particularly gamers, associate Core temperatures with CPU Utilization, which is inaccurate. As workload increases, CPU resource activity can only increase to a maximum of 100% Utilization, while CPU Power consumption can increase well above 100% TDP, especially if overclocked. So when observing thermal performance, since Core temperatures respond directly to Power consumption, it's more relevant to monitor Power than Utilization.

As per Intel’s Datasheets, TDP and Thermal Specifications are validated...

Eximo

Titan
Ambassador
Can always check yourself against siliconlottery.com They have binning statistics.

Far as I can recall, doesn't seem like it, 90C is pretty warm. If you go back to Skylake the number that could do 5Ghz was pretty low, Kabylake it was fairly achievable, and Coffeelake basically came with that capability, and this is the second revision.

Have you modified the cache speed?
 
Don't mean to burst your bubble, but 5.1 GHz isn't anything too special for a 9900k as they love to clock high. Basically every 9900k can do 5.0, and I'd imagine the majority could do 5.1 as well.

I would not consider your overclock safe. 90c in cinebench? That's quite bad. It would probably run even hotter and potentially throttle in something like prime 95 or aida 64 that is a little more unkind to your cpu. Although, this is nothing out of the ordinary as the 9900k runs very hot and this is about what I'd expect from a 240mm aio.

You might want to go back to 5ghz with less voltage to settle those temps down.
 

Phaaze88

Titan
Ambassador
Ahh! Why is iCUE Dashboard reading in Fahrenheit?


Cinebench(R15/20/23) doesn't even run AVX, plus the load fluctuates too much for a proper cooler stability test.
Prime 95, Small FFT, AVX/ AVX2/ AVX-512/ Off, 15-20mins.
OCCT, Test cpu, small data set, normal mode, steady load type, SSE instruction set, 15-20mins
^Those are the go-to thermal tests. Even though the recommendation is for AVX off, these loads will accurately replicate a real world AVX load at its peak.

Aida64 doesn't run AVX by default either, as far as I know, plus people tend to just check off everything related to the cpu, when FPU only is all that should be run.

Then there's Silicon Lottery's binning statistics(for reference):
Coffee Lake RefreshAll Core SSE FrequencyAll Core AVX2 FrequencyBIOS Vcore% Capable
9900K4.80GHz4.60GHz1.275V100%
9900K4.90GHz4.70GHz1.287VTop 91%
9900K5.00GHz4.80GHz1.300VTop 30%
9900K5.10GHz4.90GHz1.312VTop 5%

^These are most likely revision 1 chips. The bins should be a little better for rev 2 chips.

It's too hot to be hitting 90C in Cinebench(version R20 or later, at least), and no telling what the Aida settings were...
Ideally, the cpu should be below 80 and no higher than 85C under load.
 

MonsterMaxx

Distinguished
Jan 23, 2015
110
15
18,615
OK, no lottery for me :( darn, I never win.

Or at least not using the H100i. Seems like it goes faster, but then overheats.

I slowed it down to 5.0Ghz, 0AVX, 1.25V, temps are still high on Aida64 (FPU only), no throttling. First few tests it throttled and I kept lowering the voltage until it didn't throttle. Ended up at 1.25V. Seems fine there.

Still pretty hot, but I don't benchmark for a living, I do CAD & FEA and I think it'll be fine. I'm stopping for now.
2021-03-24-16_12_37-apollo.png
 

CompuTronix

Intel Master
Moderator
MonsterMaxx,

Understanding how workloads differ in stress test utilities can be a minefield of confusion.

It's important to emphasize that CPU Utilization is not actual CPU workload. Most users, particularly gamers, associate Core temperatures with CPU Utilization, which is inaccurate. As workload increases, CPU resource activity can only increase to a maximum of 100% Utilization, while CPU Power consumption can increase well above 100% TDP, especially if overclocked. So when observing thermal performance, since Core temperatures respond directly to Power consumption, it's more relevant to monitor Power than Utilization.

As per Intel’s Datasheets, TDP and Thermal Specifications are validated “without AVX.

Advanced Vector Extension (AVX) Instruction Sets were introduced with Core i 2nd Generation, then AVX2 with 4th Generation and AVX-512 with later Generations of certain High End Desktop (HEDT) X-Series, Extreme, i9s and i7s. Each AVX Instruction Set is progressively faster in calculation workloads, but can adversely affect stability by overloading your CPU, which will dramatically increase Power consumption and Core temperatures.

Concerning Core temperatures and Core voltage, do not exceed 80°C or 1.4 volts. Read paragraphs 8, 9 & 10 to see what Intel's own lab engineers have to say about it: Inside Intel's Secret Overclocking Lab

AIDA64 has 4 CPU related stress test selections (CPU, FPU, Cache, Memory) which have 15 possible combinations that yield 15 different workloads and 15 different Core temperatures. The individual FPU test is about 115% TDP workload, the CPU/FPU combination is about 90%, all 4 tests combined is about 80% and the individual CPU test is only about 70%.

The individual CPU test does NOT use AVX, however, each individual FPU, Cache and Memory test DOES invoke AVX, as well as any test combination thereof.

CineBench R15 does NOT use AVX, but CineBench R20 and R23 DO use AVX, as does Asus RealBench.

This can be easily checked by setting an AVX Offset, then running any test while monitoring CPU frequency in HWiNFO.

Utilities that don't overload or underload your processor will give you a valid thermal baseline. Here’s a comparison of utilities grouped as thermal and stability tests according to % of TDP, averaged across six processor Generations at stock settings rounded to the nearest 5%:

doQfNIZ.jpg


Although these tests range from 70% to 130% TDP workload, Windows Task Manager interprets every test as 100% CPU Utilization, which is processor resource activity, not %TDP workload. Core temperatures respond directly to Power consumption (Watts), which is driven by workload. Prime95 Small FFTs (AVX disabled) provides the correct steady-state workload for testing thermal performance.

In recent games with AVX, as well as real-world apps with AVX such as those used for rendering or transcoding, the AVX code is less intensive than Prime95 with AVX. So when heavy, fluctuating AVX workloads in games or apps spike to "peak" Power consumption, Core temperatures will typically approach, but not exceed P95 Small FFTs without AVX. The CineBench R23 Multi Core test shown above (as well as CineBench R20) is a good example of a utility which replicates heavy, real-world AVX workloads.

The CineBench Multi Core test is actually a "somewhat " steady workload, however, as it pauses between rendering cycles which allows VRMs and Core temperatures to drop, it can't be considered a true steady-state workload. As Phaaze88 alluded to, if OCCT's first test, called "CPU", is configured for Small Data Set, Normal Mode, Steady Load and SSE Instruction Set (no AVX), then it's a steady workload at more than 97% that's nearly identical to the 100% workload of Prime95 Small FFTs without AVX.

Shown below from left to right: P95 Small FFTs (No AVX), P95 Blend (No AVX), Linpack and IntelBurn Test.

gaaHaa3.jpg


Note the steady thermal signature of Small FFTs, which allows accurate measurements of Core temperatures. A steady 100% workload is key for thermal testing so the CPU, cooler, socket, motherboard and voltage regulator modules (VRM) can thermally stabilize.

Since observing numerical values alone does not reveal a complete perspective, the best way to understand workload dynamics is to observe Vcore, Package Power and Core temperatures using the graphs that are available in utilities such as Intel XTU, AIDA64 and SpeedFan, etc. HWiNFO also has graphs which can be opened by right-clicking on any sensor parameter.

If you'd like to learn more, then just click on the link in my signature.

CT :sol:
 
Solution

Karadjgne

Titan
Ambassador
Your cooler is insufficient for the task. CAD is one thing, it responds very well to high speed single core and IPC. Arnold is a whole different beast.

If your cpu scales from 0°C to 100°C, then figure it's wattage in relation. Under max OC stress at a 100% flat load, it's going to hit 250ish watts.

Your cooler capacity is right around 250w. At max OC your cpu could easily hit 100°C. A 1:1 ratio.

For those kinds of workloads, you'll need a larger wattage capacity. @ 1.5x or better is preferable. A 280mm AIO is somewhere around 300w, the largest aircoolers are 250w+ and a 360mm AIO is closer to 350w capacity.

With the larger capacity, that ratio drops, at 250w on a 350w AIO you are only @ 70% saturation. Meaning what got 100°C before is likely to be far closer to 70°C depending on efficiency.

While all that math isn't close to exacting, it's close enough to figure approximate needs.
 

MonsterMaxx

Distinguished
Jan 23, 2015
110
15
18,615
CompuTronix, so I make sure I understanding what you are saying is that P95 Small no AVX is the test we should be running to determine stability on real world applications? That and/or the latest version of Cinebench?

Karadjgne, Yes, that's my conclusion too. It'll go faster, but not with the H100i Pro. Do you have a suggestion for something else?

Thank you both for the explanations.

It seems that I'm reaching the limit of voltage under load here. That is that anything in the 1.32V range just overheats. It needs at least 1.315V to OC to the next multiplier. and 1.325 to do 5.2Ghz.

Maybe a de-lid?

Or just leave it alone...
 
Last edited:

Karadjgne

Titan
Ambassador
P95 is for temps. Stability is something else entirely. You can have a stable pc that'll run upto 100°C, or an unstable pc that never goes past 70°C.

So you'd use 2 different programs to determine the limits of both. Cinebench, Aida, occt, asus real bench etc are all good for stability, their ups/downs, AVX, ram, linpack, blender mixes will challenge the validity of stability. You'd pretty much ignore temps in these tests unless it's obviously far too high for comfort.

Prime95 small fft is a flat 100% workload. So once you feel like the pc is stable, you'd use prime to set the bar. If the stable pc comes back as 80°C under a full small fft torture, that's good enough for most ppl, it means their gaming temps (usually a 70% load at best) will be in the mid 50's - 60. If the test comes back closer to 100°C, then it's time to rethink the OC or rethink the cooling solution. That's a thermal decision, not a stability decision.

2 different tools for 2 different purposes.

As for cooling, Arctic Freezer II 280/360mm, nzxt Kraken x62/63/73/Z73, Evga 280mm, Fractal S36, Corsair H115 series, really any of the better 280/360mm aios are somewhat similar. The capacity and ability is pretty much the same, they all use the same few OEMs for the pumps and rads, the difference being the efficiency of the fan/rad combination. Arctic and nzxt enjoying a slight lead over the others, but that often has to do with others using rgb fans that aren't quite suitable for rad purposes. Asus is overpriced, even if about the best, because it comes stock with Noctua fans. And some serious Oled pump stuff.

Bigger does not necessarily mean better. Bigger simply means bigger. At 100w, all the 240/280/360mm have roughly the same performance. It's only at extreme ends of ability that a 280mm becomes markedly and obviously better than a 240mm etc. Can't put 250w into a 250w cooler and expect it to be the same as a 350w cooler.

Your cooler size/ability will be determined by your OC needs. Which brand is determined by your wants, aesthetic needs, budget, audible requirements etc.
 
  • Like
Reactions: Phaaze88

CompuTronix

Intel Master
Moderator
MonsterMaxx,

As you can see on the %TDP scale I posted above, "Stress Tests" are divided into two categories; "Thermal Test - Steady Workload" on the left and "Stability Test - Fluctuating Workload" on the right.

As Karadjgne has correctly pointed out, testing "thermal" performance and system "stability" require the use of different utilities. However, it's important to emphasize that for thermal testing, run only Prime95 Small FFTs, but do so with all AVX test selections disabled.

For a complete understanding of how to properly test cooling performance, please read our temperature guide, linked below.

CT :sol:
 
  • Like
Reactions: Phaaze88

MonsterMaxx

Distinguished
Jan 23, 2015
110
15
18,615
Well, I'm able to get the Ram up to 3466mhz with the settings in the vid above. I tried adding a little voltage, but 1.35 seems to be where it's happy.

Corsair is peddling at this point. All I was looking for was some OC help from them and they have not made this a pleasant experience.

The CPU is stuck at 5.0 0AVX 1.25V, anything more and it overheats, but it runs everything I've thrown at it at those speeds.

I think it's a good CPU in the lottery. It boots and runs at 5.2, but when I hit the gas with a stressor it overheats. Seems to like 5.2 1AVX and 1.325V, but exceeds 90°C pretty quickly, even boots w/ 0AVX at 5.2.

The case is big, but not big enough for a 360 AIO. Maybe I could jury rig it in there. Or {shudder} a custom.

I have to think about where it's going...for now, it's running, and it's running faster than it was yesterday. Thanks for the help.

Oh. Forgot to mention. Switched the CAD stations' Quadro M6000 out for the GTX 2080 Super from my daily driver. Nice bump in benchmark performance....CAD...not so much. LOST big! That's what, a 6yr old card vs best last summer?

Quadro cards still RULE when it comes to CAD. That's this machine's J.O.B.

Now shutting down to swap back GPUs. Uggg, not fun.
.
.
.
maybe tomorrow
 

Karadjgne

Titan
Ambassador
Quadro cards still RULE when it comes to CAD. That's this machine's J.O.B.
It's why they exist. It's also why that kind of specialized card cost so much. That and the tech support. I swear those guys are related to Jimmy John's sandwiches. Call'm up, tell them there's a glitch and before you can tell them exactly whats up there's a little dude in a black and green shirt on your doorstep. You don't get that with the 2080S.
 

Eximo

Titan
Ambassador
Once had a weird glitch with a Quadro driver. User's mouse pointers would disappear going over the render area in AutoCAD.

Called up Nvidia, they called up Autodesk and started working on it. We had a new driver the next day, and about a week later there was a new official driver release, presumably taking care of this problem across many applications. That is why each card is so much.

If I recall we were early adopters of Windows 7 Enterprise, the latest Kepler Quadros, and we were the first to report the problem with AutoCAD 2010? Something like that.
 

MonsterMaxx

Distinguished
Jan 23, 2015
110
15
18,615
I played musical chairs with GPUs this morning...long story short, everybody is back in their proper spot and all's running as should be.

I believe I've reached the max speed for this hardware under the H100i cooler. So until I can acquire a better cooler (and maybe de-lid) I'm stopping. It's can go faster, but it's pretty fast now
 
Last edited:

Karadjgne

Titan
Ambassador
I believe I've reached the max speed for this hardware under the H100i cooler. So until I can acquire a better cooler (and maybe de-lid) I'm stopping. It's can go faster, but it's pretty fast now
One of the most intelligent responses to advice and testing that I've read in a while. Most will argue or disseminate or look to blame something or someone else and refuse to accept reality. "But what if I do this...."