Looking for guidance on boosting the i9-10900K's performance.
Looking for guidance on boosting the i9-10900K's performance.
Hey guys, I'm just starting out with large data sets. I moved to smaller ones and temperatures hit the 90s, which caused quick BSODs. With small datasets, I can't even reach more than 5.0GHz without crashing. I think I need to learn a lot. Sorry for the mix-up, I'm still a beginner.
It's all good.
The current cooler isn't adequate for testing all core overclocks on that cpu - stock settings it's fine though.
1.40v is also too much Vcore. Dial it back down around 1.30v, or so.
Silicon Lottery - the company - gets away with even lower Vcore, but they're using settings I'm not familiar with. It's something new that came with 10th gen.
@TravisPNW
Are you familiar with these 2 columns I marked in red?
Comet Lake
All Core SSE Frequency
All Core AVX2 Frequency
Per Core Frequency
All Core Die Sense Vcore
Power Limit
% Capable
10900K
4.80GHz
4.70GHz
6C+100MHz
3C+200MHz
1.130V
210W
100%
10900K
4.90GHz
4.80GHz
6C+100MHz
3C+200MHz
1.150V
220W
Top 99%
10900K
5.00GHz
4.90GHz
6C+100MHz
3C+200MHz
1.170V
230W
Top 68%
10900K
5.10GHz
5.00GHz
6C+100MHz
3C+200MHz
1.190V
250W
Top 21%
10900K
5.20GHz
5.10GHz
6C+100MHz
3C+200MHz
1.210V
270W
Top 1%
That should've been single core boost performance - that's always high. Under heavy load, it shouldn't be as high, but if it is, there's a problem in the bios.
Yeah, there was a problem with that some months ago:
https://www.youtube(dot)com/watch?v=qQ_AETO7Fn4
Basically board vendor shenanigans, end user gets screwed.
I'm trying to look up those things I highlighted, because I don't really understand them.
Yeah, I actually have that video in my collection as one of the ones I liked before starting the build. It also explained why my setup didn't maintain auto-mobo settings for long. After checking the chart you shared... it seems the OP's chip is in the top 21% (5.1) while mine is in the top 1% (5.2), even though the spec says 1.21v and I'm at 1.34v. I use the MSI Unify board, and all the reviews praised it overall, especially for overclocking... LOL
Respectfully, not quite so. Concerning Vcore, one size does
not
fit all. Overclocking any processor is always limited by two factors; voltage and temperature. Each Microarchitecture has a “Maximum Recommended Vcore”. For example, it’s important to point out that 22 nanometer 3rd and 4th Generation processors, such as your 4790K, will
not
tolerate the higher Core voltages of other Microarchitectures.
Here's the Maximum Recommended Vcore per Microarchitecture from 14 to 65 nanometers since 2006:
Each Microarchitecture also has a "
Degradation Curve
". Here's how the Degradation Curves correspond to Maximum Recommended Vcore for 22 nanometer 3rd and 4th Generation, which differs from 14 nanometer 5th through 10th Generation:
Degradation Curves are relative to the term “
Vt
(
Voltage threshold
)
Shift
” which is expressed in millivolts (mv). Users can not monitor Vt Shift. Vt Shift basically represents the potential for
permanent
loss of normal transistor performance.
There's more detailed explanations in the
Intel CPU Temperature Guide 2021
,
Section 8 - Overclocking and Voltage
. I suggest that you check out the entire guide, especially
Section 11 - Thermal Test Basics
.
Here's the nominal operating range for Core temperature:
Core temperatures above
85
°C are not recommended
.
Core temperatures below
80
°C are ideal
.
As
Phaaze88
has pointed out, a 240mm AIO is inadequate for cooling a 10 Core 20 Thread high-end "K" processor which can consume over 250 Watts, especially when overclocked at high Core voltage. If you intend to overclock, the 10900K demands a 360mm AIO or a custom loop in order to keep it cool.
Silicon Lottery
can professionally delid your processor, which can drop temperatures by 5 to 12°C. They can also "bin" your processor so you'll know its overclocking capability and the settings needed.
Guys
,
Voltage and temperature numbers get flung around forums like gorilla poo in a cage. For example, users will often say "I ran AIDA64" ... yes ??? ... and ... ??? what exactly did you run ??? 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 temperature possibilities. When ambient (room) temperature isn't mentioned, and load test conditions aren't defined, the Core temperatures you see on various websites and forums can be highly misleading. Therefore, several points need to be clarified and emphasized.
0Artur0
, in order to provide any meaningful apples-to-apples comparisons, as
Phaaze88
alluded to in his 1st and 3rd posts, it's important to be
VERY specific
, otherwise, all we have is apples-to-oranges thermal fruit salad in a blender. When discussing thermal performance, there are 3 major variables;
environment
,
hardware
and
software
. By taking a methodical approach, variables in
environment
and
software
can be accounted for, which then leaves differences in
hardware
to sort out. This reduces the major variables to their lowest common denominators, so test results make sense, are repeatable and easier to compare.
There's been no mention of ambient (room) temperature (
environment
), for which the International Standard for "normal" is 22°C or 72°F. Ambient can be a HUGE variable. Users write into our forums who live anywhere from the Arctic Circle to the Equator, so air temperature at the computer's intake might be anywhere from 10°C (50°F) to 40°C (104°F). If you don't say, and we don't ask, then we're blundering blindly forward based on an unknown major variable.
In your 2nd post you provided very limited specifics. It's always critical to define your exact software load conditions. Most users don't realize how
much
“stress” tests
vary
, which can be characterized into two categories;
stability
tests which are
fluctuating
workloads, and
thermal
tests which are
steady
workloads. Prime95 Small FFTs
(
AVX disabled
) is ideally suited for testing thermal performance
, because it conforms to Intel's Datasheets as a
steady-state
100% workload with
steady
Core temperatures.
As per Intel’s Datasheets,
TDP and Thermal Specifications are validated “
without AVX
”
.
With respect to the %TDP scale shown below, when heavy "real-world" AVX workloads are at "peak" load, such as video transcoding apps like HandBrake (which are fluctuating workloads), the workload will typically
approach, but not exceed
P95 Small FFTs
without
AVX. The CineBench R23 CPU Render Test shown below is a good example of a utility which replicates real-world AVX transcoding workloads. Prime95 Small FFTs (all AVX test selections
enabled
) is nearly a 130% workload, which is unrealistically higher than real-world AVX workloads. This is why an AVX "Offset" is used to keep Core temperatures in check, just as
TravisPNW
pointed out.
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%:
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
actual workload.
Core temperatures respond directly to Power consumption (Watts), which is driven by Vcore and workload
. Prime95 Small FFTs
(
AVX disabled
)
provides the correct workload for testing thermal performance.
When running heavy utility tests or AVX apps, in addition to Core temperatures, always keep an eye on Package Power consumption (Watts). Download the "Portable" version of
HWiNFO
and run "Sensors Only". However, numbers alone can't reveal the big picture. The best way to visualize thermal performance is to observe how your hardware responds to software workloads
on a graph
. In addition to a few other select utilities, HWiNFO also has graphs. Just right-click for "Show Graph" on the parameters you want to see.
The "Charts" in SpeedFan span 13 minutes, and show how each test creates distinct thermal signatures.
Figure 12-1
Shown above from left to right: Small FFTs, Blend, Linpack and IntelBurn Test.
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.
Phaaze88
, in OCCT 7.3.2, if the first test, called "CPU", is configured for Small Data Set, Steady Load, SSE Instruction Set, then it's very nearly identical to Prime95's Small FFTs without AVX. When CineBench R23 is configured for MultiCore, Test Duration 10 minutes or more, although it uses AVX and is a somewhat fluctuating workload and pauses between renders, if you observe CPU thermal behavior on any utility that can display a temperature graph, you'll see that (when configured as described) P95, OCCT and CineBench R23 all provide workloads within a degree or so from one another.
CT
Good news. I frequently checked your posts while working on my December build and during overclocking sessions. The warning about core temperatures above 85°C led me to stop pushing the overclock at 5.2ghz all cores. During stress and stability tests with my 360mm AIO, I was around 80-85°C, which is acceptable for me. So I decided it was fine. This happened on January 8th, and six weeks later the system has remained stable. As discussed on upthread, I’m thinking about returning to stock speeds. The boost was beneficial for benchmarks, but with the 10900k’s capabilities, do I really need those high core speeds now? Probably not.
I’m quite satisfied with the performance of the 10900k. I actually replaced my previous CPU (7700k) back in 2017, but it wasn’t necessary here. Even though the stress/stability temperatures are close to the maximum, the everyday usage stays well within the ideal range—50-70°C. Since I do a lot of encoding with Handbrake and that pushed temps up to 80°C, I adjusted the AVX offset to -2 and now encode around 70°C without any significant slowdown. This means only a 2-3 minute difference in encoding time, which I don’t mind.
Power consumption wasn’t an issue for me either. I chose a 1000W power supply because both the CPU and the 3090 GPU perform well under load. I don’t want to compromise on quality for cost savings.
Cheers.
did you test small configurations without enabling avx512 / avx2 / avx?