Thread based boost is a lie

[phanbuey]

Here to rant about the boost algorithms Intel is using. Thread-based (1/2t 1 core boost) is total nonsense -- only a tiny % of apps can take advantage of that most overclockers are or should be using power/ temp based limits.

For example: here is my 170W limited 12600K @CPU-Z and Cyberpunk

Here is usage during cyberpunk:

Almost all modern applications, and windows, are constantly spawning multiple lower intensity threads, in this case intel CPUs run at a low all core-boost clock, and leave a ton of performance on the table instead of using the true load/power demands of the CPU. In most cases, a 5.4ghz single core boost is completely pointless, and the CPU could run all cores at that boost until it goes past a certain wattage or temperature then gear down as needed.

Does AMD Zen3 do this as well? Is it time to kill the "thread" counter boost algorithms and use true limits like wattage and temperature? The only time it seems to work is in benchmarks that force 1T/ NT at max wattage.

 

[Edwired]

That the way the CPU goes in frequency if it have temperature headroom it will boost higher same with power as both goes in hand to hand but if you limit the game to 60fps or 120fps it won't take advantage of the boosted frequency in the CPU. The only way you want the CPU to go full tilt is uncap the framerate and let it eat

 

[phanbuey]

That the way the CPU goes in frequency if it have temperature headroom it will boost higher same with power as both goes in hand to hand but if you limit the game to 60fps or 120fps it won't take advantage of the boosted frequency in the CPU. The only way you want the CPU to go full tilt is uncap the framerate and let it eat

I tested it though, if intel is loaded it boosts only to the 'all core boost' so for 12600K the single core is 4.9, and all core is 4.5 - so in the screenshot above, it would be at 4.5 regardless of limits, simply because each thread is being used (doesn't even matter by what % - could be 50%). It's a function of number of threads, not true load. Same with the 10850K. Basically that "single core boost up to" is exactly that.

 

[eidairaman1]

In the end it all relies on the schedulers in the OS,

 

[AlwaysHope]

I don't know about the intricacies of the Alderlake platform (only have experience with previous gen of Rocket Lake), just dial in manual power limits (obviously based on your cooling solution) set a manual per core or all core OC to what your system can tolerate & just game away all you like.
Heard & seen boosting algorithms can cause stutter in some games anyway. Sometimes the old school way of telling the system what & how you want it work well, even with modern hardware. You've got water cooling going by the specs in your System Specs so should easily handle more than 170w max.
I run up to 260w for PL1 but my thermal solution can handle it & it would be a poorly optimised game to push CPU load all the way up to there in the first place but in that regard its not the user's fault.

 

[phanbuey]

I don't know about the intricacies of the Alderlake platform (only have experience with previous gen of Rocket Lake), just dial in manual power limits (obviously based on your cooling solution) set a manual per core or all core OC to what your system can tolerate & just game away all you like.
Heard & seen boosting algorithms can cause stutter in some games anyway. Sometimes the old school way of telling the system what & how you want it work well, even with modern hardware. You've got water cooling going by the specs in your System Specs so should easily handle more than 170w max.
I run up to 260w for PL1 but my thermal solution can handle it & it would be a poorly optimised game to push CPU load all the way up to there in the first place but in that regard its not the user's fault.

Right and I mean that's what I did - and what most overlockers on TPU do.

But to think of your average user, reviewers, gamers etc. There is so much performance left on the table in virtually all games and applications for people who don't overclock.

There's a 10-20% IRL improvement if we could move past the fallacy that there are true "single threaded applications" that benefit from a 1core boost in the real world. Most people have multiple apps open at any given time, and almost all modern games are running multiple threads.

 

[Mussels]

Well... yeah.
Turbo doesnt care about the thread number, its about the wattages
A single core at low load can boost higher than a single core at 100% running AVX-512


This is what game mode in the OS is meant to do, sleep as much background stuff as possible so cores are less used, helping boost out.
You can do all core overclocks to bypass the issue, but that's less and less of a good idea as core counts go higher.

There is no magical cure here, if they boost to that speed for 1 millisecond, they'll advertise it - even if average users never benefit from it
(Thats why all core 'underclocks' are popular on ryzen, 5800x @ 4.6GHz is faster than stock multi threaded 4.4GHz, while ST *might* boost to 5.05Ghz while gaming.... or might sit at 4.4)



This is the fun where you get to tweak your system around your wants and needs. Static overclocks, PL1/PL2 settings, offsets, curve undervolts... I know PBO pretty well on the AMD side, but i'm not sure what the modern intels equivalent is.

 

[tabascosauz]

Does AMD Zen3 do this as well? Is it time to kill the "thread" counter boost algorithms and use true limits like wattage and temperature? The only time it seems to work is in benchmarks that force 1T/ NT at max wattage.

It kind of does...but not really? Generally across common applications you'll see a familiar drop after 2 cores, but the numbers aren't set in stone like Intel's turbo tables (where x threads gets x clock and nothing else), they vary a lot depending on what you're doing.

Ultimately the boost algorithm gets final say. It starts with the most optimistic number, the global freq limit, then starts ticking off the checklist to reduce perf. More threads? Less clock. Moderate to heavy AVX? Less clock. Bad cooling? Less clock. Bad SP (silicon quality)? Less clock. Straight up power virus wanting too much current (Prime95 at stock limits)? Murder the clocks.

Does Intel still offer the opportunity to customize your turbo tables to whatever you want on -K SKU? eg. multiplier for 1-core, multiplier for 2-core, etc.

To be honest, Windows 11 hasn't been great for clocks. Game performance is about the same, but game clocks are lower across the board for all my Ryzens. If you just go off the clocks alone, you'd be led to believe that Windows 11 has horrible game performance, but it isn't the case - it just makes do with less, apparently.

 

[phanbuey]

Does Intel still offer the opportunity to customize your turbo tables to whatever you want on -K SKU? eg. multiplier for 1-core, multiplier for 2-core, etc.

It does, but like you said, it's kind of dumb - 2 cores active at 20% each? 2 core boost. Powervirus on 2 cores? 2 core boost. It's not really an algorithm as much as a "x cores active" = "x cores boost". There isn't any more calculation past that.

I'm seeing with raptor lake the same thing... X cores go up to X MHz, and done.

When they did the 9900KS all they did was let it boost on all cores, and it was a big step up from the K for non OCers.

Zen 4 algo I think boosts all core, based on the leaked videos. If that's the case, intel is leaving that 10+% performance on the table just seems like a huge mistake for a company that's already trying to squeeze performance.

I just hope someone at intel comes across this and goes: "Hey guys, randos on the internet figured out that if we have a less stupid boost algorithm within our PL powerlimits, we can have more performance! For FREE!". To which I imagine the response would be "NO TIM! Active cores = Boost. They have to pay extra if they want thermals taken into account with TVB".

 

[Edwired]

Well at the end of the day some games aren't optimised to fully use latest and greatest cpu today. Just be happy of what you got running as some of us can't afford to buy the fastest cpu due to budget constraints. Surely I'm happy with the current system I have as it doing what I need it to do.

 

[AlwaysHope]

Right and I mean that's what I did - and what most overlockers on TPU do.

But to think of your average user, reviewers, gamers etc. There is so much performance left on the table in virtually all games and applications for people who don't overclock.

There's a 10-20% IRL improvement if we could move past the fallacy that there are true "single threaded applications" that benefit from a 1core boost in the real world. Most people have multiple apps open at any given time, and almost all modern games are running multiple threads.

I agree with your comment but when I game, that is the only "app" running on my system along with HWiNFO so I must be some kind of weirdo not conforming to the masses.
Multiple games I've played according to HWiNFO do not show only one or two cores carrying the load, it is spread out among however many threads I want to use up to the max of 16 on my i7. Turbo boost for single threaded applications is somewhat out of date these days, at least in the PC gaming world with anything from the last 10 -12 yrs.

 

[QuietBob]

Personally, I wouldn't complain about the boost if my CPU could do that with an AIO

 

[phanbuey]

yeah, the task manager in windows 11 is a little optimistic haha.

 

[fevgatos]

Thats where tvb comes in handy. Instead of using it to overclock though, i use to downclocks. I set 3t freequency at 5.7ghz and all core freequency at 5.4 ghz, and then i let TVB downclock depending on the temperatures.

 

[ThrashZone]

Hi,
I'm sure there are timers/ limiters that need to be maxed out or disabled to get turbo to function properly.

Maximum CPU Core Temperature [115]
CPU Core/Cache Current Limit Max. [255.75]
Long Duration Package Power Limit [4095]
Package Power Time Window [448]
Short Duration Package Power Limit [4095]

Think effective clock under stress tells the real story not task manager lol

 

[phanbuey]

Hi,
I'm sure there are timers/ limiters that need to be maxed out or disabled to get turbo to function properly.

Maximum CPU Core Temperature [115]
CPU Core/Cache Current Limit Max. [255.75]
Long Duration Package Power Limit [4095]
Package Power Time Window [448]
Short Duration Package Power Limit [4095]

Think effective clock under stress tells the real story not task manager lol

Task manager was just to show load distro on cores, not to measure anything. The measurement of frequencies was done using HWinfo. Stress isn't the issue - load under light stress is the issue (which is 99% of load scenarios).

Intel turbo boost works fine under full stress load on X cores, and one core full-stress load, without any additional utilization on the parked cores. Anything in between is sub optimal. It falls down hard in real life due to light load threads on all/multiple cores.

You can manually force it to scale all cores to stay within power limit (but remove the core utilization limits) and you will see a 10-15% increase in performance within the same parameters.

It's just an out of date approach as @AlwaysHope mentioned.

 

[Toothless]

yeah, the task manager in windows 11 is a little optimistic haha.

Does this on 10 too with my 4790k. I'd love for it to run at 5.5ghz.

 

[Mussels]

It's because overclocking the boards breaks the measuring algorithm they use

They'll do FSB x Multi, but when you're OCing weird shenanigans happen with dividers and windows cant see all them and gets so very very confused

 

[AlwaysHope]

It's because overclocking the boards breaks the measuring algorithm they use

They'll do FSB x Multi, but when you're OCing weird shenanigans happen with dividers and windows cant see all them and gets so very very confused

Good point, but MS should update the algorithm then. I mean don't they want to make the OS functionality please everyone in order to boost sales?

 

[ThrashZone]

Task manager was just to show load distro on cores, not to measure anything. The measurement of frequencies was done using HWinfo. Stress isn't the issue - load under light stress is the issue (which is 99% of load scenarios).

Intel turbo boost works fine under full stress load on X cores, and one core full-stress load, without any additional utilization on the parked cores. Anything in between is sub optimal. It falls down hard in real life due to light load threads on all/multiple cores.

You can manually force it to scale all cores to stay within power limit (but remove the core utilization limits) and you will see a 10-15% increase in performance within the same parameters.

It's just an out of date approach as @AlwaysHope mentioned.

Hi,
Turbo boost 3 is garbage and really should be disabled from startup
From my understanding speed step should be disabled and speed shift should be enabled in bios.

Little bit jpmboy on ocn posted a while back
Test it or not

remove Intel Turbo Boost 3.0...
Here's a proper way to disable the application (with device driver intact).
1. Right-click on the tray icon, and disable it.
2. Open Task Scheduler and disable Intel Turbo Boost Max Technology 3.0

3. Open Command Prompt and type “sc config ITBMService start= disabled”
4. Reboot.
Now back into bios and disable EIST (Enhanced Intel SpeedStep), and Enable SpeedSHIFT.
F10 to save and reboot.

 

[Mussels]

Good point, but MS should update the algorithm then. I mean don't they want to make the OS functionality please everyone in order to boost sales?

The boards are breaking the algorithms to overclock
Especially when you start altering boost algorithms on intel, they'll lie to the CPU about its current state to get it to boost higher - and then windows gets messed up data (Ah base clock is now 5.3GHz with boost of 150% so when we boost (which isnt even active any more) we now run at: GarbageHz

Here, all core OC on my ryzen.

But because the stock/turbo clocks are reported to windows, i'm always at 123% at this clock speed

But since it's not doing whatever intel boards do when they OC, task manager still gets it more or less right

(I'm testing 4.7GHz, if that wasnt obvious - current BIOS has broken curve optimiser, and i felt like it)

 

[blkout]

I tested it though, if intel is loaded it boosts only to the 'all core boost' so for 12600K the single core is 4.9, and all core is 4.5 - so in the screenshot above, it would be at 4.5 regardless of limits, simply because each thread is being used (doesn't even matter by what % - could be 50%). It's a function of number of threads, not true load. Same with the 10850K. Basically that "single core boost up to" is exactly that.

It depends on if you’re running windows 10 or 11 unless you use third party to manage your thread priority. Things like speed shift, EIST, TVB and Turbo Boost V3 also factor in. Your results are going to heavily depend on your MB, BIOS settings and OS choice/config.

“
The boards are breaking the algorithms to overclock
Especially when you start altering boost algorithms on intel, they'll lie to the CPU about its current state to get it to boost higher - and then windows gets messed up data (Ah base clock is now 5.3GHz with boost of 150% so when we boost (which isnt even active any more) we now run at: GarbageHz”

Do you mean Extended CPUID?