X3D (or is it AMD?) single core torture

III_Slyflyer_III said:
So obvious fun fact... SMT off also decreased my temps by almost 10~15C when stress testing... lol. Now I can probably push this all core OC past 4.725Ghz to 4.8Ghz+ maybe... Interestingly, my MT scores (not that it matters much for gaming) were not lowered by as much as I would have guessed. Still got over 25K in CB23 (vs 31K) and CPUz quick test was 10.5K (vs. 13.6K). However, after a few days here I can confirm gaming is even smoother. Why this is not "well known" knowledge is beyond me. Honestly, debate over in my eyes from my testing.
Click to expand...
Doing an static all core overclock removes the ramping up and down of clock speeds. This tends to be the cause of microstutter. I'm running Linux and through much of the year I have distributed computing programs running so the CPU doesn't downclock so the microstutter isn't there. I use a program called CoreCTRL for monitoring and controlling my AMD video cards but it also has a "Frequency governor" for Ryzen CPUs. I originally had this set to On Demand which seemed to be the best setting between power savings (when not running heavy loads at least) and performance. Since it's summer and I'm not running any DC projects I started to notice microstutter and changed the setting from On Demand to Performance. The cores no longer clock under 3.8Ghz but my microstutter disappeared. I have since left it on this setting especially since power usage doesn't seem to be much higher as idle temps are almost exactly the same as before.

The source of the microstutter may not have anything to do with SMT but instead power savings options.
 
SmokeRngs said:
Doing an static all core overclock removes the ramping up and down of clock speeds. This tends to be the cause of microstutter. I'm running Linux and through much of the year I have distributed computing programs running so the CPU doesn't downclock so the microstutter isn't there. I use a program called CoreCTRL for monitoring and controlling my AMD video cards but it also has a "Frequency governor" for Ryzen CPUs. I originally had this set to On Demand which seemed to be the best setting between power savings (when not running heavy loads at least) and performance. Since it's summer and I'm not running any DC projects I started to notice microstutter and changed the setting from On Demand to Performance. The cores no longer clock under 3.8Ghz but my microstutter disappeared. I have since left it on this setting especially since power usage doesn't seem to be much higher as idle temps are almost exactly the same as before.

The source of the microstutter may not have anything to do with SMT but instead power savings options.
Click to expand...
Well yes and no... the all core OC helps both with clocks ramping all over the place, the disabling of CPPC keeps single cores from being overloaded and SMT off helps with a core being divided into completing multiple tasks. In another thread, I show how this increased L3 Cache bandwidth by over 50%, likely helping in gaming as well.

Because I like the graph, from a gaming standpoint, Forza 5 shows this nicely below (ran at different points in time as I modified settings in the bios). I did not use Frame Generation for these as that removes the CPU limit, so I used what I have always used at 4K before that feature came out for Forza. First picture is with PBO and CPPC, second picture is with All Core OC, CPPC off, and third picture is with SMT disabled. As you can see, FPS average gets better and better, along with stutter count going down and 1% lows going up. I have other games I can show, but this is a good example with data points. I game at 4K, and with a 4090 notice the difference, so I imagine at lower resolutions, the situation could be exacerbated more.

PBO w/ Curve Optimizer used, SMT ON.
1686227819575.png


ALL-Core OC, PBO Off, CPPC off, SMT ON
1686227922782.png


ALL-Core OC, PBO Off, CPPC Off, SMT OFF
1686228010647.png
 
Wow informative breakdown and quite a performance gain for that game. Has this gain carried over to other games as well using the same technique?
 
learners permit said:
Wow informative breakdown and quite a performance gain for that game. Has this gain carried over to other games as well using the same technique?
Click to expand...
Yes, absolutely, that game was just the best way to show it graphically. I could show screens shots across various games I play showing similar results with numbers, just not a cool little graph like FH5 which I think highlights the CPU render the best.

I'd encourage anyone to simply try it, you have nothing to lose. It may not be the same with lower core count CPUs as I mentioned, I'd have to disable one of my CCDs to test that theory out I guess, but it is possible to do if enough people are curious. I think having the extra cores helps based on what I have done, if you take away SMT on an 8 core CPU, results may not be the same for that specific setting, but I would recommend ANYDAY going with an all core OC and CPPC off over PBO (even with a custom curve).
 
Another thing I think folks under utilize is the ability of newer hardware to store multiple bios profiles. Why not have a gaming and a productivity profile saved in bios to have the best of both?
 
learners permit said:
Another thing I think folks under utilize is the ability of newer hardware to store multiple bios profiles. Why not have a gaming and a productivity profile saved in bios to have the best of both?
Click to expand...
I do, and yeah, I think a lot of people forget this. Interestingly though, I stopped using the PBO profile all together. Other than toggling SMT for productivity work, the all-core manual OC performs better in both gaming AND productivity, while running 10C cooler on average in all types of workloads. It also uses less power while under load.
 
Haven't tried a manual OC on the 7800 yet but will give it a go this weekend. I'll comment on the result sometime next week as this weekend will be busy for me.
 
learners permit said:
Haven't tried a manual OC on the 7800 yet but will give it a go this weekend. I'll comment on the result sometime next week as this weekend will be busy for me.
Click to expand...
I'd be curious if it work (or if you even can?) on AM5. I think it worked so well for me on AM4 because the 50% plus increase in L3 cache bandwidth helped overcome the inherent latency issues with the I/O die talking to the DDR4, even when coupled with FCLK. Not sure how things will go with DDR5 and AM5, so I am very much curious.
 
You must log in or register to reply here.
Back
Top