Homebrew Homebrew app Switch-OC-Suite

[Anxiety_timmy]

Erista with samsung ram
DRAM: 1.2v running at 2.064ghz
CPU: 1.275v 2.091ghz

Post automatically merged: Feb 11, 2023

Mostly stable so far, but had crashes at stock voltage at 1.968ghz RAM
Im currently running memtester and will check back if there are any errors.

 

[swutch]

Took some time to find stable settings for my OLED Samsung RAM.
I've tested it with TinyMemBenchNX, MemTesterNX, Kirby and the Forgotten Land & BotW, with full OV and without.

First i tried "AUTO_ADJ_MARIKO_4266" timings for my 4266 Mbps Samsung K4U6E3S4AA-MGCL RAM.
But Kirby crashed really fast with every frequency above 1932 MHz.
The same with "AUTO_ADJ_MARIKO_SAFE" timings.


With "NO_ADJ_ALL" timings i could play Kirby without crashes,
up to 2131MHz. But MemTesterNX showed me some errors.
I noticed that the errors disappear when I increase RAM Vdd2 voltage by 25mV.

And I thought I finally found stable settings. But then when I raised the temperature in my fan settings a few degrees to lower the noise, the crashes came back.
So with higher temps and full overclock my Ram seems to be stable only at 2064 MHz.
And OV doesn't do any difference.



@Cooler3D
Even if you undervolt the EMC Vddq by 50mV in your 2400MHz kip,
your EMC Vdd2 is still overvolted by 150mV.

I saw your Switch only draws 800mV GPU voltage at 1305 MHz.
That's impressive if it stays stable.
Would be nice if we could change the GPU voltage in KazushiMes Switch-OC-Suite too, and experiment with it.

 

[guily6669]

KazushiMe says you can edit in the dvfs table, I think you'd have to compile your own stuff not sure. I was reading a bit on RetroNX and apparently undervolting on CPU isn't very useful as they are tuned at the factory. Undervolting on GPU on the other hand may be useful in reducing power draw but only at 768+ and there needs to be good testing involved to check that performance isn't hampered.

GPU only governor will be better for some games yes, and vice versa.

I know we are talking about probably only saving few mah, but it's not about saving power but rather the fun of finding the lowest number for our switch (the super tiny power saving is just a extra)

EDIT: Forgot to say I asked here before about that dvfs table, I don't know where is it, but anyway I would like is having the feature on the online configurator

 

[Medel87]

I'm trying now 2131,2 mhz at RAM with 1150 mV, better results at tinymembench and 37 fps at crash bandicoot, if it stay stable, will try lower voltage, I wish we can UV GPU to get the best for my switch

Pd. Have a game freeze after 30 minutes, will try more things other day, it's later here

 

[Anxiety_timmy]

1.2125v, 2.102 is kind of unstable.

Post automatically merged: Feb 12, 2023

2.1056ghz RAM, pushed up to 1.25v, it seems stable so far.

 

[ChanseyIsTheBest]

Took some time to find stable settings for my OLED Samsung RAM.
I've tested it with TinyMemBenchNX, MemTesterNX, Kirby and the Forgotten Land & BotW, with full OV and without.

First i tried "AUTO_ADJ_MARIKO_4266" timings for my 4266 Mbps Samsung K4U6E3S4AA-MGCL RAM.
But Kirby crashed really fast with every frequency above 1932 MHz.
The same with "AUTO_ADJ_MARIKO_SAFE" timings.


With "NO_ADJ_ALL" timings i could play Kirby without crashes,
up to 2131MHz. But MemTesterNX showed me some errors.
I noticed that the errors disappear when I increase RAM Vdd2 voltage by 25mV.

And I thought I finally found stable settings. But then when I raised the temperature in my fan settings a few degrees to lower the noise, the crashes came back.
So with higher temps and full overclock my Ram seems to be stable only at 2064 MHz.
And OV doesn't do any difference.



@Cooler3D
Even if you undervolt the EMC Vddq by 50mV in your 2400MHz kip,
your EMC Vdd2 is still overvolted by 150mV.

I saw your Switch only draws 800mV GPU voltage at 1305 MHz.
That's impressive if it stays stable.
Would be nice if we could change the GPU voltage in KazushiMes Switch-OC-Suite too, and experiment with it.

Do you think you could send me the loader.kip's if you have them so I can put the information of them on the first page? Even if they aren't stable its good to know for others with Mariko. I'm also curious if 'AUTO_ADJ_MARIKO_4266' would give you better performance on TinyMemBench compared to the others at lower RAM clocks because you have that type of RAM.

I found your point interesting about raising the temperature causing more RAM stability issues when OCed. (I'm assuming you changed fan settings in a .ini file for atmosphere). This is a future point that should be investigated to see if temperature needs to be investigated with RAM stability.

 

[guily6669]

I'm very interested in the temp info too and to what point it affects differently the stability\max OC on each memory specially because despite having a cooling mod my memories have crazy thick bad thermal pads that are probably not doing anything and are probably preventing the backplate from contacting the whole pipe very well...

Though it's not like the backplate will do much as a cooler, it mostly should probably only increase time to reach full temperature by a bit but max temps will probably be almost the same crap and will take longer to cool down too after long time use...

I actually wonder how much better will temps be without the backplate with the air cooling almost the entire motherboard and all its components even if it's a very tiny airflow...

 

[Cooler3D]

Would be nice if we could change the GPU voltage in KazushiMes Switch-OC-Suite too, and experiment with it.

I'm working on dynamically adjusting optimization levels for GPUs, for specific consoles, and for specific frequencies. 800mV for 1300MHz effective performance (ST4+) is stable on all consoles. The most successful samples are able to remain stable at this setting, up to 690-700mV. Adaptation will allow sacrificing stability on unsuccessful samples, and not limiting the potential of successful ones.

 

[Medel87]

1.1V on CPU has a bit less performance for me

I tried and yes, little better without UV, but not much

 

[swutch]

Do you think you could send me the loader.kip's if you have them so I can put the information of them on the first page? Even if they aren't stable its good to know for others with Mariko. I'm also curious if 'AUTO_ADJ_MARIKO_4266' would give you better performance on TinyMemBench compared to the others at lower RAM clocks because you have that type of RAM.

I found your point interesting about raising the temperature causing more RAM stability issues when OCed. (I'm assuming you changed fan settings in a .ini file for atmosphere). This is a future point that should be investigated to see if temperature needs to be investigated with RAM stability.

"AUTO_ADJ_MARIKO_4266" is slower too.

Yes fan settings can be adjust in "\atmosphere\config\system_settings.ini"
My config for an OLED:

[tc]
use_configurations_on_fwdbg = u8!0x1
tskin_rate_table_console_on_fwdbg = str!"[[-1000000, 40000, 0, 0], [36000, 43000, 51, 51], [43000, 49000, 51, 128], [49000, 54000, 128, 255], [54000, 1000000, 255, 255]]"
tskin_rate_table_handheld_on_fwdbg = str!"[[-1000000, 40000, 0, 0], [36000, 43000, 51, 51], [43000, 49000, 51, 128], [49000, 54000, 128, 255], [54000, 1000000, 255, 255]]"
holdable_tskin = u32!0xC738
touchable_tskin = u32!0xC738
tskin_pcb_coefficients_console_on_fwdbg = str!"[8051, -45213]"
tskin_pcb_coefficients_handheld_on_fwdbg = str!"[7176, -33954]"
tskin_soc_coefficients_console_on_fwdbg = str!"[7831, 57590]"
tskin_soc_coefficients_handheld_on_fwdbg = str!"[9029, 4274]"

https://switchbrew.org/wiki/System_Settings#tc

edit:
I changed it to OLED Docked standards

 

[Medel87]

@ChanseyIsTheBest update my config, it seems stable at frecuency 1968000, more than this I have some freezes with 1100000 voltage.

Pd. With stock voltage at CPU it have little better performance like @guily6669 said, will try with stock voltage at RAM if it have difference too

 

[ChanseyIsTheBest]

@ChanseyIsTheBest update my config, it seems stable at frecuency 1968000, more than this I have some freezes with 1100000 voltage.

Pd. With stock voltage at CPU it have little better performance like @guily6669 said, will try with stock voltage at RAM if it have difference too

Is performance better then 1996 stock? So it's the 2064 one just 1968 mHz your running at instead?
Sorry I've been slow updating the first post.

 

[Medel87]

Is performance better then 1996 stock? So it's the 2064 one just 1968 mHz your running at instead?
Sorry I've been slow updating the first post.

1996 is stock for Mariko, I have Erista and stock is 1862, I reduced voltage from 1125000 to 1100000 and going up for frecuency from 1862400 to 1968000, and yes, it have better performance, I want to try if I put stock voltage at RAM give more performance like happen with CPU

 

[ChanseyIsTheBest]

1996 is stock for Mariko, I have Erista and stock is 1862, I reduced voltage from 1125000 to 1100000 and going up for frecuency from 1862 to 1968, and yes, it have better performance

Is this stable at stock RAM voltage as well? I should also make it clear that there's no good reason to undervolt RAM or CPU for that matter. The power saving is basically nothing and you can seriously impact performance very easily.

 

[Medel87]

Is this stable at stock RAM voltage as well? I should also make it clear that there's no good reason to undervolt RAM or CPU for that matter. The power saving is basically nothing and you can seriously impact performance very easily.

It should be stable with stock voltage as well, my swith for example is stable at 1996, ever higher, with stock voltage, I only want to get higher frec with minimun voltage, and if you search max frecuency with stock voltage, can have same effect with performance reduced because you use more frec with same voltage, this is the reason I want to try same frec without UV to see if it have better performance like happen with CPU, I don't know if I explained well, because I'm spanish

 

[Anxiety_timmy]

I tested power draw on memtester with 3 threads and same setup as before.
These are only the maximum load readings, i'll add the minimums later.
Readings are in watts

2091	11.45
1963	10.28
1887	9.88
1785	9.35
1683	8.78
1581	8.29
1428	7.57
1326	7.22
1224	6.96
1122	6.56
1020	6.28
918	6.14
816	5.96
714	5.78
612	5.56
510	5.29
408	4.98

Post automatically merged: Feb 14, 2023

Graph version

 

[Medel87]

I'm doing other tests, I'm trying to find higher frecuency supported by my switch, seems like 1250mv make some graphical glitches, so I will try with 1225mv for now, and going up little by little, for now, seems like 2192mhz work nice, will try tomorrow if have freezes or something while playing more time

Pd. Not stable, I have a freeze

 

[ChanseyIsTheBest]

I tested power draw on memtester with 3 threads and same setup as before.
These are only the maximum load readings, i'll add the minimums later.
Readings are in watts

2091	11.45
1963	10.28
1887	9.88
1785	9.35
1683	8.78
1581	8.29
1428	7.57
1326	7.22
1224	6.96
1122	6.56
1020	6.28
918	6.14
816	5.96
714	5.78
612	5.56
510	5.29
408	4.98

Post automatically merged: Feb 14, 2023

View attachment 353369
Graph version

Wish there was something like this for GPU as well this is really good, thank you for taking your time to do this. The readings also along very closely with your analysis of the power draw of NieR Automata. I wonder how much the DRAM overvolt is adding to this if you still have it activated? I'm assuming not much but am still curious.

 

[Anxiety_timmy]

I should probably do 2 and 1 thread tests just for the hell of it lol
And idle especially.

Post automatically merged: Feb 15, 2023

I need to test those later, because im currently not close to my house.

Post automatically merged: Feb 15, 2023

Wish there was something like this for GPU as well this is really good, thank you for taking your time to do this. The readings also along very closely with your analysis of the power draw of NieR Automata. I wonder how much the DRAM overvolt is adding to this if you still have it activated? I'm assuming not much but am still curious.

Maybe one way is to use a 3d rendering devkitpro example and set the resolution to something stupid like 15360x8640

 

[guily6669]

I wish someone would make a proper benchmark with full RAM+GPU+CPU capabilities and be able to chose any of them or combined.