Horrible Ryzen 7 Overclock!

[porina]

9 hours ago, jjohnthedon1 said:

Erm thoes temps are crazy I do t go over 55 ever at 1.42v 4ghz

also up your llc to stop the droop it will yield lower idle voltages

I was going to ask under what conditions you test at, as obviously they can make a huge difference.

 

Who cares about idle voltages? Either my systems are working hard, or they're off.

8 hours ago, jjohnthedon1 said:

69 max temp at 4.1

passed Aida 64 with r15 on loop for 5 hours 

4ghz max temp 55 

no point in cooking the cpu with high voltage and temps for 1 fps gain 

There it is. Weak loads. Well, Aida64 can start some mild heating if you uncheck CPU and only run FPU stress test, but I never found it really useful as a stress. As stated, my testing was done using Prime95 29.1 as that was the first version to add Ryzen support and use AVX. Older versions detected it as an older AMD CPU and used non-AVX code. Also I was using 64k FFT to focus the load on the CPU itself.

 

I'd be interested if you could try the same and see what temps you get, if it is even stable.

6 hours ago, MageTank said:

You already know, Linpack MKL scares me, lol. If you are not using a delidded/soldered chip, and you are not on water, avoid MKL like the plague. Stock clocks will still get hot, lol.

Ooh, had to look it up, didn't know there were binaries for Linpack now... didn't fancy compiling from source in the past. This sounds like fun to run  Is it any worse than Prime95?...

6 hours ago, MageTank said:

Yup. It's why I use varying FFT lengths of Prime95 when stressing my system. 48k for stressing thermal solutions, 1344-2688 for vcore (constant load fluctuation), 512-4096k for memory and I/O lanes, etc.

I always wondered why those specific values, not some other either side of it. 8x the FFT size is the amount of ram the working data takes. 48k = 384k so would fit easily in L3 cache of most processors. Larger ones will take more than can fit on cache, and stress ram/IMC/etc. but I've not looked at if specific values are better or worse in some ways than others. Actually, that would be an interesting test I can do some time...

[er_gaming01]

8 hours ago, aisle9 said:

OP, I feel your pain. I'm bleeding voltage into my 1700 on a Gigabyte AB350 Gaming-3 and can't even get the thing to survive 8 hours of RealBench at 3.6GHz. It's bad enough that I picked up a Strix B350-F Gaming board with better VRMs and better cooling, not to mention better RAM compatibility, in hopes of getting myself to 3.8GHz at a reasonable voltage without giving up my stock cooler. It's clearly not the best binned CPU in the world based upon its relative lack of trustworthiness at 3.7GHz pulling down 1.33v, about as high as the stock cooler comfortably goes for me, but the laughably poor design of the Gigabyte board for overclocking (or, you know, using RAM that's on the QVL) could be a common thread with your MSI board.

Watch buildzoids video on all b350 vrm here 

he said that gigabytes vrm is horrible

[jjohnthedon1]

1 hour ago, porina said:

I was going to ask under what conditions you test at, as obviously they can make a huge difference.

 

Who cares about idle voltages? Either my systems are working hard, or they're off.

There it is. Weak loads. Well, Aida64 can start some mild heating if you uncheck CPU and only run FPU stress test, but I never found it really useful as a stress. As stated, my testing was done using Prime95 29.1 as that was the first version to add Ryzen support and use AVX. Older versions detected it as an older AMD CPU and used non-AVX code. Also I was using 64k FFT to focus the load on the CPU itself.

 

I'd be interested if you could try the same and see what temps you get, if it is even stable.

Ooh, had to look it up, didn't know there were binaries for Linpack now... didn't fancy compiling from source in the past. This sounds like fun to run  Is it any worse than Prime95?...

I always wondered why those specific values, not some other either side of it. 8x the FFT size is the amount of ram the working data takes. 48k = 384k so would fit easily in L3 cache of most processors. Larger ones will take more than can fit on cache, and stress ram/IMC/etc. but I've not looked at if specific values are better or worse in some ways than others. Actually, that would be an interesting test I can do some time...

Aida 64 and r15 on loop for 5 hours isn't high load ? Ok then ? 

I will never push my system harder than that for anything I do 

that test is completly overkill 

if it doesn't fail on that I do t see it failing on gaming or a light works render 

 

[porina]

8 minutes ago, jjohnthedon1 said:

Aida 64 and r15 on loop for 5 hours isn't high load ? Ok then ? 

I will never push my system harder than that for anything I do 

that test is completly overkill 

Overkill for you maybe, but Prime95 is representative of a typical workload for all my systems. It may explain our temperature differences. I'm ok knowing that I wont reach as high overclocks as others who don't demand as much.

[jjohnthedon1]

12 minutes ago, porina said:

Overkill for you maybe, but Prime95 is representative of a typical workload for all my systems. It may explain our temperature differences. I'm ok knowing that I wont reach as high overclocks as others who don't demand as much.

What do u possible do that's harder that all cores maxed rendering and Aida 64 at the same time ?

[porina]

9 minutes ago, jjohnthedon1 said:

What do u possible do that's harder that all cores maxed rendering and Aida 64 at the same time ?

I often run other prime number finding software (LLR, PFGW) which uses the same gwnum math library as Prime95, thus Prime95 is representative of that load. There is also software (genefer) that isn't based on the same library but still performs comparably.

 

Aida64 and cinebench I don't find that stressful. It's due to whatever instruction mix they happen to use, hence as said I find unchecking CPU and leaving FPU selected in aida64 can increase temps a bit more. Have you tried Asus Realbench as stress test? It sucks as a benchmark, but I find it a good 2nd choice for stability testing after Prime95 as it exercises other parts of the CPU, and can find instability way faster than aida64. I have to say I've never considered cinebench for stress testing. I'd like something better than crash/not crash, it doesn't report minor errors does it? At least aida64, realbench and Prime95 do that.

[jjohnthedon1]

6 minutes ago, porina said:

I often run other prime number finding software (LLR, PFGW) which uses the same gwnum math library as Prime95, thus Prime95 is representative of that load. There is also software (genefer) that isn't based on the same library but still performs comparably.

 

Aida64 and cinebench I don't find that stressful. It's due to whatever instruction mix they happen to use, hence as said I find unchecking CPU and leaving FPU selected in aida64 can increase temps a bit more. Have you tried Asus Realbench as stress test? It sucks as a benchmark, but I find it a good 2nd choice for stability testing after Prime95 as it exercises other parts of the CPU, and can find instability way faster than aida64. I have to say I've never considered cinebench for stress testing. I'd like something better than crash/not crash, it doesn't report minor errors does it? At least aida64, realbench and Prime95 do that.

And what do you need thies prime numbers for if u don't mind me asking ? 

I find realbench stress a little shit 

[porina]

1 minute ago, jjohnthedon1 said:

And what do you need thies prime numbers for if u don't mind me asking ? 

People have different interests and hobbies, this is one of mine. There is a list of top prime number finders (not tied to any specific project) and I'm trying to keep in the top 100 worldwide for score... the score is a blend of quantity and quality, taking into consideration finding bigger ones gets a lot harder than many smaller ones.

 

Finding some types of prime numbers can prove or disprove mathematical conjectures, which may then act as a building block to something else down the line. At most, I'll get a tiny footnote in mathematical history.

 

1 minute ago, jjohnthedon1 said:

I find realbench stress a little shit 

In what way?

[jjohnthedon1]

2 minutes ago, porina said:

People have different interests and hobbies, this is one of mine. There is a list of top prime number finders (not tied to any specific project) and I'm trying to keep in the top 100 worldwide for score... the score is a blend of quantity and quality, taking into consideration finding bigger ones gets a lot harder than many smaller ones.

 

Finding some types of prime numbers can prove or disprove mathematical conjectures, which may then act as a building block to something else down the line. At most, I'll get a tiny footnote in mathematical history.

 

In what way?

Because on my i7 system I could pass the stress for 24 hours then crash in game, fail burn test fail itx stress and fail Aida 64

[porina]

1 minute ago, jjohnthedon1 said:

Because on my i7 system I could pass the stress for 24 hours then crash in game, fail burn test fail itx stress and fail Aida 64

I don't trust any single stress test in isolation, but I did find realbench found errors for me that aida64 did not. Maybe I should have added, I don't run realbench as a stress test for more than an hour. 

[Simon771]

11 minutes ago, jjohnthedon1 said:

Aida 64 and r15 on loop for 5 hours isn't high load ? Ok then ? 

I will never push my system harder than that for anything I do 

that test is completly overkill 

if it doesn't fail on that I do t see it failing on gaming or a light works render 

As long as it's stable for you, it's fine.

When I'm testing stability, I do 2 hours of aida, 2 hours of prime95, few cinebench runs and some software for rendering (HandBrake).

Also if I don't get any crash after 1 week, I concider that stable  

[jjohnthedon1]

41 minutes ago, Simon771 said:

As long as it's stable for you, it's fine.

When I'm testing stability, I do 2 hours of aida, 2 hours of prime95, few cinebench runs and some software for rendering (HandBrake).

Also if I don't get any crash after 1 week, I concider that stable  

Fully stable ;p ?

[DildorTheDecent]

What a roller coaster this was. 

 

Not surprised that it's an MSI board having trouble though  

[WkdPaul]

Thread somewhat cleaned.

 

I'd like to remind everyone that this is about the OP, not yourselves, if you can't or won't help, there's no need to reply or argue.

 

Keep the Community Standards in mind when interacting with others.

[er_gaming01]

20 hours ago, ErrantNyles said:

LLC should be set to 1, SOC to 1.2V, DRAM to 1,35-1,5V and kept at 2400Mhz max. For CPU voltage, try 1,375.

 

If it doesn't give you any benefit, you simply got a """"bad"""" chip.

I have my ram at 3466 and cpu at 3.9Ghz with 1.425v. Haven't tried any lower voltage. The SOC 1.2v and LLC 1 helped ALLOT! Thanks! 

[er_gaming01]

20 hours ago, jdwii said:

Lot of things can come into play for one his VRM temps could be jumping above 90C and the board might be shutting off as the B350 boards don't have the best VRM. Two if he is on the stock cooler pushing anything higher then 1.25V is going to push temps. 

This also helped me with a stable OC. The vrm of my mobo sucks. goes to 120C with a fan pointed at it with prime95 1.425v

[done12many2]

4 minutes ago, er_gaming01 said:

I have my ram at 3466 and cpu at 3.9Ghz with 1.425v. Haven't tried any lower voltage. The SOC 1.2v and LLC 1 helped ALLOT! Thanks! 

 

Looks like cranking up your Vcore helped a lot too.  

[3rrant]

5 minutes ago, er_gaming01 said:

I have my ram at 3466 and cpu at 3.9Ghz with 1.425v. Haven't tried any lower voltage. The SOC 1.2v and LLC 1 helped ALLOT! Thanks! 

@MageTank

 

Now what? Happy to help.

[done12many2]

Just now, ErrantNyles said:

@MageTank

 

Now what? Happy to help.

 

Please note that he increased VCore substantially before you gloat too much.  Wow.  

[Guest]

Just now, er_gaming01 said:

This also helped me with a stable OC. The vrm of my mobo sucks. goes to 120C with a fan pointed at it with prime95 1.425v

Good glad to hear you really want to be safe man

 

Not sure why people made a big deal about 1.2V soc that is a fine voltage for SOC just don't go higher. 

 

I'd personally try and see if you can go a little lower on CPU core voltage for the sake of the CPU and board VRM. 

 

This is my setup at the moment

 

https://postimg.org/image/coovfb5of/

 

Try 1.35V 3.9Ghz see if that passes if not 3850mhz then 3825 if that fails. Should be able to get at least 3.8Ghz at 1.35V my chip can do 3.825mhz at 1.3V for example. 

 

But run prime95 blend for 2-4 hours. See if it passes. Can speed things up by using Ryzen master then putting in the final OC in the bios. Your VRM will go down probably 20-25C once you set it at 1.35V

[MageTank]

47 minutes ago, ErrantNyles said:

@MageTank

 

Now what? Happy to help.

He just mentioned overclocking his ram alongside raising that SoC voltage, rofl. I've said it countless times already. SoC helps with stabilizing ram, not the core. Until you offer any evidence showing SoC stabilizing the core, I am left to believe the countless sources I provided showing it's impact on ram stability, not core. 

 

9 hours ago, porina said:

I was going to ask under what conditions you test at, as obviously they can make a huge difference.

 

Who cares about idle voltages? Either my systems are working hard, or they're off.

There it is. Weak loads. Well, Aida64 can start some mild heating if you uncheck CPU and only run FPU stress test, but I never found it really useful as a stress. As stated, my testing was done using Prime95 29.1 as that was the first version to add Ryzen support and use AVX. Older versions detected it as an older AMD CPU and used non-AVX code. Also I was using 64k FFT to focus the load on the CPU itself.

 

I'd be interested if you could try the same and see what temps you get, if it is even stable.

Ooh, had to look it up, didn't know there were binaries for Linpack now... didn't fancy compiling from source in the past. This sounds like fun to run  Is it any worse than Prime95?...

I always wondered why those specific values, not some other either side of it. 8x the FFT size is the amount of ram the working data takes. 48k = 384k so would fit easily in L3 cache of most processors. Larger ones will take more than can fit on cache, and stress ram/IMC/etc. but I've not looked at if specific values are better or worse in some ways than others. Actually, that would be an interesting test I can do some time...

Linpack MKL is magnitudes worse than Prime95. 48k FFT's are hot, because it fits completely within cache (AVX scales with bandwidth, and cache bandwidth is worlds faster than even the fastest ram) but MKL uses both cache and your ram. Basically, your entire CPU subsystem gets stressed. We are talking an 8C difference on my system between 48k FFT and MKL. 

[done12many2]

Just now, MageTank said:

Linpack MKL is magnitudes worse than Prime95. 48k FFT's are hot, because it fits completely within cache (AVX scales with bandwidth, and cache bandwidth is worlds faster than even the fastest ram) but MKL uses both cache and your ram. Basically, your entire CPU subsystem gets stressed. We are talking an 8C difference on my system between 48k FFT and MKL. 

 

Was given to man by the Devil himself.  

[MageTank]

Just now, done12many2 said:

 

Was given to man by the Devil himself.  

You have to admit though. No other program can point out cache instability quite like Linpack MKL, lol. 

[done12many2]

2 minutes ago, MageTank said:

You have to admit though. No other program can point out cache instability quite like Linpack MKL, lol. 

 

Nope, even with CPU cache only slightly underpowered (lack of Vcache)  the PC either completely shuts off or blue screens instantly.  No real need in wasting time there.  

[Simon771]

6 hours ago, jjohnthedon1 said:

Fully stable ;p ?

Well yeah for me it's fully stable at 3,8GHz.

But since it's summer I had to turn it down to 3,7GHz.