Zen SuperPI 32M 5G/4G Challenge

[KaRtA]

1 hour ago, Hardware_Numb3rs said:

I’ll post some screen later but so far I was able to get like 6:15.xxx at 5450MHz with flck at 1800 and ram at 3600 C11 

At -96°C (cascade)

The importance of the flck is clear, maybe a sub 6 is achievable only with 5600/5650 if I can maintain 1800 flck around -140 

next recharge I’ll try ?

FYI. When I started my session the other day, I was having some strange BSOD happening (was a voltage in the end) but tested 1566/3133C10 setting at 5.5g and was at 6m15.9s. 
 

weirdly, I was checking my efficiency at -20 when I started and found that I had lost 0.6s to 1s at 4.5G from the day before. Spent maybe an hour with warming and trying different ways to retrain and couldn’t recoup that missing time. Moon phase or day can really make a difference. 
 

I had a 3600 profile set to test too but LN2 was limited. Also when trying for 1667FCLK moisture under the CPU came into play and ended the day for me, I was warming to -60 just to boot by that stage which was fraught with 00 and 07 code reboots. If I had another 50L on hand I would setup again and retry. 

[Max-Hardware_Numb3rs]

45 minutes ago, Sparky's__Adventure said:

@Hardware_Numb3rs Does having both CCDs enabled mess with efficiency?

Don’t think so, like Karta wrote SMT does, but regarding the CCDs, you hit way higher clock with only one, and if your fastest core is in the first two blocks of the remaining CCXs (2+2) is even better, since right after the flck also the CPU frequency matter, and I’m my case, the second CCD is pretty garbage ?

[_mat_]

@KaRtA You used BenchMate for your sub 6 run, I guess to make use of the Virtual RAMDisk tweak, right? Why didn't you do the screenshot with the result dialog as well? Would have been much more trustworthy to know that SuperPi wasn't tampered with and that timers were not skewed.

[KaRtA]

Hey _Mat_, I was waiting for someone to notice that. I honestly only used it for the ramdisk. I tested with LPM I think it was called in there but it made no notable difference in efficiency. 
 

Also, the previous session when I was cold and on the limit, I kept BSOD when opening everything so the less I had to do, the better. I don’t believe it was a program issue, just not all the cores on the cpu were happy with load at 5750mhz. Even this one I changed to 32bit CPUz to make the screenshots quicker and less load. 
 

I found that I could achieve 2 second more efficiency than my usual technique by opening in benchmate, then closing the main program, run my Waza and start my run. Every run was within 0.2s consecutively this way.  For example, at 4.2ghz, with my usual Waza run, I would be 7m58.6s, but using benchmate with real-time, cpu affinity to my select core (core4) and ramdisk only, would be 7m56.3s at its best (with Waza of course)
 

If I knew how to setup ramdisk separately I would try that. If anyone can explain how, I’d be happy to check the efficiency that way too while I still have the CPU. 
 

im not sure that would be considered as altered at all, but happy to share any of my findings or technique for full transparency. Can even image my OS if you would like to check from your side. 

Edited January 11, 2020 by KaRtA

[Max-Hardware_Numb3rs]

10 hours ago, KaRtA said:

FYI. When I started my session the other day, I was having some strange BSOD happening (was a voltage in the end) but tested 1566/3133C10 setting at 5.5g and was at 6m15.9s. 
 

weirdly, I was checking my efficiency at -20 when I started and found that I had lost 0.6s to 1s at 4.5G from the day before. Spent maybe an hour with warming and trying different ways to retrain and couldn’t recoup that missing time. Moon phase or day can really make a difference. 
 

I had a 3600 profile set to test too but LN2 was limited. Also when trying for 1667FCLK moisture under the CPU came into play and ended the day for me, I was warming to -60 just to boot by that stage which was fraught with 00 and 07 code reboots. If I had another 50L on hand I would setup again and retry. 

I saw that the temp alone matters for the flck stability, and therefore efficiency, I think there are still a lot to uncover with Zen2 to get the maximum performance, testing at ambient is pretty much useless since the paradigm change completely with cold.

We can't say it's a boring architecture

I will keep study with the cascade for some weeks before going LN2 again.

[KaRtA]

I tested every memory setting I could at 3133, 3200,3333 and 3600 in preparation for cold. My 3133-3333 memory is all the same in the end, with no benefit in timing changes  

 

Still today I put it back on ambient to test more and yet another day where efficiency went down the toilet. It almost feels that if it’s working well, you have to go cold straight after or you lose out. 
 

One reboot can make a change that you cannot get back. If AMD allowed access to RTL or IOL it would be far more beneficial for us. 
 

not really sure why I tortured myself with it again today, but I guess that’s what 32m does to people. Time for a rest 

[Max-Hardware_Numb3rs]

1 hour ago, KaRtA said:

I tested every memory setting I could at 3133, 3200,3333 and 3600 in preparation for cold. My 3133-3333 memory is all the same in the end, with no benefit in timing changes  

 

Still today I put it back on ambient to test more and yet another day where efficiency went down the toilet. It almost feels that if it’s working well, you have to go cold straight after or you lose out. 
 

One reboot can make a change that you cannot get back. If AMD allowed access to RTL or IOL it would be far more beneficial for us. 
 

not really sure why I tortured myself with it again today, but I guess that’s what 32m does to people. Time for a rest 

I’ve found that let it retrain with cold helps a bit, especially when there’s a restart at or near fullpot, probably something change and sometimes is for good, since the stability after a cold restart usually improves. 
Bios have various features hidden, something to try would be find what trigger those features, like playing with blck and dividers, but for a bit I’ll take a break from this, I’ve like 30/40 cpu from socket 939 to sandy bridge to bin ?

[l0ud_sil3nc3]

On 1/12/2020 at 5:03 AM, Hardware_Numb3rs said:

I’ve found that let it retrain with cold helps a bit, especially when there’s a restart at or near fullpot, probably something change and sometimes is for good, since the stability after a cold restart usually improves. 
Bios have various features hidden, something to try would be find what trigger those features, like playing with blck and dividers, but for a bit I’ll take a break from this, I’ve like 30/40 cpu from socket 939 to sandy bridge to bin ?

Retraining while cold is critical because the margin is recalculated when this occurs. Should happen at least once and usually multiple times, especially if you're direct booting at high freq.

[l0ud_sil3nc3]

Popped in the 3800X last night to try some pi, since I've never pushed fabric super hard on this cpu, still maxed out at 1600 for full pot operation. This cpu is significantly worse than the 3950x I tried a while ago at least in terms of core and IMC.

5750 no waza

5775 no waza

5775 waza + maxmem

Not the greatest waza but still trying to sort that out as I have figured it out good waza for no maxmem but with maxmem it doesn't nearly shave as much. Try again with 3950X later.

[Splave]

[Max-Hardware_Numb3rs]

Welcome to the club ??

[flanker]

how many people now in the club? 3 only?

[cbjaust]

@flanker currently it's 2.

 

http://url.hwbot.org/36qRikD

[l0ud_sil3nc3]

13 hours ago, cbjaust said:

@flanker currently it's 2.

 

http://url.hwbot.org/36qRikD

Look four posts above your post

Just a placeholder for now, but still slightly under 6 min.

[cbjaust]

yep, cool, so like 2 official and 1 unofficial ?

[Alpi]

Anyone tryed 2x4 edie ever ? Made a fast test just for fun and it wasnt such bad. (Made with an R5 1600af).

Maybe it worth a try. 

[chew*]

On 1/24/2020 at 2:24 PM, l0ud_sil3nc3 said:

Popped in the 3800X last night to try some pi, since I've never pushed fabric super hard on this cpu, still maxed out at 1600 for full pot operation. This cpu is significantly worse than the 3950x I tried a while ago at least in terms of core and IMC.

5750 no waza

5775 no waza

5775 waza + maxmem

Not the greatest waza but still trying to sort that out as I have figured it out good waza for no maxmem but with maxmem it doesn't nearly shave as much. Try again with 3950X later.

be a good 2 secs faster if you dump the 3800x and use 3900x

[l0ud_sil3nc3]

On 1/30/2020 at 5:07 PM, chew* said:

be a good 2 secs faster if you dump the 3800x and use 3900x

Was just trying the 3800x to see how high fclk went but sadly it's not over 1600 for full pot operation

Ended up testing the 16c and thinking outside of the 1:1 box, along with trying a few new idea's from KaRtA (thanks Daryl!) and Zen (thanks Davis!) and ended up breaking sub 20's while binning some retail 4800's. The kit below will be for sale and runs these settings at 1.89v vdimm set in bios, with super tight secondaries and tertiaries.

After retail binning was done, I popped in my binned modules and used the last bit of LN2, and ended up in the 56's

https://hwbot.org/submission/4344329_l0ud_sil3nc3_superpi___32m_ryzen_9_3950x_5min_56sec_930ms/

There's some sand as well

 

Edited February 3, 2020 by l0ud_sil3nc3
linkz

[KaRtA]

Killed it man, congrats!!!!

[Splave]

15 hours ago, l0ud_sil3nc3 said:

There's some sand as well

 

 

[dumo]

Awesome clockin! Nice late batches of 4800

[Bullant]

Nice work loud ?

[Alex@ro]

Excellent for sure, contemplating in joining as well, looks fun

[l0ud_sil3nc3]

On 2/3/2020 at 2:04 PM, dumo said:

Awesome clockin! Nice late batches of 4800

 

13 hours ago, Bullant said:

Nice work loud ?

 

13 hours ago, Alex@ro said:

Excellent for sure, contemplating in joining as well, looks fun

[l0ud_sil3nc3]

On 2/3/2020 at 3:35 AM, l0ud_sil3nc3 said:

There's some sand as well