Zen SuperPI 32M 5G/4G Challenge

[Max-Hardware_Numb3rs]

Check this:

https://hwbot.org/submission/4194748_

At 2993 1:1 I have CB at -142°C, no CBB, I can boot as well at almost -160 with the motherboard at default, so something strange is going on.

From today’s session I saw that timings as well cripple the fabric stability and CB.

I don’t know if this bios there’s a way to go with low timings without risking CB or CBB.

plus Turbo V doesn’t work, so I have to do all via bios and the frequency with the new software. 

I think I wait for the CH8 and better bioses, like this is pretty much a waste of LN2 

 

[KaRtA]

Ah good, I am not the only one with TurboVcore issues then. Makes it a massive PITA to reboot each time. The C5 and reset bios after failed training is an massive kick is the ass too.

 

This C6H is pretty rubbish on memory too, not too happy over 3400 on air, at least 3500C10 is possible on Cold with the old 2700X

 

Will test my 3900X chip on SS soon and see what -50 to -60 is like first before ordering a fill and wasting my money. It's not great but good enough.

[cbjaust]

yeah, Bus clock control went missing in action from TurboV Core for me. I don't know what BIOS settings did it but it did reappear in the course of testing. There are a few voltages that are 0.000 as well. I have the R5 3600X with G.Skill F4-3200C14D-16GTZ on the C6H BIOS 7106. My memory clocks have improved over my R7 1800X so that's a plus. I havent tried to run/install Windows 7 yet but Windows 10 seems to be terrible at SuperPi even compared to 7.

[flanker]

Win7 working on X570: 

 

[unityofsaints]

My humble contribution:

I don't think Ryzen 3rd gen will do too well in 32M so I've gone back to 2nd gen for now. Still a lot to test of course.

Edited July 25, 2019 by unityofsaints

[Max-Hardware_Numb3rs]

2 hours ago, unityofsaints said:

My humble contribution:

I don't think Ryzen 3rd gen will do too well in 32M so I've gone back to 2nd gen for now. Still a lot to test of course.

Zen 2 doesn't really shine in Spi32M, mostly for the maximum frequency not far from Zen+ and the CB related to the fabric that doesn't allow to push high memory frequency in 1:1, there's a big compromise to be taken, higher clock (with lower temps) and low fabric clock Vs higher fabric clock and lower CPU clocks, I'm still trying to find the best settings, but with a lucky 3800X or 3900X probably there's a chance to beat Zen+.

For your Zen+ run, I can suggest you to try C11.10 with memory at 3600/3700 (if your IMC can do it), the benefit of this combo will give you more efficiency at higher CPU clocks 

[I.nfraR.ed]

Yes, I think so, too.

On Zen 1st and 2nd gen, higher memory clock should be better on ambient, since the fclk is tied to it. 3600+ with CAS 11 or 12 should be faster.

My 1800X could do even 3800C12, but as soon as I cool it down (even on chiller) the max mem clock goes down, as you already know.

So maybe he's pretesting clocks which should work on LN2.

[Max-Hardware_Numb3rs]

2 minutes ago, I.nfraR.ed said:

Yes, I think so, too.

On Zen 1st and 2nd gen, higher memory clock should be better on ambient, since the fclk is tied to it. 3600+ with CAS 11 or 12 should be faster.

My 1800X could do even 3800C12, but as soon as I cool it down (even on chiller) the max mem clock goes down, as you already know.

So maybe he's pretesting clocks which should work on LN2.

I didn't find any particular issue with cold, in fact I usually pretest the memories on air to the limit, knowing that at full pot I can push a bit more on max memory frequency, my boot settings are 4GHz CPU - 3600 C11.10 (2,04V RAM - 1,265V SOC)  and then I raise the clock in Windows, usually from 1800 (x2) to 1860/1870.

This is the freq/timings that gave me the best efficiency at 5,5GH and above.

Another important trick is to check the performance at every boot, because the motherboard train again the memories, and it brings changes to timings not visible to us,  at 4GHz I should have around 109 with the smallest run, if higher (sometimes even 169) you must reboot and try again. 

[unityofsaints]

13 hours ago, Hardware_Numb3rs said:

For your Zen+ run, I can suggest you to try C11.10 with memory at 3600/3700 (if your IMC can do it), the benefit of this combo will give you more efficiency at higher CPU clocks 

Yes that's the target for LN2 but IMC won't do 3700 C11 on air, no way. Need that cold scaling.

12 hours ago, I.nfraR.ed said:

My 1800X could do even 3800C12, but as soon as I cool it down (even on chiller) the max mem clock goes down, as you already know.

This is a known issue with Ryzen 1000: the IMC gets worse with cold. Here is Elmor's graph (from overclocking.guide) :

Ryzen 2000 no longer has this issue, often even getting a bit better with cold - for example 4000C12 dual channel is much easier cold, on any chip. On ambient I can't consistently do that.

12 hours ago, Hardware_Numb3rs said:

Another important trick is to check the performance at every boot, because the motherboard train again the memories, and it brings changes to timings not visible to us,  at 4GHz I should have around 109 with the smallest run, if higher (sometimes even 169) you must reboot and try again. 

Yes I always get 109 for 16K. But I have focused on timings so far, still need to optimise O.S. and waza.

Edited July 25, 2019 by unityofsaints

[I.nfraR.ed]

Yes, I know about 1st gen, but was unaware that 2nd gen doesn't have this problem.

[Guest]

I'm wondering how there were at least 4 people that were around 8m20s, even some at sub 8m17s with 3200, on 1st Gen Ryzen and whatever methods they used have not been discovered by someone else yet.

Even if AMD and 32m are a niche combination it doesn't make sense that someone else doesn't know how to make them work.

[Max-Hardware_Numb3rs]

7 minutes ago, AutisticChris said:

I'm wondering how there were at least 4 people that were around 8m20s, even some at sub 8m17s with 3200, on 1st Gen Ryzen and whatever methods they used have not been discovered by someone else yet.

Even if AMD and 32m are a niche combination it doesn't make sense that someone else doesn't know how to make them work.

I tried literally everything, various waza size and types of files, cache settings, Windows tweaks, nothing to reach that level, close but they have discovered a good tweak that we can’t figure it out, one of my main goal is still to be the first sub 6 min on AMD, let’s see if I can with the 3800X

[flanker]

Im sure, for record we need Windows 7 OS? not Windows 10. Im feeling Win10 "much" slower in Superpi than Win7 on Ryzen.

[Max-Hardware_Numb3rs]

1 hour ago, flanker said:

Im sure, for record we need Windows 7 OS? not Windows 10. Im feeling Win10 "much" slower in Superpi than Win7 on Ryzen.

Win 8/10 is like 1 minute slower, absolutely non sense to use with spi 

[Alpi]

NAh, to test the new bios, I did some spi32 runs. 7301 works flawless ! No more C5 just normal boot or hang.  

Os : W7 64, try to turn on, off smt, waza but I guess lucky run is what You need.   Infrared's 8:17 is crazy clear run ! Can't go closer, always had some slower loops and Infra passes. Nice run, grats ! Simply can't hide You are genius ! ?

 

[Alpi]

Another fast test. That setup on W10. Crying and crying.... :DDDD Cpu clocks were 200+ mhz higher even...

https://hwbot.org/submission/4204154_alpi_superpi___32m_with_benchmate_ryzen_5_3600_8min_45sec_207ms?recalculate=true

[flanker]

yes, Win7 as I see is much better choice for 32M

[I.nfraR.ed]

First test with XP ACPI. Seems faster, but gradually looses time after loop 10-12 and couldn't beat my previous score with 1800X.

Will pop it in later and compare with my old 8.17.xxx score.

Timings and other bios options not optimized for sure.

Also working on SuperPi perf bias which looked promising from the first test.

Edited August 6, 2019 by I.nfraR.ed

[Max-Hardware_Numb3rs]

Xp and a well tweaked Win7 are more or less the same at the end, with Zen2 I’m in the 8.20 area as well with 1900 flck (Win7)

The problem is that is unusable with LN2, I’m working now to find an efficient 2800/2933 CL9 ram setup, or below -70/80° it suffer from CB

[I.nfraR.ed]

I think XP will be faster in the end (not by much, though), but it took me several hours to get it to this state on the Crosshair VI Hero. ACPI and USB 3.2 work.

Perhaps we need XP64 for this generation of CPUs. Anyway, will still test it and see if 1800X is faster.

Maybe the low write speeds of the 3600 is hindering the performance.

The difference between 3600 and 3800 memory/if clock isn't that great and 3600 tighter timings might get it even closer.

 

PS: Wouldn't 4000-4133 CAS12 be a better choice with IF at 1400-1500 (whatever is possible on LN2)? 4000C12 tight was damn easy, haven't tried higher.

Edited August 6, 2019 by I.nfraR.ed

[I.nfraR.ed]

Some more

https://hwbot.org/submission/4211718_i.nfrar.ed_superpi___32m_ryzen_5_3600_7min_26sec_359ms

 

Guess I will have to try Win7 again. XP doesn't seem that fast.

Edited August 7, 2019 by I.nfraR.ed

[I.nfraR.ed]

On 7/27/2019 at 2:17 PM, Alpi said:

NAh, to test the new bios, I did some spi32 runs. 7301 works flawless ! No more C5 just normal boot or hang.  

Os : W7 64, try to turn on, off smt, waza but I guess lucky run is what You need.   Infrared's 8:17 is crazy clear run ! Can't go closer, always had some slower loops and Infra passes. Nice run, grats ! Simply can't hide You are genius ! ?

 

You did a good job! I'm currently on my 5-th Windows 7 installation and sitting at 8:24.xxx.

Whatever I "tweak" it makes it slower and slower. Exactly the same situation before 2 years with Zen1, but I don't remember how I made it faster for this 8:17.xxx run.

Should have saved the OS image or write down all the changes :(. I must have found something back then which I don't remember.

PS: First few loops of "untweaked" 7 are faster than XP (even faster than what you have on that SS), but then it slows down.

Then I start "tweaking" and even these first loops get slower.

Edited August 9, 2019 by I.nfraR.ed

[KaRtA]

If you ever find the key Infrared, I’d love to hear it. 

 

My current tests have 3620c11-1t around 8m21s with Waza in W7 x64 on C6H/3900X

 

testing 2:1 ratios now ready for cold but the performance hit is big. Same 4g but 4000C12-2t/1000FCLK is at 8m34s. Huge difference . I also did a few tests at 4000C12-12-1T but the improvements over 2T weren't really there.

 

 

Edited August 11, 2019 by KaRtA

[KaRtA]

Did get 3800C12-12-1T with 1900FCLK running also.

Nice and Tight, seems to be pulling ahead now against my 2700X efficiency. Including a 4.6G run for reference on scaling.

 

At 4.2ghz its a good 12 seconds faster.

 

2933C9/1467 Performance wasn't too bad either, 4 seconds faster than my 2700X Efficiency at 4200.

[Bullant]

Nice work Karta, sub 6 on the way ?