Project HYDRA for Ryzen Zen3, 3+ and Zen 4

Seen a couple of posts on this; but no discussion since the ClockTuner thread (CTR since decommissioned for Zen3+).

I ran a quick diagnostic last night (took hours, let it go overnight), have just checked the logs and came back with this:



Reading the developers notes, appears he's full steam ahead with Zen4 support and RDNA2/3 overclocking (as well as potentially nVidia frequency/voltage support), all whilst moving between bunkers in the Ukraine, nuts.

Anyone playing with HYDRA these days?

 

Been playing with this last few days, really do like it. Some thoughts here to avoid double post.

Love how HYDRA shows real-time which core(s) are preventing the entire CCD from boosting higher; and allowing you to change in real-time easily.



On the fly PBO2 curve changes really nice too, much better than Ryzen Master imho.



Interesting to see where this goes with Zen 4 and RDNA3 curve overclocking.

 

The Beast here you go. Only change

Factory Default PBO (with 200Mhz Boost Override):


Settings for running CO tests in HYDRA:

- Set everything to auto (i.e voltage, boost, LLC etc)
- You can use any PPT/EDC/TDC limits you like (within reason/based on what you're targeting)

Important prior to running Hydra tests:
- Set Boost Override to 0
- Curve Offset (CO) on all cores to 0

The last 2 are particularly important as you dont want AMD's PBO running/kicking in whilst Hydra is trying to test/find/optimise each core's CO.

Once I got a set of stable CO's, you can then either set max frequency for each core in Hydra, or just go back and set a Boost Override in bios (max 200Mhz for each core) which will boost on top of your Hydra profiles.

I did the latter (in BIOS), obviously once I settled on max frequencies for each core.

If you do it in Hydra, looks like this:



I could only get Cores 05 & 06 stable at 5200Mhz (this chip's best cores), with the others to boost up to 5150. Core 10 wouldn't go past 5050 stable.

 

Much appreciated, now to find the time!

 

Haven't had much luck with the free version in the past but I have downloaded 1.2D Pro for a play and will let you know how it goes on my 5800X.
The old version of Hydra would set the first core to 4950Mhz and the second to 4825MHz which is below the stock 4850MHz.
This version is currently testing the first core at 5000MHz.

Been reading the Hydra big tutorial.
This is a feature that my CPU badly needs with the second core being considered the third best core when it is by far the worst but it seems the feature is still a work in progress.

But other features that are now working will be very useful to fix problems like the crashes I was getting in light multithreaded loads like Aida64 cache benchmark and games (despite being prime stable) due to the limitations of PBO\offset overclocking being problematic with my CPU.

Something seems to have gone wrong.
1T-2T profile seems to have a lower V than 3-11T profiles and cores that were testing ok at 5000MHz are now set to 4875MHz Max.
Will change some settings and run another diagnostic.


Edit: So this is what diagnostic two looks like while running Cinebench single core with hybrid mode enabled and it is running well below stock speeds.
I guess I can enter values manually but that will be fun to try and get stable.

 

Cranking up the PBO power limits in Hydra when creating the profile then limiting them with PBO2 seems to be giving me better results.

HYDRA BOOST TESTER RESULTS
CORE | AVG/MAX FREQ | POWER | TEMP | VID |
C01 4866 / 4875 11.78w 56c 1.375v
C02 4800 / 4800 11.9w 55.4c 1.375v
C03 4948 / 4950 11.78w 55.6c 1.375v
C04 4848 / 4875 11.63w 55.2c 1.375v
C05 4899 / 4900 11.44w 53.3c 1.375v
C06 4896 / 4900 11.46w 55.3c 1.375v
C07 4874 / 4900 12.71w 58.2c 1.375v
C08 4874 / 4875 12.1w 57.4c 1.375v



Best core overide seems to be working rather nicely. Although not with all programs, 3dmark CPU test for example still just works at the original core order. which was
1-8-2-6-3-5-7-4
Now set
3-6-8-5-1

 

Yeh I also found increasing limit worked well with Hydra. With lowet limits, diagnostic results were a bit odd.

 

Crashed a few times and the auto adjustments got to the point where it was running 200MHz slower than stock for 1-8 thread loads.
So I am giving up on project hydra again.

 

^^ My experience so far (and many others on the community discord) is its not at the point where the auto diagnostic test + follow up auto adjustments optimises all (if any) systems to be stable; even with "safe" settings.

I found still requires manual tuning of the cores after auto tests. Albeit it gets pretty close.

 

Any suggestions for what manually tuning I should do?
Load tests show it to be fairly stable, it is light work like watching a video, loading a game or running aida64 cache benchmark that makes it crash. This isn't new to Hydra the 5800X has always behaved this way when overclocking with curves\boost.

 

You need to go into event viewer and check the WHEA errors to see which cores are unstable.

There is a way to create a custom view to check for these errors - I don't have the instructions on hand. But basically this will aid in fine tuning the unstable cores.

 

Event Viewer (Local)" expand "Applications and Services", then "Microsoft", then "Windows", then "Kernel-WHEA", and then finally "Errors". In the top center frame of the window, you should see your WHEA Errors.

Trouble is there is no WHEA errors from this year they are all dated last year when I was overclocking the RAM.
When it crashes it just hard locks. Once I even had to turn it of at the PSU as holding the power button in wouldn't power it off.
Then the error log just shows that the system lost power.

Edit: wait just remembered I disabled the WHEA logger to stop the performance hit from running over 3733 RAM. I will try reverse that and see what happens.

 

how would this go with my 5900x

 

Sorry just seen your reply. If light work is crashing then obviously an issue with the top profiles (1-2T, 3T-4T) profiles. I won't discount an issue with your 5T-11T profile either in the event its being triggered (stats column) but I'd start with 1T-2T and then 3T-4T profile.

Plenty of ways you can go about doing this. The quick way is simply reduce clock speeds by 50Mhz on each profile and see how you fare. Then increment in 25Mhz once you find stability.

Looking at your profiles in your screenshot:
Profile - VID - CCD1
1T-2T - 1.374v - 4675
3T-4T - 1.400v - 4700

It doesnt make sense that the lightly threaded profile is at 4675 and the following 3T-4T profile is at 4700.

Lightly threaded should be higher/boost higher.

-----

The correct way as above is to use WHEA errors.

But an even simpler way is to use the HYDRA LT/MT table by looking at how each cores behaves under load.

When a core's CO lights up red that CO is working. BUT more importantly, if the same core repeatedly lights up red whilst another is in use, HYDRA is telling you that core is preventing the entire CCD package from boosting higher.

I did a video a while back here testing each individual core (Cinebench single thread looping through cores. You can see which core is active/under load by the orange highlight along the top):



Not sure what happened to audio (sorry!) but you can see where I "circle" my mouse around certain cores' CO - I'm referencing:
a) that core is preventing other cores from boosting higher and so
b) that core requires CO adjustment (+ve or -ve)

Examples here:
00:37 - 00:42 - C02 lighting up <-- increased CO value/running less aggressive
01:30 - 01:37 - C01 lighting up <-- C06 was also flashing red but best to change 1 core's CO at a time, not multiple
02:03 - 02:15 - C11 lighting up <-- this one very clear, as the 11th Core was lit up the whole time as the first CCD (cores 1-6 were looping)

That last example of Core 11 is interesting as it shows a core on another CCD can also affect cores on another CCD (for those that have multiple CCD cpu's i.e 5900x/5950x etc)

and at 03:12 - C06 also stays lit whilst other cores are looped/stressed <-- so I would move C06 down a little as the +ve offset of 25 prob bit high there.

Some notes which may be helpful to others:
- These CO numbers are virtual numbers, not the same as CO values in your motherboard bios of -30 to 30.
- I step up/down in increments of 10, dont care for finer/tiny odd numbers
- I step up/down until the core CO obviously stops flashing red. You want a consistent flashing/lit up red only for the core in use.

Hope that's of some help and apologies again about lack of audio!

 

Thanks very much for the detailed reply.

I just gave it another run then but forgot to enable the hybrid OC and it crashed the moment I launched steam.
So it may just be a bug with the software rather than a unstable OC, maybe a conflict with my fan\pump control software Agrus monitor (which I close while doing diagnostics) not sure yet.

As for enabling the WHEA errors I cannot find how I disabled them on windows 11 so I have no idea how to enable them again.

What effect does this number have anyway? it still increases the clock speed well beyond what it is set to up to the cores max boost.

 

Its important as it sets the speed of the entire CCD for that profile. Boost (PBO) is then applied on top of this up until the max clock set in your Auxillary table; from your screenshot here:



My first thoughts on your screenshots above:
- CO values are very aggressive (too negative). Its not entirely impossible to have quite large negative offsets if you've got a great chip, but its impossible to say without running through individual cores.
- Start by reducing CO on your best cores (C01 & C08). I found on both my 5600x & 5900x best cores will only take a small undervolt /CO compared to other cores, makes sense if said best cores are running more efficiently/used most often by Windows thread scheduler.

Again these are virtual numbers different to mobo bios CO values; and like you also found the auto diagnostic non stable & had to drastically reduce CO numbers. Obviously this is assuming you're prioritising CO first then PBO.

Not sure how you disabled WHEA logging, but does creating another Custom View in Event Viewer work..?

 

Hydra Pro 1.3A is out for Patreon members.

Yuri Bubliy | creating software for Ryzen community | Patreon

DECEMBER 24 AT 11:18 AM
HYDRA 1.3A PRO - RELEASED!

The Christmas renewal is ready!

I sincerely wish you and your families warmth in your homes and love in your hearts. Merry Christmas!

One more thing. This is not the last update this year. More details in a few days.

--------------------------------------------------------------

CPU PROFILE COMPATIBILITY WITH THE PREVIOUS VERSION?

- NO

ONLINE-GUIDE (big update soon) : https://docs.google.com/presentation/d/1_j6vB61koOVwEHLrvqt7lh1RYw4XRuoTIeEDtYmnFj0/edit?usp=sharing

CHANGELOG:

* Added SAFE profiles for DDR5 memory (Hynix M-die).
* Added SAFE profiles for DDR4 memory (Micron E-die, Samsung B-die).
* Profiles for DDR5/DDR4 can be applied in 1 click. The user can also set any value he wants in the preset field. Note that the DRAM settings in the BIOS must be in Auto mode, otherwise HYDRA will not be able to set the new values.
* The button to apply DRAM values becomes active if the user has entered his values.
* Improved DDR5 support.
* Added RMP undervolting profiles.
* Improved performance of RMP profiles for RX 6800 / 6800XT / 6900XT / 6950XT.
* Fixed bug that crashed the system with RX 6900 XT after applying RMP profile.
* Significant changes have been made to the AMD Curve Optimizer value search. The value check has several phases.
* Simplified interface for HYDRA CPU DIAGNOSTIC CENTER.
* Added test package for DRAM - MEMbench. It allows you to. Evaluate performance after optimization of timings, latency with random access and read/write speeds. Note that you will only get results if the memory is stable, otherwise the test will stop.
* Fixed a bug for Cezanne processors that prevented getting monitoring data.
* Improved stability of HYDRA driver loading during application startup. Now there will be no empty interface and CPU frequency drops to 550MHz when the driver starts incorrectly.
* Fixed bug when after enabling /disabling iGPU it was not possible to get monitoring data for dGPU.
* Fixed a bug with an inconsistent indication of the diagnostic step number.
* Fixed a bug with non-corrected CO reset to 0.
* PROFILES CREATION correctly fills in the values of maximum permissible frequencies (Zen 3).
* Correcting the start values for the processor diagnostics.
* Improved detection of 7600X / 7800X CPUs that have two CCDs.
* Fixed a bug where the system having an active shutdown function after diagnostics shut down if the user pressed the STOP button.
* Changed frequency limiter behavior for HYBRID OC. In some cases, the user will get a higher operating frequency.
* Removed delay during user interruption of PROFILE CREATION.
* Other minor bugs fixed.

Also the big tutorial has finally been updated: HYDRA BIG TUTORIAL - Google Slides

YESTERDAY AT 10:28 AM

HYDRA BIG TUTORIAL - UPDATE READY!
Glad to inform you that the guide has been updated!

Also sections about RMP and presets for DRAM have been added to it.

https://docs.google.com/presentation/d/1_j6vB61koOVwEHLrvqt7lh1RYw4XRuoTIeEDtYmnFj0/edit?usp=sharing

P.s. this is not the final version and I will keep adding to it.

 

First run:







Not much here for me with the memory tuning - my B.Die XMP is already better than most. Interesting it only offers an Optimised Preset but not a Fast Preset.



PBO2 Tuned Comparison



Hybrid Tuned Comparison



rg144 I've not much idea where to take this from here. With a watercooled CPU (and motherboard VRMs) I would think I could start with more aggressive EDC, PPT, TDC and THM limits?

 

Does this one tell you if you have a good quality CPU or not?

Doesn't seem to work very well with my 7600x
https://i.imgur.com/UDILHhP.png
gets to there, and then pretty much does nothing except make the system really slow

 

Most definately. I cranked mine all the way up and just started testing extreme limits of each CID, then multi core package, then setting profiles appropriately.

Those HYBRID tuned results not too shabby at all efficiency wise, pretty low voltage you got there

I've dropped off the Patreon so only got to v1.2. Auto generated profiles were much more stable than in v1.1 though. How have you found the hybrid profiles/overall stability so far Beast?