Mid-Life Crisis

[ArchStanton]

I've convinced myself that I do need to revisit my RAM and actually tune it to my system, rather than rely on the settings I plagiarized from J7.  In addition, I've got some QD's coming from Koolance itself.  I did some digging and found these:

 

 

But pricing and availability didn't seem better than what's available in the DIY PC enthusiast market currently, so I didn't bother researching whether chrome plated brass would function for our purposes or not (I'm guessing yes, but it would be prudent to be 100% sure).

 

Here are links to a brochure and the main website if anyone wishes to dig deeper:

Liquid Cooling Connectors for Thermal Management : CPC (cpcworldwide.com)

Non-Spill Liquid Cooling Connectors | CPC (cpcworldwide.com)

 

I think I will likely make more changes to my loop in the future.  Probably going LM on both GPU die and CPU IHS.  More robust cooling on GPU backplate (I do not anticipate active, just "more").  Planning to use QD's to make maintenance far easier and facilitate connection to a "chiller" (maybe on the "redneck" side of the spectrum of possibilities).  I know the gains might be minimal, but for squeezing out every possible drop of cooling performance, I think I will have one of my larger radiators between the CPU and GPU blocks (primarily for the small benefit it might provide when not using a "chiller" of any sort).  Aesthetics are going down the toilet, but I am okay with that now as the "benching bug" has bitten me and I am infected.

[ArchStanton]

In my continued RAM OC research, I came across some interesting settings posted by domdtxdissar in [Official] AMD Ryzen DDR4 24/7 Memory Stability Thread | Page 404 | Overclock.net (page 404, post 8075).  They looked like an improved version of what I had previously been using.  Of course, I still don't know WHY they work.  Hopefully, as I continue to study, things will clarify, and I can graduate from copy/paste to test/refine with some idea of what I'm doing.

 

 

L3 read/latency is still not quite what I was hoping for, but its better.  I'm unsure if the reason my L3 values tend to lag those of the "prosumers" that post the settings is related to the fact that they tend to run 4x8GB while I am running 2x16GB .  Perhaps I will read further and learn the answer.

 

 

Time to fire up HWiNFO64 and TestMem5 (x25 cycles 1usmus) and begin to confirm stability.  Hopefully HWiNFO64 doesn't bug out and I get accurate peak temperature readings.  I guess I should turn the logging function on so I can check and see if any high-ish values are momentary blips or true temperatures.  DRAM voltage is set to 1.560...

 

Edit:  If the DIMMS go over 45⁰C, I'm pretty sure I'll get errors, and I'll be hunting for improved RAM heatsinks.  If they approach 40⁰C, I'll probably still be looking for better heatsinks.  F@k it, I'll be getting better heatsinks.

 

Edit 2:. I need to revisit J7's comments here in my own build log regarding single rank/dual rank/quad rank.  Maybe some of that pertains to L3 read/latency speeds...

 

Edit 3:  No, putting my brain in gear, L3 shouldn't be tied to much in the way of DRAM other than fclk I guess...hmm.  So much to learn .

Edited April 1, 2022 by ArchStanton
additional rambling

[ArchStanton]

It passed TM5_1usmusx25.  Temps were fine.  I left it running OCCT this morning.  I'll check it when I get home.  However, none of the handful of synthetics I through at it this morning showed a nickel's worth of change in performance .

 

 

Some of the enthusiasts at OCN seem obcessed with GDM off, which this "tune" has.  I haven't thoroughly tested it yet, but scratching the surface, I see no benefit.  I may investigate trying to push fclk above 1900 in a "daily drivable" way, but based on what I've read so far, that seems largely luck of the draw and unlikely.

 

[ArchStanton]

Had a chance to read a handful more posts today.  After examining some CBR23 scores that were posted in conjunction with impressive L3 numbers, I have the impression that those "shiny" L3 read/latency values were obtained with rather aggressive static OC's.  I do now recall that I read somewhere that L3 speed is tied to Core clock in some fashion, but it had not occurred to me that AIDA64 Cache and Memory bench might utilize enough threads to benefit from a high static OC as opposed to the generally better performance offered by PBO with single/low thread count loads.  I will test this when time permits.

 

Edit:  Would anyone have a recommendation regarding DDR4 heatsinks?  My specific kit is F4-3600C14Q-64GVKA, though I currently only have 2 sticks in use.

Edited April 1, 2022 by ArchStanton
RAM Heatsinks

[ArchStanton]

OCCT checks good, but I think I will throw everything and the kitchen sink at this tune before trying to lower the voltages or up the fclk (I'm undecided on what direction I will turn next).  Maybe Membench inside Dram Calc for our next validation tool.

 

[ArchStanton]

In further reading, domdtxdissar says explicitly that his high clock speeds are directly responsible for his inflated L3 values, so that settles that.

 

Edit:  Per the advice of someone I trust, I will be s#!t-canning this particular RAM "tune" (due to long term concerns over the voltages in play).  Back to the daily driver settings I was using previously and making an effort to see if they can be improved upon for my system.

Edited April 2, 2022 by ArchStanton
Change of plans

[ArchStanton]

I hadn't attempted to boot anything higher than 1900fclk since before updating to the most recent AMD chipset drivers.  I had not been successful in my previous attempts.  Subsequent to J7's encouragement of Beregond in his build log, I decided I would give it another whorl.  I am too much of a newb to even to begin to speculate why I failed back then but succeeded now.  Maybe the new drivers, maybe more appropriate voltages, maybe other things, maybe some of all of it.  Following the suggestion in MemTestHelper/DDR4 OC Guide, I set primaries to 16-20-20-40 and tCWL to 16.  Everything else I set to Auto with the exception of power down mode, which I set to disabled.  As you more experienced folks would know to expect, memory read/write/copy went up, but memory latency took a big hit and L3 suffered terribly as well.  I didn't bother with any stability testing for now.  I'm still kinda in shock that it booted into Windows at fclk 2000.  It was stable enough to run CBR23, but scores were  naturally.

 

 

 

 

Did L3 performance tank because I am unstable without crashing (some sort of clock stretching in the cores?) or do DRAM timings affect L3 as well as memory performance?  I have no idea why the system is performing this way.  It will be interesting to find out.  I'm not sure I will put more work into "tuning" for fclk 2000 at the moment.  I think I might be better served to go back to the known stability/performance of fclk 1900 and really buckle down and learn how/if the previous settings can be improved.  So, after posting this I will reboot, load my "daily driver" BIOS profile, and go back to casting my net in the murky waters of the interwebs in the hopes of catching a few minnows of understanding . 

 

 

 

[ArchStanton]

After returning to my previous "daily" settings and running some TM5 to verify that everything snapped back into place, I now get a lot of this:

 

(Reference picture only, not my system)

 

After trying other suggested fixes, the individual who posted that picture ended up wiping and reinstalling Windows to resolve the issue.  I will try clearing CMOS first, but a "wipe" may be in the cards for me as well.  We'll see how it goes.

 

 

 

 

[ArchStanton]

Clearing CMOS did not resolve the issue.  Nor did "scf/scannow".  So, a' wiping we will go.

[ArchStanton]

I decided to try the "reset Windows" option available in the "settings" section of Windows 10 prior to doing a true wipe in the hope of confirming that I was experiencing a software rather than hardware issue.  The reset did eliminate the issue I was having in Testmem5 (50 cycles 1usmus_v3 stable), and I will proceed with a true wipe (I don't trust all the extra folders/junk the "reset" resulted in).  After the wipe and subsequent "pruning" of Windows, I'll research how to create an "image" (I think this is the correct term?) of my OS for faster recovery from the next time I corrupt some part of my OS.

 

I got my QD3's in this morning.  They are so "cute" compared to what I'm accustomed to working with.

 

Edited April 4, 2022 by ArchStanton
missing close parentheses

[Avacado]

2 hours ago, ArchStanton said:

I decided to try the "reset Windows" option available in the "settings" section of Windows 10 prior to doing a true wipe in the hope of confirming that I was experiencing a software rather than hardware issue.  The reset did eliminate the issue I was having in Testmem5 (50 cycles 1usmus_v3 stable), and I will proceed with a true wipe (I don't trust all the extra folders/junk the "reset" resulted in).  After the wipe and subsequent "pruning" of Windows, I'll research how to create an "image" (I think this is the correct term?) of my OS for faster recovery from the next time I corrupt some part of my OS.

 

I got my QD3's in this morning.  They are so "cute" compared to what I'm accustomed to working with.

 

Yea, Koolance QDC's are da bomb. I can't even tell you how many disconnect/reconnect's I have done over the past 6 months. I would have lost my mind without them. Is that a tin of Copenhagen snuff I see in the background? Used to do that about 10 years ago in the service. 

Edited April 4, 2022 by Avacado

[tictoc]

On 31/03/2022 at 21:24, ArchStanton said:

I've convinced myself that I do need to revisit my RAM and actually tune it to my system, rather than rely on the settings I plagiarized from J7.  In addition, I've got some QD's coming from Koolance itself.  I did some digging and found these:

 

 

But pricing and availability didn't seem better than what's available in the DIY PC enthusiast market currently, so I didn't bother researching whether chrome plated brass would function for our purposes or not (I'm guessing yes, but it would be prudent to be 100% sure).

 

Here are links to a brochure and the main website if anyone wishes to dig deeper:

Liquid Cooling Connectors for Thermal Management : CPC (cpcworldwide.com)

Non-Spill Liquid Cooling Connectors | CPC (cpcworldwide.com)

 

I think I will likely make more changes to my loop in the future.  Probably going LM on both GPU die and CPU IHS.  More robust cooling on GPU backplate (I do not anticipate active, just "more").  Planning to use QD's to make maintenance far easier and facilitate connection to a "chiller" (maybe on the "redneck" side of the spectrum of possibilities).  I know the gains might be minimal, but for squeezing out every possible drop of cooling performance, I think I will have one of my larger radiators between the CPU and GPU blocks (primarily for the small benefit it might provide when not using a "chiller" of any sort).  Aesthetics are going down the toilet, but I am okay with that now as the "benching bug" has bitten me and I am infected.

 

I have a bunch of the CPC QD's and they work great with not a drop spilled or any issues in the 4 years (maybe longer??) that I've been using them with the external radiators for my main workstation.  There was a time when the pricing on those was really good.  I think that all ended when EK started selling them stand-alone and with their Predator kits.

[ArchStanton]

8 minutes ago, tictoc said:

I think that all ended when EK started selling them stand-alone and with their Predator kits.

Section31 over at OCN said something similar regarding them.  Thank you for the confirmation that they are a quality product.  They presented themselves as such (though the cynic in me suspects EK convinced them to cut some corners for the EK versions, similar to the difference between buying a John Deere at Lowes and buying one at a John Deere dealership, they are usually a totally different grade of product), but it's good to have a firsthand account.

[ArchStanton]

A user even less experienced than myself had requested some representative timings for B-die 16GB DIMMs @ 3800Mt/s over in the big Ryzen RAM thread on OCN.  I threw up a picture of my settings for his perusal.  A few days later another user, Taraquin, suggested some "easy" changes to my settings.  After implementing most of them and validating stability I have yet to see much of a change in performance.  Though I do not think I have tested performance with the right tool (whether the settings will make a difference or not) as temperatures in my "lab" are too high at the moment to make a fair comparison to tests I had run previously.

 

 

I'll probably continue implementing most of his remaining suggestions very slowly and re-verifying stability with each step.  I did learn one DDR4 timing "rule" (actually I don't know if this rule is applicable to only DDR4, or how "firm" a rule it is, or if it matters in the least ).  Per Taraquin, tRP + tRAS = tRC.  I keep hoping I will stumble upon a repository of all these "rules".  I also hope that finding said repository will improve my understanding of how all the different timings interact.

 

I do still plan to limit vSOC to 1.1v or less going forward.  This I believe will take many of the more "extreme" profiles I've seen off the table, but that is likely a good thing for a novice like myself.

 

 

 

 

[ArchStanton]

I've been significantly busier at work the past few days (it's about time, I was starting to get a hair nervous).  Consequently, I haven't had the opportunity to experiment as much recently.  Some impressions I've formed in the meantime.  Zentimings/Aida64 screenies are quick, easy, and over-rated.  They don't provide any sort of "useable" performance delta between one configuration and another.

 

I did install MannIX-ITA's new utility early this morning.  Seems pretty nice.  I didn't have time to study my results (or figure out why my max wattage/temp occurred on a 24t test rather than a 32t test)

 

[ArchStanton]

Made some decent progress on a RAM OC profile that keeps voltages at or below self-imposed caps (mostly just implemented suggestions from Kedarwolf of OCN, they were unsolicited, but fruitful).  I have started using y-cruncher bench 2.5B as a RAM OC/performance diagnostic tool.

 

MannIX-ITA continues to refine his benchmarking utility.  It helped me eliminate some clock stretching that I had missed.  I strongly encourage folks to check his utility out.

 

 

 

Release v1.0.25 Alpha · mann1x/BenchMaestro

GITHUB.COM

---

v1.0.25 Alpha Fix: Logs directory creation issue v1.0.24 Alpha Add: Bold and colored live stats Add: Live running and remaining runtime Add: Unload of WinRing DLL at exit Add: Logs moved to Log...

 

BenchMaestro - CPU & GPU benchmarking and Tools Utility | Overclock.net

 

BenchMaestro - Post your results | Overclock.net