Mid-Life Crisis

[ArchStanton]

It lives!

 

After the obligatory 24hr leak test and redoing the cable management for the whole system (again) we have successfully booted into Windows .  I will be starting from scratch on performance testing and tweaking.  I will still provide test data for the Corsair/Xylem D5 vs "real" D5 in the coming weeks.  I have everything at hand to test those items externally of this build with the exception of one fitting that Lenz/Motion Industries should be providing in the not-too-distant future.

 

I have made some significant changes to the loop.

 

We're now enjoying the benefits of triple D5's in series.  Two of them in the EK XTOP Revo and an additional unit in the FLT 120.  The FLT 120 was a pump-top and res combo that fit in the space available to me and had mounting options I found appealing.  I also added a "swing out" upper fill port with a combination of rotary fittings (upper right-hand corner in the above picture).  I like the added safety and convenience of ball valve + plug combinations for fill/drain ports.  I made allowance for a pressure test port at the outlet of the XTOP Revo.  There's a coolant temperature sensor located in the rear port of the XTOP REVO.  Both the temp sensor and pressure ports are configured in a way that should prevent protrusion of the sensors into the main flow of the loop (initially I threaded the temp sensor into the rear port on the XTOP without the pumps installed and I estimate that the sensor body blocked about 65% of the passage for coolant flow).

 

I decided to place less of an emphasis on aesthetics this time around.  So here is our jankily mounted RAM/VRM fan .

The rubber dampers are sufficient to hold the 140mm fan stationary at is full 3000 RPM, but something would be terribly wrong if the VRM hit the required 85⁰ Celsius on the control curve for full throttle.

 

Close up of the optional pressure test port.

 

Here is a block diagram detailing the current layout of my loop.

 

As indicated by the green text in the block diagram, the pumps are circulating coolant at 1.47gpm in this new iteration.  I suspect we have fallen off the benefit/$ cliff, but as Bastiaan says, "overkill" is not a term we bandy about.

 

I'll have performance tests and tweaks to share in the days and weeks to come.  My next few days will be spent cleaning/straightening up my "laboratory" .

[Avacado]

14 minutes ago, ArchStanton said:

It lives!

 

After the obligatory 24hr leak test and redoing the cable management for the whole system (again) we have successfully booted into Windows .  I will be starting from scratch on performance testing and tweaking.  I will still provide test data for the Corsair/Xylem D5 vs "real" D5 in the coming weeks.  I have everything at hand to test those items externally of this build with the exception of one fitting that Lenz/Motion Industries should be providing in the not-too-distant future.

 

I have made some significant changes to the loop.

 

We're now enjoying the benefits of triple D5's in series.  Two of them in the EK XTOP Revo and an additional unit in the FLT 120.  The FLT 120 was a pump-top and res combo that fit in the space available to me and had mounting options I found appealing.  I also added a "swing out" upper fill port with a combination of rotary fittings (upper right-hand corner in the above picture).  I like the added safety and convenience of ball valve + plug combinations for fill/drain ports.  I made allowance for a pressure test port at the outlet of the XTOP Revo.  There's a coolant temperature sensor located in the rear port of the XTOP REVO.  Both the temp sensor and pressure ports are configured in a way that should prevent protrusion of the sensors into the main flow of the loop (initially I threaded the temp sensor into the rear port on the XTOP without the pumps installed and I estimate that the sensor body blocked about 65% of the passage for coolant flow).

 

I decided to place less of an emphasis on aesthetics this time around.  So here is our jankily mounted RAM/VRM fan .

The rubber dampers are sufficient to hold the 140mm fan stationary at is full 3000 RPM, but something would be terribly wrong if the VRM hit the required 85⁰ Celsius on the control curve for full throttle.

 

Close up of the optional pressure test port.

 

Here is a block diagram detailing the current layout of my loop.

 

As indicated by the green text in the block diagram, the pumps are circulating coolant at 1.47gpm in this new iteration.  I suspect we have fallen off the benefit/$ cliff, but as Bastiaan says, "overkill" is not a term we bandy about.

 

I'll have performance tests and tweaks to share in the days and weeks to come.  My next few days will be spent cleaning/straightening up my "laboratory" .

Looks fantastic! Great work. My favorite color is Orange, so I'm drooling here. Can't wait for the test data on the D5's. 

[ArchStanton]

The little jewels have finally arrived .

 

Also, "Tell me you're a half-arse machinist without telling me you're a half-arse machinist" .

 

[ArchStanton]

Let the games begin!

[ArchStanton]

Pump testing completed.

 

The pump included with the Corsair XD5

 

The veteran

 

Test setup:  Koolance INS-FM19 flow sensor & ADT-FM03 frequency adapter reporting through W+ rpm header on ASUS ROG C8DH, Unijin P255 Series 2.5" Dial, Oil Filled pressure gauge 0-10psi ±1.6% accuracy, data recorded/averaged by HWiNFO64 7.20.4700.0, pumps set to 100% RPM, XD5 pump top/res used for both units, loop restriction created by ball valve at reservoir return, pressure measured at pump outlet, flow meter in middle of loop, approximately 6' of 16/10mm ZMT tubing used to complete circuit.

 

 

 

 

 

The Verdict?  Pretty much a draw.  The differences displayed in the chart are likely within test variance for my homebrew setup.

 

 

Now back to our regularly scheduled programing 

[ArchStanton]

On a whim, I decided to check the pressure coming out of the XTOP Revo while I had the stuff at hand.  I was a little surprised to see 5.75psi.  More than a single pump can even begin to approach (in my testing), and I believe confirmation that my loop is on the "high" side of the restriction spectrum.  Still, very nice to see 1.46-1.47gpm (5546mL/min) circulating continuously, even if it takes 3 pumps to make it happen.

[Avacado]

6 hours ago, ArchStanton said:

Pump testing completed.

 

The pump included with the Corsair XD5

 

The veteran

 

Test setup:  Koolance INS-FM19 flow sensor & ADT-FM03 frequency adapter reporting through W+ rpm header on ASUS ROG C8DH, Unijin P255 Series 2.5" Dial, Oil Filled pressure gauge 0-10psi ±1.6% accuracy, data recorded/averaged by HWiNFO64 7.20.4700.0, pumps set to 100% RPM, XD5 pump top/res used for both units, loop restriction created by ball valve at reservoir return, pressure measured at pump outlet, flow meter in middle of loop, approximately 6' of 16/10mm ZMT tubing used to complete circuit.

 

 

 

 

 

The Verdict?  Pretty much a draw.  The differences displayed in the chart are likely within test variance for my homebrew setup.

 

 

Now back to our regularly scheduled programing 

Well look at that. I'm sure @ciarlatano would like to see this. Seems Corsair needs to change its verbiage. 

[Bastiaan_NL]

Great work @ArchStanton!
I'm glad they are about the same

[Sir Beregond]

4 hours ago, Avacado said:

Well look at that. I'm sure @ciarlatano would like to see this. Seems Corsair needs to change its verbiage. 

Very interesting results. @ciarlatano since your tag didn't work.

[J7SC_Orion]

16 hours ago, ArchStanton said:

On a whim, I decided to check the pressure coming out of the XTOP Revo while I had the stuff at hand.  I was a little surprised to see 5.75psi.  More than a single pump can even begin to approach (in my testing), and I believe confirmation that my loop is on the "high" side of the restriction spectrum.  Still, very nice to see 1.46-1.47gpm (5546mL/min) circulating continuously, even if it takes 3 pumps to make it happen.

 

...one thing to watch out for with 3x D5 is to avoid cavitation (you'll notice it when bleeding air bubbles takes a lot longer). As you know from earlier posts, I have been running triple D5s since March '21 on my 5950X / Dark Hero build, and it took a while to synchronize all 3x D5s and thus avoid cavitation. It was even worse with 5x D5 in an older loop. 

 

I'm not sure it's a perfect method, but I usually fill  and start a new 3x D5 loop with just one D5 running then bring the other ones online. If they are slightly out of synch, you can hear that, never mind see the result of cavitation. 

 

 

Edited March 2, 2022 by J7SC_Orion
typo

[ArchStanton]

1 minute ago, J7SC_Orion said:

 

If they are slightly out of synch, you can hear that, never mind see the result of cavitation. 

Thank you for the tips.  Thankfully I had no issues with cavitation.  I'm pretty familiar with the phenomenon.  The hydraulic pumps on mobile equipment are at risk of it when other components are faulty (usually a clogged intake filter/strainer or a collapsed supply line).  It makes a very distinctive sound, and when we disassemble the pumps, the internals will look like a blacksmith when to work on them with a ballpeen hammer.

[ArchStanton]

I was able to put in some quality tweaking and testing time this weekend.  My initial efforts have been concentrated on single core performance, as I plan to use a static OC in conjunction with the C8DH's dynamic OC feature.  I'm using BIOS 4004, so EDC must be set to 140 (default) if we don't want VID capped at 1.425.  PPT is set to 150 (higher values didn't seem to affect things one way or the other, which makes sense under the test constraints).  Overclocking the RAM was the toughest part for me, and I still have work to do in that area, but I did make decent progress (feels that way to me at least).  Testing done at 20⁰C ambient (#$%@ CB23 and CPU-Z scores made me long for a chiller to cheat with so badly, could taste 1680 and 700).  Overclocks confirmed stable with CoreCycler(P95/720-720/auto runtime/SSE/hyperthreading disabled/100 iterations) and TestMem5 (extreme@anta777 6 cycles).

 

 

 

 

 

 

 

 

 

 

[ArchStanton]

Forget to mention, I checked TDC from 80-140 in CPU-Z and CBR23 ST.  The default value of 95 gave the best performance for single thread high boosting loads (I thought @Sir Beregond might find this handy as he has a similar config to myself and @J7SC_Orion, though J7SC runs the older 3501 BIOS with the uncapped EDC I believe).

 

Edit:  Err, excuse me. It's not an uncapped EDC.  On BIOS 3801 and older, the VID isn't limited to 1.425 IF EDC is set above 140.  There, that's more accurate.  

Edited March 7, 2022 by ArchStanton
Clarification

[J7SC_Orion]

I have all PBO values such as TDC and EDC on default...picked the Dark Hero / 5950X combo up from the store we do a lot of business with, and it came with bios 3501 which works just fine re. 'everything, all the time', including r_BAR. I'm a bios update Luddite anyway (unless it is for a security fix) if I haven't got any issues in the first place 

[ArchStanton]

I've learned quite a bit tonight, I think.  I started out tinkering with a static OC for use in conjunction with the C8DH's dynamic OC feature.  What I believe I learned is that it's relatively easy to setup a static OC that will throw out some big numbers in CBR23 (31,500+ for MT) or CPU-Z, but if we try to tune the static OC for a worst-case scenario (I used prime95/small FFT/with AVX/AVX2 enabled), by the time we get the overclock stable and thermally manage able our increased performance has evaporated.  Since AMD based systems don't have an AVX offset, that I'm aware of, the options seem somewhat limited.  I have the impression that at the end of the day, PBO2 and its management algorithm keep all the really handy adjustment dials hidden behind the curtain.

 

My self-imposed goal of a maximum overclock that is 100% daily driving safe and stable (even under less than likely usage cases like the prime95 configuration indicated above) hasn't changed.  So, at this point I plan to chuck ASUS's DOC feature out with the bath water, rollback to an older driver, and try again with just PBO2.  I can imagine specific use cases where DOC switching might be advantageous, but I don't believe they currently apply to me.

 

I'll try to summarize some of the behavior I observed from my system tonight.  I maxed out my ambient cooling capabilities at 303W.  This loading was achieved using the indicated prime95 configuration.  Other settings included 1.35 VID set / 1.203 VID get (LLC 2).  HWiNFO64 indicated a peak "CPU" temperature of 82⁰C, but I had a peak core temperature of 94.5⁰C.  CCD0 clocked @ 4.5GHZ and CCD1 @ 4.45GHZ.  At these settings, I was scoring around 30,150 in CBR23.  I think my previous best was 30,500ish with PBO2 (admittedly with a less than 100% stable CPU, but no RAM tune to speak of).  I think I can claw my way back to the 30,500 neighborhood with rock solid stability given time.  Wish me luck .

 

P.S.  I'm don't doubt a liquid metal application between CPU die and waterblock could assist me, but I'm not in the mood to drain and dismantle the loop again currently. I certainly do get the appeal of quick disconnect fittings now, even if they do introduce a little more flow restriction.

[Sir Beregond]

On 24/02/2022 at 11:08, ArchStanton said:

 

 

 

If that's an EK reservoir, I'd be careful with that black "foam" anti-cyclone thing. After a couple years they always broke apart so these days I opt for the plastic insert instead.

[ArchStanton]

3 minutes ago, Sir Beregond said:

 

If that's an EK reservoir, I'd be careful with that black "foam" anti-cyclone thing. After a couple years they always broke apart so these days I opt for the plastic insert instead.

Thanks for the tip.  I'll make "other arrangements" at my next loop maintenance.  

[Sir Beregond]

3 minutes ago, ArchStanton said:

Thanks for the tip.  I'll make "other arrangements" at my next loop maintenance.  

Yeah I don't know if that was me just having bad luck or not, but I did maintenance one year and noticed it looking kinda weird. It had kinda disintegrated as in all the little pieces of it were breaking apart. Had to "pour" it out.

 

Since then, I just started using the plastic insert instead.

Edited March 8, 2022 by Sir Beregond

[ArchStanton]

It's only stable enough to run the ST bench, but at least I can die now in the 700 club.

[ArchStanton]

A "state of the union" type update.

 

I've arrived at my "fully stable" daily driver OC settings.  DRAM settings verified with 50 cycles of TestMem5 (1usmus_v3), 6 cycles of TestMem5 (extreme@anta777), and 48hrs OCCT (80%/Auto instruction set).  PBO2 settings verified with more time in corecycler than an Intel fanboy would be willing to endure (SSE/AVX/AVX2/ALL FFTs), 24hrs looped CBR23 MT, 24hrs OCCT CPU stress (small/extreme/variable/auto/auto).

 

PBO2 PPT 300, TDC 160, EDC 225. +25MHz (5075), scalar x10.  Curve optimizer cores/counts:

 

BIOS 3801 (AGESA 1203_C).  Chipset drivers Rev. 4.03.03.431.  20⁰C ambient air / 22⁰C coolant / "daily driver" fan profile for the benchmarks below.

 

 

 

 

 

 

 

 

I've elected to forgo dynamic overclock switching for daily driving for the time being.  As such, I left some performance on the table.  While PBO2 alone forces one to choose between max single thread or max multithread or something in the middle (what I opted for), I believe it does a better job than I can at managing the CPU under all the conceivable loads that it might face.  Were this a system intended for a specific task, I think DOC switching could be beneficial, but the user would need to keep a tight rein on the programs being launched.

 

Now it's time for me to make a decision.  Do I spend the next month trying to improve on the DRAM settings that I plagiarized from J7, or do I jump ship, download More Power Tool, and have fun squeezing my GPU?   I'm pretty confident as to which one will be the most fun, but I think long term I'd benefit more from learning the ins and outs of all the various DRAM settings and how they interact.  I'll sleep on it for at least tonight.

 

 

[ArchStanton]

A "stock" (for GPU settings anyway) Time Spy run for s#!*s and giggles.

 

 

P.S.  There's chatter over at OCN about those very fresh chipset drivers I'm running resulting in about 150-200 points less in CBR23 MT.  So, we'll see how all that shakes out in the days to come.

[ArchStanton]

Spent quite a bit of time in Time Spy this evening.  This was as far as I could go without "More Power Tool".

 

[ArchStanton]

I left the window open in my laboratory last night in the hopes of some lower ambient temperatures.  Despite what some supposed "experts" have said elsewhere, J7 was correct in his assertions (the bowels of a thread over at OCN) that Radeon 6000 does scale with temperature to a measurable amount.  Also, I have the impression that marginal stability on the core OC cannot support as much memory OC as a slightly slower clocked core.  This run was completed at 2700-2800 core, factory PL +15% from Adrenaline, 1.2v core, fast memory timings, 2150MHz memory clock slider.  I have completed a couple runs at 2725-2825, but it was touch and go with about a 35% success rate and lower scores.  Moving the memory slider above 2000MHz at those core clocks was almost certain to generate a crash.

 

 

I plan to read the remainder of [Official] AMD Radeon RX 6900 XT Owner's Club | Page 324 | Overclock.net (I'm on page 115 currently) before proceeding to play with MPT.

 

Edit:  Studying the readings in HWiNFO64, I suspect there is a significant performance increase when we keep that "hotspot" temp below 70⁰C (somewhat independent of the delta to "GPU temp").  If I am correct, keeping my hotspot below 70⁰C with an ambient around 20⁰C while using MPT will not be possible without a remount/LM.

Edited March 27, 2022 by ArchStanton
additional musings

[ArchStanton]

Instead of continuing to study the ins and outs of MPT as I should, I got side-tracked this evening and played around with a BIOS profile for Time Spy (to increase CPU score).  My efforts bore some fruit.  I will try these settings again tomorrow in the hopefully chilly pre-dawn .

 

 

Edit:  I did get a run in early this morning.  Scores were slightly higher, but I believe within margin of error.  Based on what I've read elsewhere, there may be temperature dependent thresholds for things like Vdroop on this GPU.  As long as I can stay below those thresholds there is limited scaling with decreasing temperature (still guessing 70⁰C hotspot is one such threshold).

Edited March 28, 2022 by ArchStanton
additional info

[ArchStanton]

The grapevine says that Time Spy gets confused by CPU's with >=16 physical cores.  Temporarily disabling SMT/HT did net me another 1000 points in the CPU portion of the test.  @Sir Beregond the "trick" is not supposed to work on a 5900X, but d@#n if I wouldn't check just to be sure.