What is your max infinity fabric on ryzen 5000? September 2021 edition

[Mussels]

So according to the table it should do ~665 tRFC?

Sigh, lets see if i can corrupt my BIOS again...
Hail Satan TRFC

 

[Taraquin]

Any recommendations for stability testing? sfc /scannow returned two corrupted files (both links to the Onedrive app, which ... okay?), all fixed. Nothing from DISM. Haven't used the PC much since I tuned things yesterday, but it did fine through some online gaming (Rocket League, in my experience can show memory instability as network errors/latency variation/connectivity issues when playing), plus it woke up from sleep this morning record fast. Still at zero WHEA errors, and everything seems to work fine.

Also, question: I set my vSOC to 1.2V, it's reading as 1,169V stable/peak with drops to 1,156V in HWinfo64. Am I right in thinking this is safe long term? And should I still try stepping it down some?


Welp, that did something. 12172 TS CPU score, that's another +300 points! Currently up 10.7% from my baseline score (10991). Not bad! No WHEA errors, but obviously need stability testing. Should I aim for lowering my tRFC towards the timings from the table (~280)?

If you are still at 1.35V then don`t go lower on tRFC. It scales with voltage. At 1.4V it can typically do 150ns\300 tRFC at 4000 and at 1.5V it can do 136ns\272 TRFC. If you are at 1.35V and 320 that is 160ns which is about the best you can do at 1.35V. You can try 304 if you set volt to 1.4V and 288 if you set it to 1.45V. tRFC should be lowered by 16 ticks on dual rank and 8 ticks on single rank. My single rank kit is medicore and I need 1.46V to do 280.

 

[Mussels]

666 was a fail, trying 690


690 posted. First suggestion so far to get that far.

Basic stability says yay

 

[tabascosauz]

@Valantar I do HCI first. I did buy a HCI Pro key for $5 though, so I don't have to open a billion instances (each "instance" is limited to like 3GB). The functionality is the same just that the Pro version handles all that so you can manage everything in one place.

For HCI, 100% coverage no errors is good enough to keep testing, 600% is good enough for me to do my daily work on during the day, and I test overnight till 2000-3000% for final stability.

If that passes, I move onto TM5 with anta777 Extreme1 config. It has a MT.cfg file in its folder that determines the config - should be a page on OCN with all the popular configs in one zip (there's 1usmusv3 and a variety of other anta configs), just rename your choice to MT.cfg to replace the default, then delete cfg.link (or something like that) in the same folder, before running TM5 as admin. If I already tested HCI before, I'll just run 3 back to back runs of TM5.

If that passes, I'll run a bit of LinpackXtreme. iirc it's either the bench or stress option @ either 8GB or 10GB I think. At that point I'd call that daily stable.

For all of these I disable internet and close everything except HWInfo on the second monitor, to keep tabs on DIMM temps.

Since you are pushing IF to the limit, it might be worth running some other benches at the end like P95 Large FFT or OCCT, but honestly imo LinX is plenty good, and any Bus/Interconnect WHEAs will make themselves known to you in due time with normal use.

Basically, while both get the RAM equally warm:

• HCI: appears to be the less stressful bench that is slightly lighter on CPU, but has a knack for finding sporadic errors way down the line (1000-2000%) that slip through shorter tests
• TM5: more demanding test, generally faster at finding errors, but annoying to have to restart so not for overnighting

 

[Taraquin]

666 was a fail, trying 690


690 posted. First suggestion so far to get that far.

Basic stability says yay
View attachment 222725

As you see 3800 with 690 tRFC is superior to 4000 1000+ tRFC. 4000 is generally 2ns faster than 3800 on same timings so just lowering tRFC by 300+ lowered latency by 5ns as for tRFC try 688 or 672, it's better than 690 due to time intervals. Dual rank kits should do increments in 16 ticks and numbers divideable by 16 as said, tRFC is most impactful on performance, the next in line are CL, RCDRD, RP, TRRD/FAW, WR/RTP, but if you are satisfied now I understand

 

[Mussels]

As you see 3800 with 690 tRFC is superior to 4000 1000+ tRFC. 4000 is generally 2ns faster than 3800 on same timings so just lowering tRFC by 300+ lowered latency by 5ns as for tRFC try 688 or 672, it's better than 690 due to time intervals. Dual rank kits should do increments in 16 ticks and numbers divideable by 16 as said, tRFC is most impactful on performance, the next in line are CL, RCDRD, RP, TRRD/FAW, WR/RTP, but if you are satisfied now I understand

I'll try those

Prior to these numbers everyone always said numbers well under 500 and they had zero chance of working on this RAM

672: no post, the usual frozen and no recovering
688:

You know your RAM - and thank you, getting sub 60ns is huge for 2x32GB

TRFC was an outlier on this RAM with a crazy high value, but as i just demonstrated it always wants values far higher than people expect, it's so unstable it wont even POST (not even the automatic failed overclock stuff)

Last thing i need is to spend a lot of effort tuning it in, only to forget the settings, lose them, or have one small change make it all incompatible

 

[Taraquin]

I'll try those

Prior to these numbers everyone always said numbers well under 500 and they had zero chance of working on this RAM

672: no post, the usual frozen and no recovering
688:

View attachment 222739


You know your RAM - and thank you, getting sub 60ns is huge for 2x32GB

TRFC was an outlier on this RAM with a crazy high value, but as i just demonstrated it always wants values far higher than people expect, it's so unstable it wont even POST (not even the automatic failed overclock stuff)

Last thing i need is to spend a lot of effort tuning it in, only to forget the settings, lose them, or have one small change make it all incompatible

As you see even minor changes can have a major impact. Your ram is the worst performer on tRFC of all dies so no wonder most people think you can do tighter tRFC since all other dies can

If 688 turns out to be unstable I'm 99% sure 704 will be stable.

 

[Valantar]

@Valantar I do HCI first. I did buy a HCI Pro key for $5 though, so I don't have to open a billion instances (each "instance" is limited to like 3GB). The functionality is the same just that the Pro version handles all that so you can manage everything in one place.

For HCI, 100% coverage no errors is good enough to keep testing, 600% is good enough for me to do my daily work on during the day, and I test overnight till 2000-3000% for final stability.

If that passes, I move onto TM5 with anta777 Extreme1 config. It has a MT.cfg file in its folder that determines the config - should be a page on OCN with all the popular configs in one zip (there's 1usmusv3 and a variety of other anta configs), just rename your choice to MT.cfg to replace the default, then delete cfg.link (or something like that) in the same folder, before running TM5 as admin. If I already tested HCI before, I'll just run 3 back to back runs of TM5.

If that passes, I'll run a bit of LinpackXtreme. iirc it's either the bench or stress option @ either 8GB or 10GB I think. At that point I'd call that daily stable.

For all of these I disable internet and close everything except HWInfo on the second monitor, to keep tabs on DIMM temps.

Since you are pushing IF to the limit, it might be worth running some other benches at the end like P95 Large FFT or OCCT, but honestly imo LinX is plenty good, and any Bus/Interconnect WHEAs will make themselves known to you in due time with normal use.

Basically, while both get the RAM equally warm:

• HCI: appears to be the less stressful bench that is slightly lighter on CPU, but has a knack for finding sporadic errors way down the line (1000-2000%) that slip through shorter tests
• TM5: more demanding test, generally faster at finding errors, but annoying to have to restart so not for overnighting

Wow, that's quite the procedure. I know memory stability testing is time consuming, but that sounds like a 4-5-day procedure. I'll give it a try though. It seems like HCI has made Memtest "pay what you want", as the free version defaults to "All unused RAM" and there's a Flattr link to pay anything from $3-50 for it - but I tried paying and it just doesn't work, so ... I'll see how I fare with the free version. For now, I'm using the system as normal, we'll see where that gets me. At least I did a system image backup a couple of days ago


Oh, and here's an interesting observation: According to HWinfo64, my RAM after tuning runs 4-5 degrees cooler than previously. I've always thought it ran a bit warm, typically in the mid-40s. There isn't that much airflow in my case, especially at idle as only one fan is running then, it's densely packed, and the only fans blowing anywhere near the RAM is the top radiator fan that's off at idle. Yet now one DIMM is sitting at 41.6° and the other at 42.3°, with that fan off, and the bottom radiator fan at 900rpm. So, by increasing the RAM voltage from 1.35V to 1.38V, increasing clocks from 3200 to 4000, and tuning the timings much more tightly than before, all of a sudden I've reduced the RAM power draw? I guess I'll see how it behaves under load, but either way this is pretty nice.

 

[tabascosauz]

Wow, that's quite the procedure. I know memory stability testing is time consuming, but that sounds like a 4-5-day procedure. I'll give it a try though. It seems like HCI has made Memtest "pay what you want", as the free version defaults to "All unused RAM" and there's a Flattr link to pay anything from $3-50 for it - but I tried paying and it just doesn't work, so ... I'll see how I fare with the free version. For now, I'm using the system as normal, we'll see where that gets me. At least I did a system image backup a couple of days ago

Oh, and here's an interesting observation: According to HWinfo64, my RAM after tuning runs 4-5 degrees cooler than previously. I've always thought it ran a bit warm, typically in the mid-40s. There isn't that much airflow in my case, especially at idle as only one fan is running then, it's densely packed, and the only fans blowing anywhere near the RAM is the top radiator fan that's off at idle. Yet now one DIMM is sitting at 41.6° and the other at 42.3°, with that fan off, and the bottom radiator fan at 900rpm. So, by increasing the RAM voltage from 1.35V to 1.38V, increasing clocks from 3200 to 4000, and tuning the timings much more tightly than before, all of a sudden I've reduced the RAM power draw? I guess I'll see how it behaves under load, but either way this is pretty nice.

It doesn't take that long even if you're exhaustive about it. Normally if you can make it to like 700% in HCI, OCers already consider that daily stable. Or a single overnight pass in HCI if you're extra careful. I've just had some negative experiences with CJR where instability hides out till the 4th TM5 run or into the 2000% range in HCI, so I do more. 700% HCI is done after like 2 hours on dual rank, TM5 about 2 hours per run on dual rank.

With free HCI just make sure you open enough instances. Each is limited to about 3GB so set it so that all your RAM is used save for about 500MB to 1GB remaining physical free memory in HWInfo. Same for HCI Pro, you gotta specify usage and CPU threads. TM5 you don't need to worry about that, but you do need to cone back and restart it, so it's not an overnight tool.

Is 41C in game? If that's idle then it's a bit warm, dunno if you'll make it thru a memtest. The only number that matters is your max logged temp during HCI or TM5, at 1.52V I'm at roughly 47.5C. At idle I'm sitting at about 33C, in games it tops out at 38-42C.

 

[Mussels]

x

As you see even minor changes can have a major impact. Your ram is the worst performer on tRFC of all dies so no wonder most people think you can do tighter tRFC since all other dies can

If 688 turns out to be unstable I'm 99% sure 704 will be stable.

688 did just give me some fun bsods, trying 704

 

[Taraquin]

x

688 did just give me some fun bsods, trying 704

Since you have 32gb sticks, I actually wonder if you should use 32 as ticks multiplier, 704 should be best then

 

[Mussels]

Since you have 32gb sticks, I actually wonder if you should use 32 as ticks multiplier, 704 should be best then

beyond me, but kinda makes sense - 704 has been stable all day.
I suppose keep that knowledge tucked away, as 32GB sticks are gunna become more common

 

[Valantar]

Is 41C in game? If that's idle then it's a bit warm, dunno if you'll make it thru a memtest. The only number that matters is your max logged temp during HCI or TM5, at 1.52V I'm at roughly 47.5C. At idle I'm sitting at about 33C, in games it tops out at 38-42C.

Yep, it's idle, and I've seen mid-40s after long-term desktop use - but as I said, this is with near zero airflow in a densely packed case:

But good news! This finally got me off my butt and made me install the J-Hack M2426 unit (read: 24-to-6, as in reducing the 24-pin cable to just 6), which definitely opens up some room for airflow around the RAM (and in the bottom of the case too).

That's quite the reduction in cable clutter!

From the couple of minutes I've had the system running, if anything it seems to increase idle RAM temperatures - which kind of makes sense, as the warm air off the radiator can rise up past the RAM more easily now. Still, it should make for much better cooling under load. And if that's not enough I could always stick a slim 92mm fan in there (I've got a free header on the Quadro) and set it to turn on if the RAM exceeds something like 47 degrees. That ought to do the trick. But for now, I'll run some stability tests today and see where I end up.

 

[Pictus]

My rev E could run tRFC at 528 at 3733 (good bin), my B-die can do 254. In certain games this could give performancegains of almost 10%, but the rev E compensated somewhat since it could run tRP, tWTRS and tWTRL lower than B-die. Actual performancedifference is around 5-6% when max-tuned.


Is that your highest stable/bootable IF? Micron rev E or B ram? tWR equal to tCL and tRTP half of tWR should work and improve performance tRFC 560 might be doable, maybe lower, my rev E can do 522 at 3733.

Micron rev E, not the max bootable, but I do not want to use higher DRAM voltages.

 

[Valantar]

Well that's disappointing. Especially after >600% memtest and two runs of TM5 anta777 with no errors whatsoever. Looking at the event log it seems to have been while gaming yesterday. Guess I'll either have to bump some voltages or drop a bit below 4000/2000.

 

[tabascosauz]

View attachment 223081
Well that's disappointing. Especially after >600% memtest and two runs of TM5 anta777 with no errors whatsoever. Looking at the event log it seems to have been while gaming yesterday. Guess I'll either have to bump some voltages or drop a bit below 4000/2000.

Bus/Interconnect can surface in TM5 and HCI (seen it only once in TM5) but they aren't meant to test for it directly. Can try Prime95 Large FFT, LinpackXtreme, not sure if OCCT has a dedicated test, there was another alleged dedicated IF tester script somewhere on OCN randomx is the name...........but daily IF stability is best tested through long-term daily use. 38 errors is quite a bit.

Send a zentimings screenie? Let's see those volts.

First time I've seen that kind of relatively slow accumulation, actually. It seems to either be stable (0), almost stable (1-2 after a day), or very unstable (1 every other second):

 

[Valantar]

Bus/Interconnect can surface in TM5 and HCI (seen it only once in TM5) but they aren't meant to test for it directly. Can try Prime95 Large FFT, LinpackXtreme, not sure if OCCT has a dedicated test, there was another alleged dedicated IF tester script somewhere on OCN randomx is the name...........but daily IF stability is best tested through long-term daily use. 38 errors is quite a bit.

Send a zentimings screenie? Let's see those volts.

First time I've seen that kind of relatively slow accumulation, actually. It seems to either be stable (0), almost stable (1-2 after a day), or very unstable (1 every other second):

View attachment 223083

Seems to be only when gaming (across two different games yesterday), and relatively intermittent:

It's also clearly limited to the times I was playing games. Guess it's somehow related to stressing the PCIe link? Looking at the event viewer, it seems like there were some errors on the 28th as well that I apparently didn't spot. Not many though, similar to this.

Zentimings:

 

[Taraquin]

Seems to be only when gaming (across two different games yesterday), and relatively intermittent:
View attachment 223090
It's also clearly limited to the times I was playing games. Guess it's somehow related to stressing the PCIe link? Looking at the event viewer, it seems like there were some errors on the 28th as well that I apparently didn't spot. Not many though, similar to this.

Zentimings:

View attachment 223091

Going higher on voltages probably won' t fix this. You should try 3933/1966, 3866/1933 or 3800/1900, that is the only viable fix :/ Usually 5600X/5800X and ITX/mATX with 2 dimms has the highest chance of hitting 4000/2000+ without whea 19, but sometimes they don't.

3800/1900 usually work, and you can tighten timings then so performance could be close to 4000cl17. You can also lower voltages and then clockspeeds will improves since IO-die uses less power.

 

[Mussels]

3800/1900 usually work, and you can tighten timings then so performance could be close to 4000cl17. You can also lower voltages and then clockspeeds will improves since IO-die uses less power.

Oh yeah i should look into that, ye olde 5800x loves its nice high temps

 

[tabascosauz]

Oh yeah i should look into that, ye olde 5800x loves its nice high temps

IO die really doesn't make any heat at all to affect your temps. If anything it's the opposite, the less IOD draws, the more out of PPT can go to the CCD, so the cores run faster and draw more power and run hotter.

 

[Mussels]

IO die really doesn't make any heat at all to affect your temps. If anything it's the opposite, the less IOD draws, the more out of PPT can go to the CCD, so the cores run faster and draw more power and run hotter.

9V iod, got it

 

[Taraquin]

IO die really doesn't make any heat at all to affect your temps. If anything it's the opposite, the less IOD draws, the more out of PPT can go to the CCD, so the cores run faster and draw more power and run hotter.

At 3800cl15 tuned IO uses 19W, at 4000cl16 it uses 21W on my setup.

 

[tabascosauz]

At 3800cl15 tuned IO uses 19W, at 4000cl16 it uses 21W on my setup.

Yes I agreed with you lol, only corrected the misconception that it's to do with improving temps

Though the power reduction really is only significant if one is able to reduce quite a bit of VSOC (~0.1V). 14-15W @ 3600 vs 16-18W @ 3800, but VSOC is also 1.03V vs 1.125V. Is that 19W when you're hammering the IF (TM5) or just idle? If it's the latter it seems kinda high for 1CCD, but I also don't know your VSOC

 

[Mussels]

Even moving 1W across could help for people chasing benchmark scores, so it was a good reminder

Later tonight i'm using the derbauer offset and liquid metal, i doubt it'll help my IF clocking but it better damn well help those temps

 

[Taraquin]

Yes I agreed with you lol, only corrected the misconception that it's to do with improving temps

Though the power reduction really is only significant if one is able to reduce quite a bit of VSOC (~0.1V). 14-15W @ 3600 vs 16-18W @ 3800, but VSOC is also 1.03V vs 1.125V. Is that 19W when you're hammering the IF (TM5) or just idle? If it's the latter it seems kinda high for 1CCD, but I also don't know your VSOC

2W in CB23/games, a bit more in TM5, don't remember exact, 1W more in idle, total pack pwr idle is 18-25W, IO die uses less than half The sole difference is from 0.04V higher SOC and IOD volt.