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

[Deleted member 215115]

After fixing my minor PEBKAC issue I'm currently at ~160% memtest with 0 errors after going to 2t (stepped down to 1.4V). Current settings:
View attachment 224183
Previous settings that gave errors almost immediately, which from what I can tell are identical except for the command rate:
View attachment 224184
Should I try 1T with GDM enabled instead? Is there a meaningful difference in the performance tradeoff between the two?

Also, does anyone know if ASRock B550 boards lack a software VDIMM readout? I can't find anything resembling that in HWinfo at least.

Use this to calculate your tRFC: https://docs.google.com/spreadsheets/d/1A7G97QOL0dNMwJZa9SYEq2RElJ5T6Hcx9WdReTsnIWw/edit#gid=0 . You should input the correct values for tRFC2 and tRFC4 too even though they are almost never used.

1T + GDM is sort of like a 1.5T performance-wise and is definitely better than 2T. I would stick to 1T+GDM. Since you have no uneven timings like CAS 15, GDM doesn't really affect you much if at all.

There's nothing wrong with the timings in the second screenshot except for the GDM. As I said, GDM off with dual rank memory (2x16 or 4x8) is very tricky to get stable and is ultimately not worth it. Other than that, perhaps tRDWR and tWRRD could be set to 9-3 or 9-1 to improve stability. Generally you want tRDWR to be tRCDRD/2 +1 with as little tWRRD as possible. Sometimes tWRRD 1 works, sometimes only 3 will work. tWTRL will go down to 8, tRRDL will go down to 4 and maybe the SCLs will go down to 2 but keeping the SCLs high in order to lower the primary timings is more desirable.

In my opinion, your kit can do 3800 CL14 with a bit more voltage and with the right rtt/drvstr values. Don't be afraid to set 1.6V to test things. Nothing bad will happen to your memory. You might also want to go higher on the vSOC. It's safe up to 1.2V but you have to keep VDDG IOD within 50/75/100 mV of the SOC and the CCD within 50/75/100 mV of the IOD. They sort of go together from my experience.

As far as stability testing is concerned, the best way to test is by using your PC for a while. Sure, you can run TM5 or Karhu or MemTest but ultimately you will need to check the long-term stability at some point. Attaching TM5 with the 1usmus_v3 config. Run 20 cycles of this bad boy and you should be golden if there are no errors.

Here's me being super super tight if you feel like having some fun.

 

[schidddy]

hi people,
i just cant get my 5900x over 1900mhz IF, 1933 makes Aida64 Cache and Mem Benchmark crash after a few seconds. I may have found the problem but the problem is also my problem.
My gigabyte Aorus Elite wont change/set the VDDG CCD and IOD Settings. CCD and IOD sticking at 0.99v
Soc Voltage is set to 1.13v and should be fine.
RAM is clocked down for IF testing and shouldnt make any problems. 1900 stable, 1933 IF, crash on cache testing.
My Gigabyte lets me change the VDDP and VDDG on the tweakings menu and also in AMD Overclocking in the Settings tab. in the tweaking tab changes doesnt have effect at all. on AMD Overclocking it changes but not to the voltage i set it to it takes whatever it wants but is higher now if i set it to manual.
Somebody experiences on gb boards? Or maybe my 5900x doesnt like anything higher then 1933mhz, on 1966mhz it crashes on the same cache Testing, but also bootable

 

[Taraquin]

Use this to calculate your tRFC: https://docs.google.com/spreadsheets/d/1A7G97QOL0dNMwJZa9SYEq2RElJ5T6Hcx9WdReTsnIWw/edit#gid=0 . You should input the correct values for tRFC2 and tRFC4 too even though they are almost never used.

1T + GDM is sort of like a 1.5T performance-wise and is definitely better than 2T. I would stick to 1T+GDM. Since you have no uneven timings like CAS 15, GDM doesn't really affect you much if at all.

There's nothing wrong with the timings in the second screenshot except for the GDM. As I said, GDM off with dual rank memory (2x16 or 4x8) is very tricky to get stable and is ultimately not worth it. Other than that, perhaps tRDWR and tWRRD could be set to 9-3 or 9-1 to improve stability. Generally you want tRDWR to be tRCDRD/2 +1 with as little tWRRD as possible. Sometimes tWRRD 1 works, sometimes only 3 will work. tWTRL will go down to 8, tRRDL will go down to 4 and maybe the SCLs will go down to 2 but keeping the SCLs high in order to lower the primary timings is more desirable.

In my opinion, your kit can do 3800 CL14 with a bit more voltage and with the right rtt/drvstr values. Don't be afraid to set 1.6V to test things. Nothing bad will happen to your memory. You might also want to go higher on the vSOC. It's safe up to 1.2V but you have to keep VDDG IOD within 50/75/100 mV of the SOC and the CCD within 50/75/100 mV of the IOD. They sort of go together from my experience.

As far as stability testing is concerned, the best way to test is by using your PC for a while. Sure, you can run TM5 or Karhu or MemTest but ultimately you will need to check the long-term stability at some point. Attaching TM5 with the 1usmus_v3 config. Run 20 cycles of this bad boy and you should be golden if there are no errors.

Here's me being super super tight if you feel like having some fun.

View attachment 224229

I mostly agree with you, BUT on Ryzen 5000 it seems 2T often gives better perf than 1T gdm. I get slightly better fps in SOTTR and 0.5-1ns lower latency in aida with 2T.

 

[Deleted member 215115]

I mostly agree with you, BUT on Ryzen 5000 it seems 2T often gives better perf than 1T gdm. I get slightly better fps in SOTTR and 0.5-1ns lower latency in aida with 2T.

Sure, I'm not going to argue your findings on this topic. My experience has been different but I'm not an expert on the matter. From my understanding of how GDM works, it wouldn't make sense for it to be worse unless you've got some uneven timings. These below seem to be affected by GDM:

• tCL

• tWR

• tCWL

• tRTP

These have always been even numbers for me except tCWL which is affected by tRDWR/tWRRD values so I had some fun while testing the performance. Lowest I could set it to was 11 (uneven) but that required high tRDWR and was worse performance-wise.

 

[Valantar]

2T is often a bit faster than gdm, but slightly less stable. Stick with 2T is my advice. You can try 15 15 15 30 45, but you will need 1.43-1.5V to do that. If you set it to 1.45V you can lower trfc to 280 or 272 as it scales with voltage. 1.5V allows for 264 or 256.

Use this to calculate your tRFC: https://docs.google.com/spreadsheets/d/1A7G97QOL0dNMwJZa9SYEq2RElJ5T6Hcx9WdReTsnIWw/edit#gid=0 . You should input the correct values for tRFC2 and tRFC4 too even though they are almost never used.

1T + GDM is sort of like a 1.5T performance-wise and is definitely better than 2T. I would stick to 1T+GDM. Since you have no uneven timings like CAS 15, GDM doesn't really affect you much if at all.

There's nothing wrong with the timings in the second screenshot except for the GDM. As I said, GDM off with dual rank memory (2x16 or 4x8) is very tricky to get stable and is ultimately not worth it. Other than that, perhaps tRDWR and tWRRD could be set to 9-3 or 9-1 to improve stability. Generally you want tRDWR to be tCL/2 +1 with as little tWRRD as possible. Sometimes tWRRD 1 works, sometimes only 3 will work. tWTRL will go down to 8, tRRDL will go down to 4 and maybe the SCLs will go down to 2 but keeping the SCLs high in order to lower the primary timings is more desirable.

In my opinion, your kit can do 3800 CL14 with a bit more voltage and with the right rtt/drvstr values. Don't be afraid to set 1.6V to test things. Nothing bad will happen to your memory. You might also want to go higher on the vSOC. It's safe up to 1.2V but you have to keep VDDG IOD within 50/75/100 mV of the SOC and the CCD within 50/75/100 mV of the IOD. They sort of go together from my experience.

As far as stability testing is concerned, the best way to test is by using your PC for a while. Sure, you can run TM5 or Karhu or MemTest but ultimately you will need to check the long-term stability at some point. Attaching TM5 with the 1usmus_v3 config. Run 20 cycles of this bad boy and you should be golden if there are no errors.

Here's me being super super tight if you feel like having some fun.

View attachment 224229

I ran HCI Memtest until ~150% with zero errors, so I decided to switch to TM5 anta777 to see whether it triggers anything. If not, I'll run HCI overnight and probably leave it here for now. At the settings @Taraquin suggested above, trfc272, 1.45V, 0 errors in HCI.

TM5 has the added advantage of running a consistent 100% CPU load with a relatively low heat output, so my Quadro switches into 100% fan mode (because my default profile is very tuned for silence) meaning TM5 actually makes my DIMM temperatures go down

I'm well aware that I'll have to do long-term in-use testing to determine absolute stability and how memory instability can be really sporadic and weird,, but at the same time I think @tabascosauz's suggestions previously about testing before this make sense (though I certainly don't have the patience for the suggested 600% HCI + 3 runs of TM5 + 2000% HCI - judging by the time it takes me to get to 100% HCI that would likely take more than 24 hours!). I have no qualms whatsoever with stepping back to some more conservative settings if there's an indication of instability, but for now I'm really just having fun seeing where this can go.

 

[Taraquin]

Sure, I'm not going to argue your findings on this topic. My experience has been different but I'm not an expert on the matter. From my understanding of how GDM works, it wouldn't make sense for it to be worse unless you've got some uneven timings. These below seem to be affected by GDM:

• tCL

• tWR

• tCWL

• tRTP

These have always been even numbers for me except tCWL which is affected by tRDWR/tWRRD values so I had some fun while testing the performance. Lowest I could set it to was 11 (uneven) but that required high tRDWR and was worse performance-wise.

Yeah, on Ryzen 3000 I got better perf with 1T gdm, but on 5000 I get better perf with 2T. Difference is small though.

I ran HCI Memtest until ~150% with zero errors, so I decided to switch to TM5 anta777 to see whether it triggers anything. If not, I'll run HCI overnight and probably leave it here for now. At the settings @Taraquin suggested above, trfc272, 1.45V, 0 errors in HCI.

TM5 has the added advantage of running a consistent 100% CPU load with a relatively low heat output, so my Quadro switches into 100% fan mode (because my default profile is very tuned for silence) meaning TM5 actually makes my DIMM temperatures go down
View attachment 224236
I'm well aware that I'll have to do long-term in-use testing to determine absolute stability and how memory instability can be really sporadic and weird,, but at the same time I think @tabascosauz's suggestions previously about testing before this make sense (though I certainly don't have the patience for the suggested 600% HCI + 3 runs of TM5 + 2000% HCI - judging by the time it takes me to get to 100% HCI that would likely take more than 24 hours!). I have no qualms whatsoever with stepping back to some more conservative settings if there's an indication of instability, but for now I'm really just having fun seeing where this can go.

Try twr 14, trtp 7, iod will probably work fine at 1.00V, if not try 1.04V.

 

[schidddy]

hello over there, i got my problem solved with the vddp/vddg volts
now i have some further questions if u people maybe able to help i would be glad. I see some people here running their b-die at 4000 on cl14 and such crazy settings. Currently im stuck at 3800 at 16-16-16-32-50
Any step up on the RAM to 3866 will not even let me post, and i am already at 1.45v with the RAMs (hwinfo reads 1.47v, RAM is cooled with an additional fan and doesnt go over 43c
i did attach my Zen Timings, any suggestion or just unlucky with a shit sample. IF 1900 runs stable
edit its trident z neo 3600 cl16 rated kit

 

[tabascosauz]

I mostly agree with you, BUT on Ryzen 5000 it seems 2T often gives better perf than 1T gdm. I get slightly better fps in SOTTR and 0.5-1ns lower latency in aida with 2T.

Sure, I'm not going to argue your findings on this topic. My experience has been different but I'm not an expert on the matter. From my understanding of how GDM works, it wouldn't make sense for it to be worse unless you've got some uneven timings. These below seem to be affected by GDM:

Neither of you have to be "wrong". GDM vs 2T is not set in stone. The two are tied on my 5900X+CJR, GDM is slightly faster on 5900X+Bdie DR but only in some instances, GDM was faster on 3700X+CJR, the two are the same on 4650G+Bdie SR, GDM is slightly faster on 5600G/5700G+Bdie SR but only in some instances, etc.

I ran HCI Memtest until ~150% with zero errors, so I decided to switch to TM5 anta777 to see whether it triggers anything. If not, I'll run HCI overnight and probably leave it here for now. At the settings @Taraquin suggested above, trfc272, 1.45V, 0 errors in HCI.

TM5 has the added advantage of running a consistent 100% CPU load with a relatively low heat output, so my Quadro switches into 100% fan mode (because my default profile is very tuned for silence) meaning TM5 actually makes my DIMM temperatures go down

I'm well aware that I'll have to do long-term in-use testing to determine absolute stability and how memory instability can be really sporadic and weird,, but at the same time I think @tabascosauz's suggestions previously about testing before this make sense (though I certainly don't have the patience for the suggested 600% HCI + 3 runs of TM5 + 2000% HCI - judging by the time it takes me to get to 100% HCI that would likely take more than 24 hours!). I have no qualms whatsoever with stepping back to some more conservative settings if there's an indication of instability, but for now I'm really just having fun seeing where this can go.

Must be something up with how you ran HCI, it's not that slow. Are you using the paid or free version? On the free version make sure to set each instance to about 3GB then open enough instances to add to fill all of your available memory leaving ~0.5-1GB free (that last instance should be set to whatever number you need).

On the paid version make sure Low Priority Threads is unchecked. Not sure how it affects 8C/16T but I run 24 threads with per thread Auto size.

imo TM5 anta777 should run much harder than HCI ever will. If I leave mine at stock 142/95/140, TM5 can spike to 87C, about 10C higher than any game or stress test (even LinX is like 85C max!).TM5 anta777 runs "100%" load yes just like HCI - but there's a lot of types of "100%" load......which is why HWInfo's per-core power metrics are so important. TM5 tends to ebb and flow in all-core power and temp, but usually cycles an unusually heavy load through a single core that jumps around - with HWInfo you can track that single 11-20W anomaly as it moves around despite everything being at 100%.

 

[Deleted member 215115]

hello over there, i got my problem solved with the vddp/vddg volts
now i have some further questions if u people maybe able to help i would be glad. I see some people here running their b-die at 4000 on cl14 and such crazy settings. Currently im stuck at 3800 at 16-6-16-32-50
Any step up on the RAM to 3866 will not even let me post, and i am already at 1.45v with the RAMs (hwinfo reads 1.47v, RAM is cooled with an additional fan and doesnt go over 43c
i did attach my Zen Timings, any suggestion or just unlucky with a shit sample. IF 1900 runs stable
edit its trident z neo 3600 cl16 rated kit

Instead of pushing for higher FCLK, I would focus on 3800 CL14. Your kit can certainly do that. Feel free to copy my timings from the ZenTimings screenshot above and increase the voltage to 1.5 or 1.55V.

I ran HCI Memtest until ~150% with zero errors, so I decided to switch to TM5 anta777 to see whether it triggers anything. If not, I'll run HCI overnight and probably leave it here for now. At the settings @Taraquin suggested above, trfc272, 1.45V, 0 errors in HCI.

TM5 has the added advantage of running a consistent 100% CPU load with a relatively low heat output, so my Quadro switches into 100% fan mode (because my default profile is very tuned for silence) meaning TM5 actually makes my DIMM temperatures go down

I'm well aware that I'll have to do long-term in-use testing to determine absolute stability and how memory instability can be really sporadic and weird,, but at the same time I think @tabascosauz's suggestions previously about testing before this make sense (though I certainly don't have the patience for the suggested 600% HCI + 3 runs of TM5 + 2000% HCI - judging by the time it takes me to get to 100% HCI that would likely take more than 24 hours!). I have no qualms whatsoever with stepping back to some more conservative settings if there's an indication of instability, but for now I'm really just having fun seeing where this can go.

HCI is a walk in the park compared to TM5 with 1usmus_v3 config. I strongly advise testing with TM5 because if you pass 20 cycles of that, you really don't need to test with any other tool and can begin using your PC normally to check long-term stability. Again, you can find the proper version of TM5 attached to my post #176 above. All you need to do is restart your PC, kill all unnecessary garbage and run the thing as administrator. It will fill your RAM and should take ~1h for 16 GB and maybe ~2h for 32 GB.

By the way, there's a sheet in the tRFC calculator linked above which has more details on each TM5 error and potential fixes. Yet another benefit over HCI.

 

[tabascosauz]

@schidddy I don't think it's your Bdie. What revision is your board? rev.1.0 is a launch day 4-layer board, that doesn't benefit from the later improved B550-era memory topology improvements. Most average 4-layer boards can hit 4266-4600 single rank, but dual rank like yours may cap out at 3733-3866.

 

[Taraquin]

hello over there, i got my problem solved with the vddp/vddg volts
now i have some further questions if u people maybe able to help i would be glad. I see some people here running their b-die at 4000 on cl14 and such crazy settings. Currently im stuck at 3800 at 16-16-16-32-50
Any step up on the RAM to 3866 will not even let me post, and i am already at 1.45v with the RAMs (hwinfo reads 1.47v, RAM is cooled with an additional fan and doesnt go over 43c
i did attach my Zen Timings, any suggestion or just unlucky with a shit sample. IF 1900 runs stable
edit its trident z neo 3600 cl16 rated kit

Trc 48, tRFC 264, 272 or 280, twr 16, trtp 8, wtrs 4, wtrl 8 or 10, scl 4 on both. Try that, it should improve your scores quite a bit.

 

[Deleted member 215115]

@schidddy I don't think it's your Bdie. What revision is your board? rev.1.0 is a launch day 4-layer board, that doesn't benefit from the later improved B550-era memory topology improvements. Most average 4-layer boards can hit 4266-4600 single rank, but dual rank like yours may cap out at 3733-3866.

I think pcie gen 4 requires at least 6 layers for signal integrity.

 

[tabascosauz]

I think pcie gen 4 requires at least 6 layers for signal integrity.

Nah that was early speculation. Aorus Elite is a 4-layer board, Aorus Pro is a 6-layer board. All the AM4 mATX boards except for the new Aorus Pro AX are 4-layer. All Steel Legend/Extreme4 is 4-layer. Gigabyte is one of the most transparent companies about this, all it takes a glimpse at the bottom left of the board.

Most vendors don't advertise different QVL for single rank and dual rank, but MSI does. They put the limit at 4400 SR/3733 DR, I'd say 3800 is already pretty good for a 4-layer. My TUF mATX couldn't do 3866 either. Steel Legend os one of the best 4-layers and iirc in practice it was 4666/3866 for that board.

@rares495 definitely, TM5 is the much harder test to pass. HCI is good for overnight testing because it can catch intermittent errors hours into testing for barely-barely-stable setups, and same goes for IF.

Was lopping off some unnecessary VDIMM and VSOC at the same time. I know I usually need 1.125V + mild LLC for this, but wanted to see if 1.118V + weak LLC would work. No problem for about 2 weeks, and ran TM5 about 20 times, but look what the cat dragged in:

 

[purecain]

I tried the suggestion of lowering ram timings, unfortunately I also changed the settings for the curve optimizer etc to negative 20 on all cores at the same time.
After being stable for 20hours running almost any benchmark, when I started streaming I ran into my first bsod after 10m.

After undoing the settings it was fine streaming for 3 1/2 hrs. I'm not sure weather it was the system or the game for the last crash.

I'm going to add a nice new AIO very soon.

I will be trying again though, I need to read up a little more on what voltages affect which settings and i should be able to get a decent set up.

 

[Deleted member 215115]

Nah that was early speculation. Aorus Elite is a 4-layer board, Aorus Pro is a 6-layer board. All the AM4 mATX boards except for the new Aorus Pro AX are 4-layer. All Steel Legend/Extreme4 is 4-layer. Gigabyte is one of the most transparent companies about this, all it takes a glimpse at the bottom left of the board.

Most vendors don't advertise different QVL for single rank and dual rank, but MSI does. They put the limit at 4400 SR/3733 DR, I'd say 3800 is already pretty good for a 4-layer. My TUF mATX couldn't do 3866 either. Steel Legend os one of the best 4-layers and iirc in practice it was 4666/3866 for that board.

@rares495 definitely, TM5 is the much harder test to pass. HCI is good for overnight testing because it can catch intermittent errors hours into testing for barely-barely-stable setups, and same goes for IF.

Was lopping off some unnecessary VDIMM and VSOC at the same time. I know I usually need 1.125V + mild LLC for this, but wanted to see if 1.118V + weak LLC would work. No problem for about 2 weeks, and ran TM5 about 20 times, but look what the cat dragged in:

Yeah, that's exactly why no tool can guarantee 100% stability. Manual testing over a longer period is always the final step.

I tried the suggestion of lowering ram timings, unfortunately I also changed the settings for the curve optimizer etc to negative 20 on all cores at the same time.
After being stable for 20hours running almost any benchmark, when I started streaming I ran into my first bsod after 10m.

After undoing the settings it was fine streaming for 3 1/2 hrs. I'm not sure weather it was the system or the game for the last crash.

I'm going to add a nice new AIO very soon.

I will be trying again though, I need to read up a little more on what voltages affect which settings and i should be able to get a decent set up.

Use Ryzen Master to find out which cores are the best. Apply -5 negative offset to all the other cores, -10 to the second best core, -15 to the best core. Test that for a while. If all's well, try -15 and -20 on the better cores. Leave the others alone.

 

[Dar1usElite]

2 timings have the name scl in them. Set those to 4. Vddg Iod and ccd is under amd overclocking menu, not in timings.

alrighty ill give that a crack and let you know how it goes

 

[Taraquin]

Yeah, that's exactly why no tool can guarantee 100% stability. Manual testing over a longer period is always the final step.


Use Ryzen Master to find out which cores are the best. Apply -5 negative offset to all the other cores, -10 to the second best core, -15 to the best core. Test that for a while. If all's well, try -15 and -20 on the better cores. Leave the others alone.

On my setup the best core (according to RM) is the one of the 2 that does not UV to -30 at +200 pbo.

 

[Dar1usElite]

Try bios F34, I use the same (agwsa 1.2.0.3b). Try 36 tras, 56 trc yes.

Since it does not boot try a safe approach:
16 22 22 22 38 trc 60, trrds/l 5/8, tfaw 20, twr 16, trtp 8, wtrs/l 4/12, cwl 16, trfc 512, scls 4, rdwr 12, wrrd 3. Disable spread spectrum. 1T and gear down mode on. Let me know if this works, it should

Let your cpu run stock, don't use fixed OC or pbo/co until you have stable ram oc, llc medium is usually fine then.

alright this booted, I found those 2 voltages aswell listed as VDDG IOD and VDDG CCD, both are set in mv aswell.
So im running your safe options and 1.4v on the dram, i guessed this since no mention of ram voltage.
The original timings just gave me a c3 stop code followed by b6 stop code.

 

[Taraquin]

alright this booted, I found those 2 voltages aswell listed as VDDG IOD and VDDG CCD, both are set in mv aswell.
So im running your safe options and 1.4v on the dram, i guessed this since no mention of ram voltage.
The original timings just gave me a c3 stop code followed by b6 stop code.

Okay, thats good. I don't know errorcodes so don't ask Run Testmem 5 with usmus1 config for 5-10 rounds to check stability. If stable try lowering the following: tRFC 496, 480, 464, 448, 432 you probably can't go lower. Start at 496, if stable drop 16 etc, this timing affect gaming the most, lowering it from 512 to 464 for instance might give you 2% performance alone. If your binning us very good, below 450 might work.

When you find lowest try lowering tRCDRD/WR to 21 or 20, next tRP to 21 or 20. If that works lower tRAS by the same and make tRC=tRAS+tRP. If you can run them at 20, set tRDWR to 10. For instance if you can do tRCDs at 20 and tRP at 21 then set tRAS to 36 and tRC to 57. tRAS=tCL+tRCDRD on B-die and CJR/DJR, often does not work on Micron rev E/B and most others.

When you found lowest try trrds/l 4/6 and tFAW 16. If it works try tWTRL 10, 9 or 8.

Hynix DJR only scales CL from voltage, rest don't scale if I remember correct. Since 16 works at 3800 then 1.4V is okay.

Please post a zentimings screenshot.

 

[Valantar]

Running HCI overnight gave me four errors, so I guess those settings aren't 100% stable. Should I try bumping the voltage further (currently at 1.45V) or changing something else?

Must be something up with how you ran HCI, it's not that slow. Are you using the paid or free version? On the free version make sure to set each instance to about 3GB then open enough instances to add to fill all of your available memory leaving ~0.5-1GB free (that last instance should be set to whatever number you need).

On the paid version make sure Low Priority Threads is unchecked. Not sure how it affects 8C/16T but I run 24 threads with per thread Auto size.

imo TM5 anta777 should run much harder than HCI ever will. If I leave mine at stock 142/95/140, TM5 can spike to 87C, about 10C higher than any game or stress test (even LinX is like 85C max!).TM5 anta777 runs "100%" load yes just like HCI - but there's a lot of types of "100%" load......which is why HWInfo's per-core power metrics are so important. TM5 tends to ebb and flow in all-core power and temp, but usually cycles an unusually heavy load through a single core that jumps around - with HWInfo you can track that single 11-20W anomaly as it moves around despite everything being at 100%.

I might need to run more HCI instances, at least 1/thread? I'm just opening 10, which allows me to run 9 at 3000MB and the last one at 1000-2000 depending on what Windows feels like is enough to hit 31GB in-use for the current load. I would assume the pro version automatically allocates a worker to every thread? As I mentioned previously I tried to buy the pro version but their payment system just plain refused to work, so I'm stuck with free for now.

As for the loads, this HCI setup runs at ~80-90% CPU load, 4.7GHz and ~80W. TM5 anta777 runs ~100% load at 4.8GHz, but at or below 100W, hence why I called it a relatively light load. Looking into it now it seems to be hitting my 120A EDC setting. Kind of weird given the package power reading? Especially seeing how this isn't a current spike at all, but rather a sustained load, and the highest TDC reading is 65A. I'm currently running 120/95/120, I guess I could always raise my EDC some.

HCI is a walk in the park compared to TM5 with 1usmus_v3 config. I strongly advise testing with TM5 because if you pass 20 cycles of that, you really don't need to test with any other tool and can begin using your PC normally to check long-term stability. Again, you can find the proper version of TM5 attached to my post #176 above. All you need to do is restart your PC, kill all unnecessary garbage and run the thing as administrator. It will fill your RAM and should take ~1h for 16 GB and maybe ~2h for 32 GB.

By the way, there's a sheet in the tRFC calculator linked above which has more details on each TM5 error and potential fixes. Yet another benefit over HCI.

I know, but as @tabascosauz said above it's nice to have something that can run unattended for long periods. I'll try the 1usmus config of TM5 after deciding what I'll try tweaking next.

 

[Taraquin]

Running HCI overnight gave me four errors, so I guess those settings aren't 100% stable. Should I try bumping the voltage further (currently at 1.45V) or changing something else?

I might need to run more HCI instances, at least 1/thread? I'm just opening 10, which allows me to run 9 at 3000MB and the last one at 1000-2000 depending on what Windows feels like is enough to hit 31GB in-use for the current load. I would assume the pro version automatically allocates a worker to every thread? As I mentioned previously I tried to buy the pro version but their payment system just plain refused to work, so I'm stuck with free for now.

As for the loads, this HCI setup runs at ~80-90% CPU load, 4.7GHz and ~80W. TM5 anta777 runs ~100% load at 4.8GHz, but at or below 100W, hence why I called it a relatively light load. Looking into it now it seems to be hitting my 120A EDC setting. Kind of weird given the package power reading? Especially seeing how this isn't a current spike at all, but rather a sustained load, and the highest TDC reading is 65A. I'm currently running 120/95/120, I guess I could always raise my EDC some.

I know, but as @tabascosauz said above it's nice to have something that can run unattended for long periods. I'll try the 1usmus config of TM5 after deciding what I'll try tweaking next.

If you are satisfied with current timings then bump voltage to 1.42V, that should be enough. You should try tWTRL 8 as this should be equal to tRTP, but might not work. If 272 tRFC works at 1.4V your bin is good. Going to 1.45V can allow tRFC of 264 or maybe even 256.

 

[Valantar]

If you are satisfied with current timings then bump voltage to 1.42V, that should be enough. You should try tWTRL 8 as this should be equal to tRTP, but might not work. If 272 tRFC works at 1.4V your bin is good. Going to 1.45V can allow tRFC of 264 or maybe even 256.

That failed test was at 1.45V, currently testing at 1.48V to see if that makes a difference. Currently into my fourth run of TM5 1usmus with no errors so far, but I won't be here to keep it running for much longer, so I'll just switch over to HCI when I leave and let that run. I finally managed to get their payment system working, so at least now it should be a bit faster.

If this seems error free at 1.48V later today I'll try adusting the timings to your suggestions here

 

[Dar1usElite]

Okay, thats good. I don't know errorcodes so don't ask Run Testmem 5 with usmus1 config for 5-10 rounds to check stability. If stable try lowering the following: tRFC 496, 480, 464, 448, 432 you probably can't go lower. Start at 496, if stable drop 16 etc, this timing affect gaming the most, lowering it from 512 to 464 for instance might give you 2% performance alone. If your binning us very good, below 450 might work.

When you find lowest try lowering tRCDRD/WR to 21 or 20, next tRP to 21 or 20. If that works lower tRAS by the same and make tRC=tRAS+tRP. If you can run them at 20, set tRDWR to 10. For instance if you can do tRCDs at 20 and tRP at 21 then set tRAS to 36 and tRC to 57. tRAS=tCL+tRCDRD on B-die and CJR/DJR, often does not work on Micron rev E/B and most others.

When you found lowest try trrds/l 4/6 and tFAW 16. If it works try tWTRL 10, 9 or 8.

Hynix DJR only scales CL from voltage, rest don't scale if I remember correct. Since 16 works at 3800 then 1.4V is okay.

Please post a zentimings screenshot.

alright ill run testmem5, this is the zentimings screenshot so far

 

[tabascosauz]

That failed test was at 1.45V, currently testing at 1.48V to see if that makes a difference. Currently into my fourth run of TM5 1usmus with no errors so far, but I won't be here to keep it running for much longer, so I'll just switch over to HCI when I leave and let that run. I finally managed to get their payment system working, so at least now it should be a bit faster.

If this seems error free at 1.48V later today I'll try adusting the timings to your suggestions here

Yeah I can't say I've ever seen HCI CPU usage so low even in the free version. I never ran as many instances as CPU threads with HCI free. HCI Pro isn't much different, it just gives you the option to set # of threads, as well as capacity per thread (leaving it on Auto is fine), so I set it to CPU threads and it loads to 100% (well, "100%" as I explained earlier). Just make sure Low Priority Threads is not checked.

Obviously no user input needed to run HCI overnight, but if you're already through 4 runs of TM5 don't fret too much about HCI. Daily stable is generally regarded as 600-700%, but I just have some adverse experiences between 1000-2000% so OCD.

 

[Taraquin]

That failed test was at 1.45V, currently testing at 1.48V to see if that makes a difference. Currently into my fourth run of TM5 1usmus with no errors so far, but I won't be here to keep it running for much longer, so I'll just switch over to HCI when I leave and let that run. I finally managed to get their payment system working, so at least now it should be a bit faster.

If this seems error free at 1.48V later today I'll try adusting the timings to your suggestions here

If you only had 4 errors in HCL overnight then 1.46-1.47V is probably enough. If I pass 20 TM5 usmus1 I call it rock solid Seems your bin is very similar to mine, I need 1.47V for about the same timings at 3800, but 1T.

alright ill run testmem5, this is the zentimings screenshot so far

View attachment 224326

Looks good, try my suggestions 1 by 1 if stable. If unstable set voltage a bit higher. Lots of errors try 1.45V, very few errors try 1.42V. Another boost for fps is trying 1T without gear down mode, but that us trickier, can help you after you have tuned the other timings if you want?