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

[Cowboystrekk]

Screwing around with these Micron E die sticks with my Ryzen 1600. So far so good.

View attachment 226454

That is awesome, I thought IMC on Ryzen gen 1 maxed at 3600ish. If you wanna improve a bit further you can compare to what I ran on rev E with my 3600, I set relatively safe values for you behind if needed, but you may be able to run what I did on some timings:
11 rp (16)
56 rc (60)
4/6/16 rrds/l/faw (5/8/20)
Wtr 3/6 (4/12)
Wr/rtp 12/6
528 rfc (560)
4 rdrdscl/wrwrscl
9 rdwr (10)
1 wrrd (3)

A 2400G I build managed 3533 on rev E with 16-18-18-16.

 

[damric]

That is awesome, I thought IMC on Ryzen gen 1 maxed at 3600ish. If you wanna improve a bit further you can compare to what I ran on rev E with my 3600, I set relatively safe values for you behind if needed, but you may be able to run what I did on some timings:
11 rp (16)
56 rc (60)
4/6/16 rrds/l/faw (5/8/20)
Wtr 3/6 (4/12)
Wr/rtp 12/6
528 rfc (560)
4 rdrdscl/wrwrscl
9 rdwr (10)
1 wrrd (3)

A 2400G I build managed 3533 on rev E with 16-18-18-16.

Looks like 3733 got an error, and errored again even with more voltage so I rolled it back to 3600MT/s. 3600MT/s passed both AIDA64 memtest and memtest 64. I'm going to try tighter timings. I'll see if yours work.

This is one of the 1600 AFs I snatched for $85 last year from Amazon. I sold my 3600 to a friend and this is a holdover until I see what the 3D cache Ryzens can do next year. This one seems to overclock very well and doesn't mind pushing these 4 Micron E DIMMs on this Taichi x470 board.

 

[Cowboystrekk]

Looks like 3733 got an error, and errored again even with more voltage so I rolled it back to 3600MT/s. 3600MT/s passed both AIDA64 memtest and memtest 64. I'm going to try tighter timings. I'll see if yours work.

This is one of the 1600 AFs I snatched for $85 last year from Amazon. I sold my 3600 to a friend and this is a holdover until I see what the 3D cache Ryzens can do next year. This one seems to overclock very well and doesn't mind pushing these 4 Micron E DIMMs on this Taichi x470 board.

Ah, thought you meant the original 1600, AF has better IMC, but if you manage to stabilize 3733 that is still impressive!

 

[damric]

Ah, thought you meant the original 1600, AF has better IMC, but if you manage to stabilize 3733 that is still impressive!

It seems like the best it will do is running slightly above 3600MT/s (using 101.0 base clock) with 1T and GDM disabled. That's still way better than I was able to do with the other 1600AFs and various other RAMs like Hynix DJR and Samsung B-die. I'm still figuring out which timings I can tighten up without throwing errors, but you know how that goes, slow process.

Edit: Ok fooled around a bit and this seems stable. This is a mishmash of timings from the calculator, xmp tables, and your suggested timings, and it seems to work.

 

[DLD]

You latency is kinda high. Try the following: set soc volt to 1.1v, see if latency improves, try 1.12V and 1.06v iod, see if that improves more. You should be getting around 55ns with current settings. As for further tweaks cl 15 should be possible at 2T gdm off, but you probably need to change cad, procODT etc. You can try 48 proc, 40 20 30 20 or 60 20 30 20 and see if it helps. If you can boot 2T we have room for improvement and maybe cl15 can work Let us know how this goes.

I agree it should be lower. Not sure why I'm always around 58ish. I changed the soc and iod to what you suggested, but still seems to be the same.

First bench gave just over 60 ns, then ran it again right after and got 57 ns.

 

[Ibizadr]

What causes whea 19 errors? Frequency?

 

[Cowboystrekk]

What causes whea 19 errors? Frequency?

Generally if infinity fabric runs above 1900 (ram at 3800+) many motherboards/cpus gets whea 19, some very many of them. If you have 5600X or 5800X you have a smaller chance of whea vs 5900X/5950X. Also small motherboards with only 2 dimmslots (ITX and some mATX) seems less likely to get whea above 3800/1900.

 

[Ibizadr]

Generally if infinity fabric runs above 1900 (ram at 3800+) many motherboards/cpus gets whea 19, some very many of them. If you have 5600X or 5800X you have a smaller chance of whea vs 5900X/5950X. Also small motherboards with only 2 dimmslots (ITX and some mATX) seems less likely to get whea above 3800/1900.

That's make sense I'm running 3866mhz I today I see I got some whea and lower my frequency to 3800MHz but didn't test it yet. I will go for 3800MHz in my 5800x and dont have any bsod or reboot but I want to get rid of that type of errors. Ty for your help one more time.

 

[Cowboystrekk]

I agree it should be lower. Not sure why I'm always around 58ish. I changed the soc and iod to what you suggested, but still seems to be the same.

First bench gave just over 60 ns, then ran it again right after and got 57 ns.

Low on first run can be background processes interfering. 57 is still a bit high, do you get whea 19 in event viewer? Set soc/iod back to what you had if performance is the same.

 

[DLD]

Low on first run can be background processes interfering. 57 is still a bit high, do you get whea 19 in event viewer? Set soc/iod back to what you had if performance is the same.

0 events in event viewer. I don't know why my latency is high or what the cause is. I was hoping to get into the low 50s with this kit.

 

[Cowboystrekk]

0 events in event viewer. I don't know why my latency is high or what the cause is. I was hoping to get into the low 50s with this kit.

Try running aida in safe mode, that should verify if background processes is the issue. Aside from that you could try stabilizing 2T. Try cad 40 20 30 20 and start with changing:
15 cl
14 cwl
14 wr
7 rtp

Set voltage to 1.5V. See if it boots and is stable. If it works try
15 rcd
15 rp
30 ras
45 rc

 

[DLD]

Try running aida in safe mode, that should verify if background processes is the issue. Aside from that you could try stabilizing 2T. Try cad 40 20 30 20 and start with changing:
15 cl
14 cwl
14 wr
7 rtp

Set voltage to 1.5V. See if it boots and is stable. If it works try
15 rcd
15 rp
30 ras
45 rc

Tried the first values you mentioned. Did not like that and required CMOS reset.

 

[Cowboystrekk]

Tried the first values you mentioned. Did not like that and required CMOS reset.

Okay, then it seems your ram dislikes low CL\CWL or 2T-settings. Can you try again with cad at 60 20 40 20? What is the highest voltage you can run on the ram without issues?

You could try at 1.45V at 1T\gdm
29 ras
45 rc
12 wr
6 rtp

If you can run 1.5V on ram without problems try lowering rfc. You can try 256. You can also try 4\8 wtr and 9 rdwr. If you are able to do all that you should get a bit better latency.

 

[DLD]

Okay, then it seems your ram dislikes low CL\CWL or 2T-settings. Can you try again with cad at 60 20 40 20? What is the highest voltage you can run on the ram without issues?

You could try at 1.45V at 1T\gdm
29 ras
45 rc
12 wr
6 rtp

If you can run 1.5V on ram without problems try lowering rfc. You can try 256. You can also try 4\8 wtr and 9 rdwr. If you are able to do all that you should get a bit better latency.

I'll give it a try. Not sure on the voltage. I only ever went to 1.5, sometimes my ambient temp gets a little warm and have had RAM temps get up to around 52c while gaming.

 

[Cowboystrekk]

I'll give it a try. Not sure on the voltage. I only ever went to 1.5, sometimes my ambient temp gets a little warm and have had RAM temps get up to around 52c while gaming.

Okay, stick with 1.45V then, and 272 rcf, but you can try all the other advice.

 

[DLD]

Okay, then it seems your ram dislikes low CL\CWL or 2T-settings. Can you try again with cad at 60 20 40 20? What is the highest voltage you can run on the ram without issues?

You could try at 1.45V at 1T\gdm
29 ras
45 rc
12 wr
6 rtp

If you can run 1.5V on ram without problems try lowering rfc. You can try 256. You can also try 4\8 wtr and 9 rdwr. If you are able to do all that you should get a bit better latency.

Still a no go. Did not require a CMOS reset, but it kept restarting over and over and then required me to go in bios and revert back.

So as of right now this is the current stable settings at 1.45v.

 

[tabascosauz]

@DLD just an idea, can you make sure TSME is disabled in BIOS? Just use the search function. It shouldn't theoretically be possible, but it's super quick to go in there and change, and honestly you never know with AMD.

 

[DLD]

@DLD just an idea, can you make sure TSME is disabled in BIOS? Just use the search function. It shouldn't theoretically be possible, but it's super quick to go in there and change, and honestly you never know with AMD.

Just checked. It was on Auto.

 

[Cowboystrekk]

Still a no go. Did not require a CMOS reset, but it kept restarting over and over and then required me to go in bios and revert back.

So as of right now this is the current stable settings at 1.45v.

Okay, did you test all of the above tweaks? They should be tested one by one. Start with 29 ras 45 rc, if it works try 272 rfc, if it works try 8 wtrl, if it works try 12 wr/6 rtp, if it works try 9 rdwr.

Also: Disable spread spectrum, that can affect stability.

Next up is testing curve optimizer, than can boost performance by a few percent.

 

[DLD]

Okay, did you test all of the above tweaks? They should be tested one by one. Start with 29 ras 45 rc, if it works try 272 rfc, if it works try 8 wtrl, if it works try 12 wr/6 rtp, if it works try 9 rdwr.

Also: Disable spread spectrum, that can affect stability.

Next up is testing curve optimizer, than can boost performance by a few percent.

Only tried what you suggested to get 2T working. Going to start testing the other stuff now.

Just ran the benchmark in safemode. Much better latency there.

 

[Cowboystrekk]

Only tried what you suggested to get 2T working. Going to start testing the other stuff now.

Just ran the benchmark in safemode. Much better latency there.

It seems you have a lot of background processes etc in windows slowing you down. Difference between safe mode andcregular is 1ns on my setup. 4ns must be many apps etc, try cleaning up

 

[DLD]

It seems you have a lot of background processes etc in windows slowing you down. Difference between safe mode andcregular is 1ns on my setup. 4ns must be many apps etc, try cleaning up

29 ras and 45 rc ran an hour with 0 errors.

As far as background stuff goes, not sure what's left to get rid of. I'm on a fresh OS install and only Asus GPU tweak and Corsair ICUE. Nothing stands out to me in task manager either. So not sure what's running to cause that.

So all the things you suggested to try passed with a few hours of stress test and a few hours of gaming.

This is where we are at currently.

 

[Cowboystrekk]

29 ras and 45 rc ran an hour with 0 errors.

As far as background stuff goes, not sure what's left to get rid of. I'm on a fresh OS install and only Asus GPU tweak and Corsair ICUE. Nothing stands out to me in task manager either. So not sure what's running to cause that.

So all the things you suggested to try passed with a few hours of stress test and a few hours of gaming.

This is where we are at currently.

Looks good man, you can try windows 10 debloater to speed up windows. Also try curve optimizer

As for further tweaking you might be able to run RAS and RC lower. Remember that RC=RAS+RP. You can try 28 44, 27 43, 26 42 etc and see how that works. You may be able to run RP lower aswell, maybe 15 or 14 but it might need more voltage.

 

[tabascosauz]

Only tried what you suggested to get 2T working. Going to start testing the other stuff now.

Just ran the benchmark in safemode. Much better latency there.

Those are good results. Spot on if not even faster than expected for CL16.

iCUE seems to be pretty notorious for hogging resources in the background. At least past revisions, not sure about currently.

If GPU Tweak is still Afterburner-based then it's still a monitoring program, which usually affect AIDA latency a little bit. Same goes for HWInfo. Not a problem, just exit them when you're benchmarking.

 

[Cowboystrekk]

There's a tRFC calculator made by a very knowledgeable person from OCN. Use the pre-release sheet.

DDR4/5 Helper by Veii

docs.google.com

I've used that before, but according to anta777 it's very simple: 8GB modules operates in 8 ticks, 16GB in 16 ticks so a good timed tRFC should be divideable by 8 if 2x8 or 4x8 and divideable by 16 if 2x16 or 4x16. This will sometimes match the tRC x 6 or 8. I don't know what is right/perfect, but if he's correct (he made several presets for TM5 so I guess he knows much) then tRFC 280 and 265 on 2x16 would get identical performance since you need to run 264 or lower for previous tick-cycle. The performacediff between 264 and 280 is marginal at best anyways, but voltage requirements on B-die is a factor. If run at 4000 265 is 132.5ns while 280 is 140ns which translate into 0.05v which is quite a difference on B-die which is very temp sensitive.

The calc itself seems solid for finding a baseline, but I believe anta777 is correct so finding the nearest divideable by 8 or 16 can be good advice