Share your AIDA 64 cache and memory benchmark here

[zx128k]

Yes, I agree.
I dialed down the frequency to 3800 and 1800 fclk and it looks good now.
My system really doesn't like anything above 1800.

Unstable IF with Ryzen caused me no end of problems with me 3800x system.

 

[lawood]

Twr =2 X trtp


1800 wasn't 1=1 to 3800. You need to run fclk==mclk. Só mclk at 1900 should do ram at 3800. If it boot 38xx with whea errors it can do 3800 stable for sure

It was a typo, i meant to say 1900. math is hard.

 

[freeagent]

Just default board timings..

Edit:

Except tRFC

 

[zx128k]

Just default board timings..

View attachment 245435


Edit:

Except tRFC

Getting GDM off is a big bump in performance with 1t command rate.

 

[freeagent]

Getting GDM off is a big bump in performance with 1t command rate.

With 4 sticks and straight 1T is hard to get stable past 34xx or something like that for my setup.. could just be the operator though.

 

[zx128k]

With 4 sticks and straight 1T is hard to get stable past 34xx or something like that for my setup.. could just be the operator though.

Totally worth it if you can.

 

[freeagent]

Totally worth it if you can.

I could drop 2 sticks and run 3600 12-12-12 1T but its not really worth it..

 

[zx128k]

I could drop 2 sticks and run 3600 12-12-12 1T but its not really worth it..

If the RAM does not do 1T then thats fine, its just very nice when it does.
Final RAM overclock results, passed TestMem5 v0.12 extreme@anta777 three passes. tXP and Precharge power down (PPD) has a major impact on AIDA64 memory latency. PPD the most. tXP to 4 and PPD to 0 decreases latency a lot. I have tXP to auto and PPD to 0. That was enough to drop me down from >40ns to <40ns in aida64. Maximum speeds for DDR4 4000 is 64000. Timings are tight but not extreme. IOLs etc are tightened, these increase performance and reduce latency. RAM heat in tests feels hot, so got a temperature measuring multimeter to see what they are currently.

 

[Dia01]

3933C16 @ 1.480V DIMM
4000 still a working progress....

 

[freeagent]

My Strix -F topped out at 1900 1:1 with 4 sticks and everything in my setup.. wouldn't even try to start..

The XE pushes just a little bit harder.. That capture is not stable.. yet?.. But it is still pretty neat to me

I thought the XE was just a fancy -F, but so far everything about it is just a little nicer.

 

[tabascosauz]

I think the last time I experimented with this was in AGESA 1202? Back then it was accumulating Bus/Interconnects every second, now it only seems to come in waves when running tests, so maybe time has helped Fabric stability a little bit?

that one is an actual error on core 0 of your cpu; whea 19 warnings show as self correcting with no relevant source attached to them

the whea19 error on core0 was typical on my 3900x if I used undervolt on the vcore, maybe 1 or 2 every month would show, also 100% stable under testing,apps,games.

Yes, it's true that the Bus/Interconnects at 2000MHz Fabric are actually warnings/correctable errors. But it's a bit of a stretch to call that stable. The WHEAs that ailed my 3700X @ 1800Mhz and 1866Mhz were only "correctable errors". After that hellhole I learned to play it safe with Fabric. Before you ask, the RAM itself was 24 hours+ stable in HCI - any shenanigans were coming exclusively from the CPU.

Can't wait for my replacement 5700G - will see if SP have improved 8 months later.

 

[zx128k]

I think the last time I experimented with this was in AGESA 1202? Back then it was accumulating Bus/Interconnects every second, now it only seems to come in waves when running tests, so maybe time has helped Fabric stability a little bit?

View attachment 245536



Yes, it's true that the Bus/Interconnects at 2000MHz Fabric are actually warnings/correctable errors. But it's a bit of a stretch to call that stable. The WHEAs that ailed my 3700X @ 1800Mhz and 1866Mhz were only "correctable errors". After that hellhole I learned to play it safe with Fabric. Before you ask, the RAM itself was 24 hours+ stable in HCI - any shenanigans were coming exclusively from the CPU.

Can't wait for my replacement 5700G - will see if SP have improved 8 months later.

Looking at the CR2, you can try for CR1.

AMD:

• Getting GDM disabled and CR 1 stable can be pretty difficult but if you've come this far down the rabbit hole it's worth a shot.
• If you can get GDM disabled and CR 1 stable without touching anything then you can skip this section.
• CR 1 becomes significantly harder to run as the frequency increases. Oftentimes, running CR 2 can help with achieving higher frequencies.
• On AMD, Gear Down Mode will override Command Rate. For this reason, disabling Gear Down Mode in order to set CR 2 may be beneficial to overall stability.

1. One possibility is to set the drive strengths to 60-20-20-24 and setup times to 63-63-63.
   • Drive strengths are ClkDrvStr, AddrCmdDrvStr, CsOdtDrvStr and CkeDrvStr.
   • Setup times are AddrCmdSetup, CsOdtSetup and CkeSetup.
2. If you can't POST, adjust the setup times until you can (you should adjust them all together).
3. Run a memory test.
4. Adjust setup times then drive strengths if unstable

• My stable GDM off CR 1 settings

1. Oftentimes, a drive strength above 24 ohms may hurt stability. Furthermore, running non-zero setup times is rarely needed, however may aid in the stabilization of CR 1.

 

[tabascosauz]

Looking at the CR2, you can try for CR1.

AMD:

• Getting GDM disabled and CR 1 stable can be pretty difficult but if you've come this far down the rabbit hole it's worth a shot.
• If you can get GDM disabled and CR 1 stable without touching anything then you can skip this section.
• CR 1 becomes significantly harder to run as the frequency increases. Oftentimes, running CR 2 can help with achieving higher frequencies.
• On AMD, Gear Down Mode will override Command Rate. For this reason, disabling Gear Down Mode in order to set CR 2 may be beneficial to overall stability.

1. One possibility is to set the drive strengths to 60-20-20-24 and setup times to 63-63-63.
   • Drive strengths are ClkDrvStr, AddrCmdDrvStr, CsOdtDrvStr and CkeDrvStr.
   • Setup times are AddrCmdSetup, CsOdtSetup and CkeSetup.
2. If you can't POST, adjust the setup times until you can (you should adjust them all together).
3. Run a memory test.
4. Adjust setup times then drive strengths if unstable

• My stable GDM off CR 1 settings

1. Oftentimes, a drive strength above 24 ohms may hurt stability. Furthermore, running non-zero setup times is rarely needed, however may aid in the stabilization of CR 1.

Did you just copy paste the memtesthelper guide? I've tried the whole shebang on both 5700G and 5900X, the CAD_BUS stuff doesn't always work without upping VDIMM. It was just a quick trip to 4000 grab some more up to date numbers, I usually run 3800.

 

[zx128k]

Did you just copy paste the memtesthelper guide? I've tried the whole shebang on both 5700G and 5900X, the CAD_BUS stuff doesn't always work without upping VDIMM. It was just a quick trip to 4000 grab some more up to date numbers, I usually run 3800.

Better than writing the story myself and then leaving something likely important out. Where you lucky enough to get IC's with a builtin temp sensor? I have to use the one with my multi-meter. In aida64 you can enable temperature monitoring of the DIMMs as you stress test if the ICs support it.

What temps are you all getting with your RAM tuning and overclocks. I use a 120mm CFM 63 fan to suck the hot air from between the DIMMs. I do this because In have four DIMMs and there is not much space between them.

Did you get a RAM kit rate to 1.5 volts or did take the risk without water cooling?

Voltages exceeding 1.45v is only recommended for Samsung B-die for daily use up to 1.5v. To consider is that DDR4 kits with a rated 1.5v also have PCB's that can run that voltage as a daily driver. Extreme voltages for daily use is only wise if you can water cool your DDR4 kit.

 

[Dia01]

Did you just copy paste the memtesthelper guide? I've tried the whole shebang on both 5700G and 5900X, the CAD_BUS stuff doesn't always work without upping VDIMM. It was just a quick trip to 4000 grab some more up to date numbers, I usually run 3800.

At least you're getting that far, I can't get it to boot @ 4000 no matter what

 

[tabascosauz]

Where you lucky enough to get IC's with a builtin temp sensor? I have to use the one with my multi-meter. In aida64 you can enable temperature monitoring of the DIMMs as you stress test if the ICs support it.

What temps are you all getting with your RAM tuning and overclocks. I use a 120mm CFM 63 fan to suck the hot air from between the DIMMs. I do this because In have four DIMMs and there is not much space between them.

Did you get a RAM kit rate to 1.5 volts or did take the risk without water cooling?

My dual rank has sensors but the two single rank B-die are Viper Steel so unfortunately don't. 4DIMM is hard to cool, sounds like you already got the cooling down pat

Some say 1.6 or 1.65 is okay for B-die depending on temps, but 1.65 daily waterblock is a good idea. That, or isolate the RAM from GPU heat (ie. sandwich case) and blast air at it.

At least you're getting that far, I can't get it to boot @ 4000 no matter what

It is a 4-layer board; 3933 dual rank is already excellent

 

[Dia01]

It is a 4-layer board; 3933 dual rank is already excellent

Yeah but its fun trying

 

[tjtremor999]

I think the last time I experimented with this was in AGESA 1202? Back then it was accumulating Bus/Interconnects every second, now it only seems to come in waves when running tests, so maybe time has helped Fabric stability a little bit?

View attachment 245536



Yes, it's true that the Bus/Interconnects at 2000MHz Fabric are actually warnings/correctable errors. But it's a bit of a stretch to call that stable. The WHEAs that ailed my 3700X @ 1800Mhz and 1866Mhz were only "correctable errors". After that hellhole I learned to play it safe with Fabric. Before you ask, the RAM itself was 24 hours+ stable in HCI - any shenanigans were coming exclusively from the CPU.

Can't wait for my replacement 5700G - will see if SP have improved 8 months later.

Any amount of hours passing in hci-memtest or any crazy config out there for another memtest is 100% worthless because those are not heavy loads on the SOC , IF,mem voltage & certain timings.

I can set crazy memory settings that pass 24h+ of hci-memtest that fail in Prime95 due to low soc voltage, low vdimm & too low timings. Those same settings would then fail gaming past the 6h mark or randomly reboot during desktop usage.

11h of Prime95 blend or large for memory validation is better on my side. Even a quick 1h-3h hci-memtest to validate new settings is a waste because Prime95 will catch those in 3mins.

 

[zx128k]

It can take over an hour for your memory to reach maximum temperature. B-die can read with an external K Thermocouple approx. 50c or above when overclocked. If you worry about errors over long stress tests being heat. Add a decent fan over the RAM, 4xDIMM you should do this. Temperature under load drops from 50+ to 22c and idle at 18c. Note this is a thermocouple stuck between two DIMM modules.

The best memory test appears to be TestMem5 v0.12 and a good custom config. extreme@anta777.cfg appears to work. Each test can be run individually 12 times and each error appears to match certain settings in BIOS. Was able to dail in more stable settings very quickly.

Also memory temperatures should be monitored. People who state b-die is not stable about 40c are not always correct. I have tighten settings and appear to work fine at higher temps below 85c over long stress tests. If temps are an issue there could be temperature sensitive settings.

I dont think one test rules them all. I have past hours of one test to fail prime 95 large ffts in one hour. Some memory problems can take 6-8 hours to sho in prime 95 but you can find the test that errors within a few minutes to a few hours in TestMem5 v0.12. Then focus on that test untill you find the setting in bios that passes that test. Making trounleshooting so much easier and quicker.

 

[tjtremor999]

These dual ccds are hard to get sub 53ns latency. For some reason forgot trtp on auto. Also turned off L2 HW prefetcher: -400MB/s bandwidth but gives out lower latency on my 5900x.

flat 16-16-16-16-33 50 at 4000 easily scores 52.x but wasn't stable on these kits

 

[zx128k]

These dual ccds are hard to get sub 53ns latency. For some reason forgot trtp on auto. Also turned off L2 HW prefetcher: looses ~400MB/s bandwidth but gives out lower latency on my 5900x.

flat 16-16-16-16-33 50 at 4000 easily scores 52.x but wasn't stable on these kits

View attachment 245767

Copy and latency appear to increase time spy cpu scores the most. I also use Cinebench r20 or 21 to see how performance is affected. There is also linpack. The goal is to tune for maximum performance and not for one Aida64 output.

 

[The King]

These dual ccds are hard to get sub 53ns latency. For some reason forgot trtp on auto. Also turned off L2 HW prefetcher: looses ~400MB/s bandwidth but gives out lower latency on my 5900x.

flat 16-16-16-16-33 50 at 4000 easily scores 52.x but wasn't stable on these kits

Impressive but this goes to show that the Patriot Viper 4000 CL16 kit is a much better bin than the 4400 C19. I also have seen the 4000 CL16 selling cheaper than 4400 C19 kits on both Amazon and on Newegg.

https://www.amazon.com/Patriot-4000MHz-16-16-16-36-Low-Latency-Memory/dp/B09NLBNLZ4/ref=pd_lpo_1?pd_rd_i=B09NLBNLZ4&psc=1

I would grab one of those 4000 CL16 if it was available locally but no surprise its not! Probably would cost more than the C19 kit here anyway.

 

[tjtremor999]

Copy and latency appear to increase time spy cpu scores the most. I also use Cinebench r20 or 21 to see how performance is affected. There is also linpack. The goal is to tune for maximum performance and not for one Aida64 output.

Sure, Aida64 numbers don't really represent any real workload for memory bandwidth or latency

sisoftsandra for example under 4000cl16 62.4GB/s total memory bandwidth(Aida numbers) will score something like 5GB/s or less per core available.

5GB/s x 12 cores doesn't equal 62.4GB/s total from Aida64

I didn't do game fps or 3dmark runs yet but I don't suspect anything will show between these:

3800cl14 ~54.x latency ~59GB/s
4000cl16 ~52.x latency ~ 62GB/s with 1ns less for ccd to ccd latency
4000cl16 ~53.x latency with ~62.4GB/s (L2 HW Prefetch -on)

 

[zx128k]

The King is correct that 4000 CL16 is better than 4400 CL19 for latency. There are also 4400 CL16 kits, I believe G-skill makes them. (16-19-19-39) at 1.5V F4-4400C16D-16GTZR
View attachment RAM-Latency-Table.webp

Sure, Aida64 numbers don't really represent any real workload for memory bandwidth or latency

sisoftsandra for example under 4000cl16 62.4GB/s total memory bandwidth(Aida numbers) will score something like 5GB/s or less per core available.

5GB/s x 12 cores doesn't equal 62.4GB/s total from Aida64

I didn't do game fps or 3dmark runs yet but I don't suspect anything will show between these:

3800cl14 ~54.x latency ~59GB/s
4000cl16 ~52.x latency ~ 62GB/s with 1ns less for ccd to ccd latency
4000cl16 ~53.x latency with ~62.4GB/s (L2 HW Prefetch -on)

Aida64 measures some cache performance in the memory tests. For Aida64 with 4000 memory its possible to get above the maximum of 64000 as a result.
5GB/s x 12 cores doesn't equal 62.4GB/s total from Aida64 but it does get very close at 60GB/s. Remember cache appears to be counted in Aida64.

sisoftsandra is different. Note the lower latency value on the memory.

Memory Bandwidth	47.50GB/s
Data Cache/Memory Latencies (In-Page Random Access)	19.9ns
Data Cache Bandwidth	465.91GB/s

MemMax2 is different results as well, even lower than sisoftsandra at approx. 41GB/s copy and >30GB/s read and write. Latency is different as well.

Linpack appears to be a great way to measure performance after changing settings.

 

[QuietBob]

Benching for a baseline before Zen 3 upgrade: