Ryzen DRAM Calculator by 1usmus

[Cowboystrekk]

This is what I mean @Taraquin, I had em tighter but scored worse when I went to CL16, over 100 sec, hence why I said its interesting


These changes brought it back

tRTP should be 1\2 tWR, that timing is tighter in your second atempt, it might be the silver bullet tRRDL is best at 4 and is barely affected by voltage, tWTRL sits better at 12 than 14. Also your tRAS and tRC-equation is off at the first atempt and might degrade performance, due to a possible latency penalty as something is "off" clockwise

 

[jesdals]

Try tRAS=tCL+tRCDRD and tRC= tRAS+tRP. Then try tRFC=tRCx6. tRRDL works better as 4 or 6, tWTRL should be 12, 10 or 8, tWTRS should be 4 or 3. tWR 12, tRTP 6. Try that and see how it goes

Its a bit odd that raising tRFC actually gives a better result. Will test stability at these settings

 

[Cowboystrekk]

View attachment 196938

Its a bit odd that raising tRFC actually gives a better result. Will test stability at these settings


View attachment 196939

Might be due to tRC multiplier. If tRC doesnt add up to tRCF that can cause recharges etc to be off. It doesn't have to be a problem, but it can sometimes be a problem.

 

[jesdals]

Decided to try to crunch both settings with Aida64

With the tRFC 240 settings i got latency at 54.4ns

with tRFC 252 and the matching settings i got 54.5ns

so more or less the same - will do some testing of stability with the later

 

[Cowboystrekk]

Decided to try to crunch both settings with Aida64

With the tRFC 240 settings i got latency at 54.4ns
View attachment 196947

with tRFC 252 and the matching settings i got 54.5ns
View attachment 196948

so more or less the same - will do some testing of stability with the later

Your read speed, L2 and L3 cache results are slightly better at the second config

 

[Space Lynx]

Decided to try to crunch both settings with Aida64

With the tRFC 240 settings i got latency at 54.4ns
View attachment 196947

with tRFC 252 and the matching settings i got 54.5ns
View attachment 196948

so more or less the same - will do some testing of stability with the later

@tabascosauz

any idea why his L3 cache is so much better than mine? any recommendations on changes I should make? its stable so far... temps are good too.

 

[Zach_01]

Latency aside, every other (Read, Write, Copy) is tied to number of cores/threads CCXs/CCDs.

 

[Space Lynx]

Latency aside, every other (Read, Write, Copy) is tied to number of cores/threads CCXs/CCDs.

do my numbers in zentimings look ok overall? I was sort of winging it LOL

 

[Zach_01]

do my numbers in zentimings look ok overall? I was sort of winging it LOL

Well, we cant see it all, but from what it is shown I think L1/2/3 are ok.
But I would expect Memory bandwidth to be higher (Read= ~55000MB/s) for 3800MHz DRAM frequency

This is an old AIDA test with my R5 3600

 

[thesmokingman]

Well, we cant see it all, but from what it is shown I think L1/2/3 are ok.
But I would expect Memory bandwidth to be higher (Read= ~55000MB/s) for 3800MHz DRAM frequency

This is an old AIDA test with my R5 3600

View attachment 201284

L3 cache speeds are not accurate on the earlier aida benches and its not until the recent bios that fixed cache readings specifically in aida that that bench is now meaningful with regards to L3.

 

[Zach_01]

When you say earlier? It is 18months old.

And this is 7 months old
I don't have a recent one with 3800MHz

 

[thesmokingman]

When you say earlier? It is 18months old.

And this is 7 months old
I don't have a recent one with 3800MHz

View attachment 201287

Read my post. Per the bios fix for AIDA, that means it wasn't accurate before said bios.

For ex.

Version 3602

2021/03/12 20.29 MBytes

ROG STRIX X570-E GAMING BIOS 3602
"- Update AMD AM4 AGESA V2 PI 1.2.0.1
- Support Smart Access Memory for Ryzen 3000 Series Processors
- Fix the L3 cache performance in AIDA64
- Improved system performance

 

[Space Lynx]

Read my post. Per the bios fix for AIDA, that means it wasn't accurate before said bios.

For ex.

so my scores are ok considering I have latest updates across the board?

 

[Makaveli]

Your scores are fine.

This is what I see on my system with the updated bios which includes the aida fix.

Single CCD chips always score lower than Dual CCD chips in this benchmark.

 

[jesdals]

any idea why his L3 cache is so much better than mine? any recommendations on changes I should make? its stable so far..

The difference i L3 cache between 5900/5950 and the 5800/5600 makes it impossible to compare directly.

 

[Space Lynx]

How do I know if I am running 1:1 in real time? My FCLK is 1900 in BIOS, 3800 ram speed in BIOS, but I noticed in CPU-Z my ram speed is reported as 1899.6mhz does this mean Infinity Fabric is broken since it doesn't get read perfectly even? Or am I overthinking it?

 

[Tomorrow]

How do I know if I am running 1:1 in real time? My FCLK is 1900 in BIOS, 3800 ram speed in BIOS, but I noticed in CPU-Z my ram speed is reported as 1899.6mhz does this mean Infinity Fabric is broken since it doesn't get read perfectly even? Or am I overthinking it?

CPU-Z reports MT/s not Mhz. Since DDR is Double Data Rate. 1900 x 2 is 3800. So that's correct. As long as Zentimings reads all 3 as 1900 then you're fine. Beside if you were not at 1:1:1 the latency would be far higher than 54,9ns.

 

[Space Lynx]

CPU-Z reports MT/s not Mhz. Since DDR is Double Data Rate. 1900 x 2 is 3800. So that's correct. As long as Zentimings reads all 3 as 1900 then you're fine. Beside if you were not at 1:1:1 the latency would be far higher than 54,9ns.

thanks! that makes sense.

I ran aida64 stability test, prime95, memtest, cpu-z stress and bench, 3dmark timespy stress, everything passed with flying colors. excellent.

one interesting thing though, my 3dmark scores were a little lower, and the only variable that change was my 3200 cas 14 ram going to 3800 cas 16, but all the synthetic benches like aida64 latency test were better on 3800. weird. i'm leaving it at 3800 anyway. I lost about 700 points in timespy with that single variable change. doesn't make sense, but oh well

 

[GorbazTheDragon]

Pretty sure you should be able to run 14-14/9-14 (CAS-RCDRD/RCDWR-RP) primaries on that stuff, might need closer to 1.5v or so but it's b-die so it will run that fine...

I'd tighten up RAS, RRDs, FAW, and WTRL, you should also be able to drop RFC down to 275 or so. DRAM calc has a useful tool for calculating RFC2/4.

 

[tabascosauz]

@tabascosauz

any idea why his L3 cache is so much better than mine? any recommendations on changes I should make? its stable so far... temps are good too.

Like others have said, single CCD. I'd be much more concerned with the bandwidth, 3800 should be somewhere in the 53-55GB/s range on R and W. tRRDS to tWR are loose, but I don't think I've ever pulled that much bandwidth gain out of any combination of timings alone.

You sure Infinity Fabric is actually stable? How long is "so far"? My Renoir always does all sorts of strange shit in memory benches when it's running on the IF edge.

If you don't have paid AIDA (honestly worth just buying a licence), you can run each of the tests (mem read, write, copy, latency) individually from the main AIDA window, you won't need to gamble on what the TRIAL VERSION text covers.

Not sure why VDDP is so high. You sure you actually need it that high? In relation to your VSOC you shouldn't need more than 0.9V or 0.95V.

As for "memtest", I'm guessing you tested with HCI? How much coverage?

 

[GorbazTheDragon]

Not sure why VDDP is so high. You sure you actually need it that high? In relation to your VSOC you shouldn't need more than 0.9V or 0.95V.

RE voltages in general:

SOC has a big range 980-1280mV, lots of chips appear to not like much over 1050mV though, often audio crackling/hitching and high dpc latency are symptomatic of too much SoC.

VDDG IOD has similar scaling behaviour to Matisse, typically between 980-1080mV it should be good

VDDG CCD has negative scaling beyond 950-980mV, scales positively from at least 850mV to 950.

cLDO_VDDP limits of 1050mV are the same, 850-900mV being optimal, over 950mV a waste.

 

[Zach_01]

How do I know if I am running 1:1 in real time? My FCLK is 1900 in BIOS, 3800 ram speed in BIOS, but I noticed in CPU-Z my ram speed is reported as 1899.6mhz does this mean Infinity Fabric is broken since it doesn't get read perfectly even? Or am I overthinking it?

You are. You can see it on ZenTimings and HWiNFO.

All the following 3 software are on the same time. All are reporting the same with slight differences with each other, but its 1:1:1.

 

[Cowboystrekk]

thanks! that makes sense.

I ran aida64 stability test, prime95, memtest, cpu-z stress and bench, 3dmark timespy stress, everything passed with flying colors. excellent.

one interesting thing though, my 3dmark scores were a little lower, and the only variable that change was my 3200 cas 14 ram going to 3800 cas 16, but all the synthetic benches like aida64 latency test were better on 3800. weird. i'm leaving it at 3800 anyway. I lost about 700 points in timespy with that single variable change. doesn't make sense, but oh well

Did you set 2th/3th timings on auto or manually tuned? On auto they tend to be very lax, but with manual you can run almost every timing except the 5 primaries and tRFC the same at 3200 as 3800.

 

[bfollett]

Spent way more time than I expected on this. Ended up going back to version 1.7.0 of Dram Calculator to get these timings for G.Skill f4 3600c16 8GVKC ram now running at 3400 cl14. It's Hynix cjr or djr. The older version of Thaiphoon burner reported djr, the current version says cjr. How does this look. I couldn't seem to get above 3400 even at reduced cas latency. I have 2600x cpu so 3400 may be the highest frequency my memory controller is going to handle with this ram/motherboard (AsRock B450m Pro4).

 

[Cowboystrekk]

Spent way more time than I expected on this. Ended up going back to version 1.7.0 of Dram Calculator to get these timings for G.Skill f4 3600c16 8GVKC ram now running at 3400 cl14. It's Hynix cjr or djr. The older version of Thaiphoon burner reported djr, the current version says cjr. How does this look. I couldn't seem to get above 3400 even at reduced cas latency. I have 2600x cpu so 3400 may be the highest frequency my memory controller is going to handle with this ram/motherboard (AsRock B450m Pro4).
View attachment 201619

Looks good. Newest bios? You could try tRTP 6 or tWR 10 and tRTP 5, tWTRL might do 10 or 8. Have you tried lower tRP or is 19 lowest?