3800X build bad performance - what am I doing wrong?

[TheLostSwede]

Wait if I read this correctly. You have 2x16gb sticks of ram as in 32 gb total. Then it's normal for it to take longer to boot.

A few second maybe, not a minute.

 

[ecopsorn]

Wait if I read this correctly. You have 2x16gb sticks of ram as in 32 gb total. Then it's normal for it to take longer to boot.

no, not anymore - I have 4x8 G.skills now 3600-16-16-16-32. - With the old windows install I did have the same 60sec boot time

 

[tabascosauz]

@ecopsorn those CB R20 runs are all within a roughly 50pt margin of error, and the scores look fine. Ryzen 3000 can be pretty unpredictable in CB R20 scores from one run to the next. How are your CPU temps and drive performance looking now?

If your load temps are still up in the 80s, there may be a bit of room for those CB scores to improve. Whether that's a worthy endeavour, I don't know. We would be talking about a few dozen points, maybe 100pts at the very most. And no noticeable gain in games. Same goes for PBO, maybe a 200pt gain at most if you are willing to spend the time to optimize your PBO parameters; nothing significant like the massive 600pt improvement from getting rid of all of that background crud.

For 16-16-16-36, the latency still seems a tad bit high, though absolutely no concern compared to your horrendous results before reinstalling. Can you run this program and get us a screenshot of your subtimings? https://github.com/irusanov/ZenTimings/releases

 

[ecopsorn]

@TheLostSwede @tabascosauz
I've installed the nvme driver now and Samsung Magician.
Look at those values now!! - they seem alright, don't they?

here are timings - I didn't load up DRAM calc yet:

These are temps now:

IDLE (CPU temp jumps between 35 and 47 constantly right now)

Runninc CB20

 

[tabascosauz]

@ecopsorn entirely normal idle temperatures on the CPU, then. That's good news. Also, use HWInfo for monitoring. Load temperatures look good at 65, no issues there. Now that's what a NH-D15 is really capable of, even with just one fan.

Well, you have the rather high-end (and expensive(!)) B-die kit now, so now you can focus on working on those timings now. For good quality B-die at 3600, DRAM calc is recommending 14-15-14-30 or 14-15-14-28. The former setting has a tRFC setting of 468, which is rather loose for B-die. The latter's 288 tRFC is more in line with what B-die can do, but the former is probably a more surefire way to boot. Either is quite a bit better than the stock 630 tRFC. You could start just by lowering tRFC and seeing what the does for your latency. The RDC recommended profiles also suggest higher 1.37V.

I think 600+ tRFC usually comes in XMP alongside a 2T command rate. 600+ is really, really high, especially for B-die.

EDIT: Actually, RDC lists profiles for A0, lower quality B-die as well. Those are 16-17-16-34 and 16-17-16-32. Much closer to your XMP, and a safer place to start. Just do the tRFC for now.

Those profiles suggest 468 and 302 respectively for tRFC. Maybe give those a try.

 

[ecopsorn]

468 gives me error 8d when I leave the timing on 16-16-16-32
I tried 500 instead and error didn't show up. This is the Aida bench with this, and 4905pts cb20:
Bringing the latency down should help the score as well, right?

 

[tabascosauz]

468 gives me error 8d when I leave the timing on 16-16-16-32
I tried 500 instead and error didn't show up. This is the Aida bench with this, and 4905pts cb20:
Bringing the latency down should help the score as well, right?
View attachment 155471

If you leave HWInfo monitoring in the background for some time while you're benching, what does the DRAM voltage show under your board sensors section? Does it sometimes max out above 1.35V, does it drop below 1.35V at times?

If that's with 500 tRFC, that's already looking much healthier a latency result for B-die at 3600. It won't affect CB much, CB likes clockspeed much more than RAM speed. B-die should be doing better than that on tRFC, but if it's stable you can just leave it at that for now. Maybe give the 3600/B-die/A0/Safe profile a try with all of its timings suggestions (the 16-17-16-34 one) and see if it'll be stable.

Are the default 16-16-16-36 settings stable through other tests like P95 Large, MT86, membench?

 

[ecopsorn]

If you leave HWInfo monitoring in the background for some time while you're benching, what does the DRAM voltage show under your board sensors section? Does it sometimes max out above 1.35V, does it drop below 1.35V at times?

If that's with 500 tRFC, that's already looking much healthier a latency result for B-die at 3600. It won't affect CB much, CB likes clockspeed much more than RAM speed. B-die should be doing better than that on tRFC, but if it's stable you can just leave it at that for now. Maybe give the 3600/B-die/A0/Safe profile a try with all of its timings suggestions (the 16-17-16-34 one) and see if it'll be stable.

Are the default 16-16-16-36 settings stable through other tests like P95 Large, MT86, membench?

It's funny, another guy said don't use hwinfo haha.
Yes results above are with 500.
Ok, I will try the A0/Safe Profile - but can you tell me what all to enter? Just the basics (values until and incl. tRC (as well as tRFC and tRFC (alt) and the Rec. voltage? - So many ppl recommended also to adjust SOC voltage and in other forums I read that procODT is also important and I about the power/gear down mode thing I'm completely lost

EDIT: Cannot tell yet if they are stable, haven't run any long time tests yet. Do you recommend a short one to see if it's stable until I can ran a membench overnight?
EDIT2: I punched in the safe values - it did give me an 8d at first but after shutting pc down it booted normaly

Latency starting to get nice

 

[moproblems99]

what is not fine is a 47degree CPU in bios

That isn't really a problem per se and I wouldn't worry about that right now.

But good to see that everything seems to be heading in the right direction.

It's funny, another guy said don't use hwinfo haha.

I don't have any problems with hwinfo so not sure why not to use it. It gives you lots of useful information.

 

[ecopsorn]

But good to see that everything seems to be heading in the right direction.

Indeed I think it starts to get interesting. Already wanted to ditch the mobo for a Gigabyte Aorus Xtreme

 

[TheLostSwede]

@ecopsorn SSD is looking a lot better, but still not quite there. That said, I don't know how their software tests, so please run CrystalDiskMark again and throw that up here.
You can change to IOPS where it says MB/s.

HWInfo is good, take no notice of people dissing it, however, any software that checks the CPU clock speed is going to agitate the CPU, which means that the clock speed is going to change slightly, or maybe even by a lot for a few milliseconds. This is why only Ryzen Master is suggested, but the logging isn't very helpful there in comparison.

Temps are looking fine.

Whatever went wrong last time around seems to be resolved now.

 

[tabascosauz]

Indeed I think it starts to get interesting. Already wanted to ditch the mobo for a Gigabyte Aorus Xtreme

Bravo! You made it below 68ns. Now we're most definitely into B-die territory at 3600.

Only thing I have against HWInfo is that it does poll your CPU every time it updates, which brings Ryzen 3000 out of idle for a split second (but that's just a Matisse-too-smart, software-too-dumb thing again) so over time your CPU activity looks a bit like an electrocardiogram, lol. It's a great tool to have for overclocking and testing. I just don't run it all the time, as with any other monitoring utility, if my system is already 100% stable and where I want it. You've got a ways to go, so HWInfo is perfect.

If your system doesn't hang or crash for some odd reason, you'd probably want to enter as many subtimings as you can out of that middle section of the window in DRAM calc. But tCL, the 2 tRCDs, tRP, tRAS, tRC, tRFC and Command Rate are the important ones. I dare to say tFAW should also be among those important timings, but I'm personally unfamiliar with the rules regarding tFAW and what what relationship it must have to other timings, etc.

How's that DRAM voltage looking in HWInfo?

Alternatively, instead of using DRAM calc, you can follow this guide: https://github.com/integralfx/MemTestHelper/blob/master/DDR4 OC Guide.md#tightening-timings

It's got lots of info, and is geared towards higher end kits such as B-die, and appears to be up to date for recent developments. It is a lot to take in, though. But if you want to start probing the limits of your B-die, this is probably the way to go.

 

[ecopsorn]

How's that DRAM voltage looking in HWInfo?

Ok getting into the settings.
But I cannot find DRAM voltage in hwinfo - is it called differently?

@ecopsorn SSD is looking a lot better, but still not quite there. That said, I don't know how their software tests, so please run CrystalDiskMark again and throw that up here.
You can change to IOPS where it says MB/s.

HWInfo is good, take no notice of people dissing it, however, any software that checks the CPU clock speed is going to agitate the CPU, which means that the clock speed is going to change slightly, or maybe even by a lot for a few milliseconds. This is why only Ryzen Master is suggested, but the logging isn't very helpful there in comparison.

Temps are looking fine.

Whatever went wrong last time around seems to be resolved now.

Here we go - is DiskMark64 the right one and does it matter what I choose in the dropdown? 5/1GiB?
As I wrote, I didn't switch the M2 yet to second position because of all the talk about the thermal pads. I'm too lazy right now to take the GPU out again to check the thermal pad in the first slot, but let me know if I should do it anyway.

 

[TheLostSwede]

Ok getting into the settings.
But I cannot find DRAM voltage in hwinfo - is it called differently?



Here we go - is DiskMark64 the right one and does it matter what I choose in the dropdown? 5/1GiB?
As I wrote, I didn't switch the M2 yet to second position because of all the talk about the thermal pads. I'm too lazy right now to take the GPU out again to check the thermal pad in the first slot, but let me know if I should do it anyway.
View attachment 155473

IOPS are still way off, no idea what's going on here. Obviously as this is the OS drive, they will be lower than the numbers Samsung claims, but this doesn't look quite right on a brand new install of Windows.
I'm afraid I'm out of suggestions though, but moving the drive isn't going to help you, as then it'll go via the chipset instead of being connected directly to the CPU, which should be the slot that offers the best performance.
I guess it's possible that Samsung over hyped the IOPS in their marketing materials, but not by that much.
Based on this review, at least in Anvil's Storage Utilities, the drive is capable at numbers close the marketing numbers if you look at the 4K QD16 results.

Samsung 970 EVO Plus 1TB SSD Review - Legit Reviews

www.legitreviews.com

I have a slower OEM Samsung drive in this loaner computer, and I get better write results than you do.

 

[tabascosauz]

@ecopsorn DRAM voltage on mine is right before system and CPU fan speeds. It's not in the same block as the CPU sensors.

 

[ecopsorn]

IOPS are still way off, no idea what's going on here.

damn, that sucks

@ecopsorn DRAM voltage on mine is right before system and CPU fan speeds. It's not in the same block as the CPU sensors.

I managed to get the FAST settings on A0 running, still at 1.35V:

But let me ask this: in another guide I read, it is better to first raise frequency with loose timings until no POST is possible anymore and than tweak timings? Shouldn't I try to get higher frequencies before messing with the timings on 3600?

man I'm still trying to find that dram voltage in hwinfo, gonna google it lol

apparently - my mobo is not able to readout dram voltage https://www.hwinfo.com/forum/threads/missing-dram-voltage-asus-strix-x570-e-nuvoton-nct6798d.6079/

 

[Waldorf]

@ecopsorn
turn gear down/power down for ram to off/disables.
from rare EXCEPTIONS, its usually better to turn them off, same for fast boot.
the 2s you save will be nothing if u have probs and need to troubleshoot etc, and can't get "out of it"..

 

[tabascosauz]

I managed to get the FAST settings on A0 running, still at 1.35V:

But let me ask this: in another guide I read, it is better to first raise frequency with loose timings until no POST is possible anymore and than tweak timings? Shouldn't I try to get higher frequencies before messing with the timings on 3600?

man I'm still trying to find that dram voltage in hwinfo, gonna google it lol

Just keeps getting better, huh?

Each board reports its sensors differently, maybe yours isn't visible to HWInfo? Maybe another monitoring software will report DRAM voltage for you, but the nice thing about HWInfo is that it reports all the little fluctuations in the voltage. If it can't read out DRAM voltage, just always make sure you're setting the right voltage in BIOS.

No rule says that you have to max out the frequencies. If you were on an Intel platform with a high-end SKU, you'd already be shooting for 4400-4800MT/s already if you wanted to push the kit to the max. We can't do that on Ryzen, so we're limited to speeds like 3733 and 3800 max, where the kit can do much more than the CPU and its memory controller allows. But as we know, 3600 is the sweet spot that 99% of Ryzen 3000 CPUs can do, whereas 3733 and 3800 are progressively more difficult for all CPUs to hit. And getting to 3800 for example introduces stability concerns because it's right on the edge of what the IMC is capable of; sometimes, on a less-than-stellar day, you may run into stability problems.

Some people don't even go past 3200. They spend their days with a crazy good B-die kit getting it down to CL12 at 3200, which is pretty damn tight (and also into the territory of 1.5V DRAM voltage, which B-die just eats up no problem, but is temperamental as hell). And some others like us just stay at 3600. With good B-die, I'm confident you can probably do CL14 at 3600. Hell, there are B-die owners on Ryzen out there that are doing 3800 CL14. But it's up to you what you want to do. There's a million ways to get there.

As to Gear Down Mode, it usually should be disabled, but it won't make a difference anyways. It only does anything if your tCL or CAS latency is an odd number, whereupon it'll kick that tCL to the next number up (say 17 will become 18) for stability. That's all it does. Always have Power Down Mode disabled, however.

 

[TheLostSwede]

Simply put, there are two ways to get performance out of RAM, higher clock speeds and lower latencies, but one tends to negate the other.

Ryzen 3000 CPUs are a bit of a peculiar beast, in that the Infinity Fabric runs at the same clock as the RAM (the RAM frequency you're looking at is x2, due to Dual Data Rate or DDR). The IF can only run so fast until you start having other problems and that peak seems to be around 1900MHz on a good chip. As such, 3800MHz is as fast as you really want to run your RAM at, as otherwise the IF ends up clocking down. For example, if you were to run your memory at 4000MHz, the IF will run at a mere 500MHz. This might actually reduce the performance in a lot of programs, although some things seemingly care more about memory clocks than anything else, but in reality, most programs like a balance.

AMD obviously went out with 3200MHz as the official top memory clocks speed, but said that up to 3600MHz should be doable quite easily on most CPUs and even put out a graph showing it as the sweet spot, with 3733MHz being where it topped out before the IF dropped down a gear. However, improvements to the AGESA has shown that the limit is 3800MHz if you've got the right components and got a little bit lucky in the silicon lottery.

Does it make sense to hit 3800MHz? Maybe not, but once you're happy with how your system performs, it's something you can tinker with. For now, I'd leave it where you're at for a week or so, play around with all your software and make sure everything works well. Then you can start playing around and see what works out best for you. It might in fact be better with slightly lower clocks and tighter timings, as sometimes that works out as the optimal balance. I can't go down in latency with my RAM, so I went up in clocks instead.

No rule says that you have to max out the frequencies. If you were on an Intel platform with a high-end SKU, you'd already be shooting for 4400-4800MT/s already if you wanted to push the kit to the max. We can't do that on Ryzen, so we're limited to speeds like 3733 and 3800 max, where the kit can do much more than the CPU and its memory controller allows. But as we know, 3600 is the sweet spot that 99% of Ryzen 3000 CPUs can do, whereas 3733 and 3800 are progressively more difficult for all CPUs to hit. And getting to 3800 for example introduces stability concerns because it's right on the edge of what the IMC is capable of; sometimes, on a less-than-stellar day, you may run into stability problems.

One small point here, you can clearly go much higher than that on a Ryzen CPU in terms of the memory speed, it just doesn't make sense from a platform performance standpoint. I mean, one of the pro guys his 6024MT/s on some Micron/Crucial memory back in October (albeit with terrible latencies) on a Ryzen 3600X. Is it practical? Absolutely not, but someone did it, because it was possible.

Micron Memory Sets New DDR4 Overclocking World Record

Ballistix, Micron's global brand of high-performance gaming memory, has set a new overclocking world record for the fastest DDR4 memory frequency at 6024 MT/s. Leveraging performance-tuned Micron die and the innovation behind the new Ballistix Elite 4000 memory, the ASUS motherboard R&D team set...

www.techpowerup.com

 

[ecopsorn]

Appreciate it - will see with the ram. Somehow I already can't get the A0 FAST to work anymore - gives me F9 error into 02 and I gotta restart, than it applies JEDEC.
Before I go any further now - there was a third important bench, heaven 4.0
I just did a run and holy cow, it is even worse than with the old windows

I mean this value is embarrassing. - second run is with "prefer optimal performance" set in nvidia control panel. still like 30fps lower than to be expected!

 

[Waldorf]

it only shows 4gb vram. did you check other progs if they show correct 8gb?
i think it only looks worse as the min fps drags all numbers down.
unigine is a bit different than 3DMark and the like, as it can dip pretty hard when you start benching right after tool starts.
let it run until fps "levels out a bit", then hit "bench".

 

[tabascosauz]

@ecopsorn I think the 4GB VRAM thing might be a bug that is also shared with other, different cards. Doesn't matter, because your scores are less than half of what a 2070 Super should do in Heaven 4.0.

Does GPU-Z monitoring in the background while you bench show anything peculiar going on with your GPU clockspeeds, voltages, temps, or perfcap reasons?

You're on the latest WHQL driver, I'm guessing?

 

[ecopsorn]

it only shows 4gb vram. did you check other progs if they show correct 8gb?
i think it only looks worse as the min fps drags all numbers down.
unigine is a bit different than 3DMark and the like, as it can dip pretty hard when you start benching right after tool starts.
let it run until fps "levels out a bit", then hit "bench".

problably a bug, in cpu-z it is 8.

The heaven scores might actually be normal. I didn't realize that most test at 1920x1080, so the score there is this (all stock so far - no OC)

so it is damn late again here - appreciate all the help @Fry178 @tabascosauz @TheLostSwede

to sum it up - please correct any rumbling down here if it is wrong, no idea how I'm still able to type:

• booting time issue seems resolved - (gonna closely watch it during my program installs. gonna keep fast boot disabled until all is installed
• temperatures of CPU is still not great but acceptable under load
• Samsung evo has too little IOPS - so far no solution
• GPU bench not sure cause I'm tired as f
• rams still need tweaking but it looks like that the latency is where it should be
• I still don't have PBO enabled and the common understanding in this thread is that it should be kept disabled??
• chuck norris doesn't have an ESC key on his computer
• I should not try to manually OC CPU cause it won't do much
• I should stick to 3600 so far and try to tighten timings.

The goal of this long troubleshooting today was to figure out whether I need a new motherboard or PSU or ssd and whether I should warranty claim something. To be honest, I'm still not sure, not sure if another better mobo would solve more issues or if it would be just a wast of money.
I'd appreciate your opinion.

good night

@ecopsorn I think the 4GB VRAM thing might be a bug that is also shared with other, different cards. Doesn't matter, because your scores are less than half of what a 2070 Super should do in Heaven 4.0.

Does GPU-Z monitoring in the background while you bench show anything peculiar going on with your GPU clockspeeds, voltages, temps, or perfcap reasons?

You're on the latest WHQL driver, I'm guessing?

not sure what I should look at, here's the screen while having heaven running - ofc latest driver 445.87 - I don't even have my two other monitors plugged in yet and no other software installed

 

[Waldorf]

@ecopsorn
sounds good so far.

-set your min cpu perf to 20-30%. will lower volt/temps, but not really impact fps.

-you only need to install the samsung driver, magian only if you want to update FW, dont use any caching (or have lots of fun recovering the drive).
test with stock settings of the tools, 1gb file size.

250gb 970 pro:






-haven't tested heaven past 7xx series.
run valley or better superposition (both 1080p/extreme, change to custom if the stock setting wont match that).
did u install gfe/shadowplay? (usually eats a bit fps even if not recording (but present/buffering).

3700+2080FTW hybrid on stock clocks and "dirty" win:


-you're right, the computer has a chuck norris key

 

[ecopsorn]

Won’t have time until evening to test. 11am here. I will dl the other benches.

this is probably also an important info: this 2070s is already the second 2070s I have in the sysrem. The first one was too loud. With both cards I got almost the same results!