Originally Posted by damric
We must think alike and read the same stuff. I actually already have that memory guide bookmarked.
So mostly I've been stress testing, using linpacks for quick tests, and AIDA64 for longer tests. Linpack seems to eat my CPU and VRM, and quickly shows if my CPU clock is unstable. AIDA64 seems to be a better indicator of if my memory is unstable. Right now I seem rock solid and good temperatures with just 3825MHz CPU and 3200CL14 RAM. I'm still testing though. I might be able to get a little bit more at the costs of a lot of volts and heat.
What's cool is the way I have my fans on my tower heat sink oriented. I had extra room enough that I was able to push them way down hanging off the bottom of the cooler so that they nearly sit flush on my motherboard. This way the Push fan also pulls air over the RAM, and the Pull fan pushes air directly onto the VRM.
This motherboard has extensive sensors to monitor, and lets me assign fan speed curves to the various temperatures, so I have my Pull fan and my rear exhaust ramp up according to the VRM, and I set all the others according to CPU. Very quiet if I'm not benching or stressing. These are some 2500RPM PWM fans I bought for the CPU cooler a few months ago with supposedly over 4mmH2O of static pressure. I also have a 1200RPM Cougar for the rear exhaust, and 2 Cooler Master Air Flow 140mm PWM fans in the front intake.
Holy cow man, dust that HD 7850!
Forreals though wipe that case down and that fan filter next to your drive cage behind the shroud... will probably help temps... pull it out and wash it in the sink and let it dry overnight by a box fan or something if you don't have alcohol/electronics wipes/canned air. If you have a Shop Vac that you can set to blow air those work well too.
I see what you mean about the fans on your cooler being offset to be flush with the board. Yes, this would help cool the memory. Have you ever tried orienting the sink the other way (facing up/down)? This is generally doable with most tower coolers today. The idea is, you orient it with the heatsink fans aligned up/down, then put your fans push/pull, as they are now, with the bottom fan pulling and the top pushing, and the air moving up through the sink and out the top of the case. Since heat rises. Since you have an NVMe drive, if it is directly above the GPU, this will cool the NVMe drive as well as the GPU since heat naturally rises anyway. The only problem is that your CPU may run a few C hotter since now, heat from your graphics card will be passing through the sink too. But overall you may not see much difference other than your NVMe and GPU will run cooler, since having the sink the way you have it now with the fans moving air front to back causes a dead zone of hot air from the GPU (when gaming or doing 3d benches) directly below the heatsink, right over where most boards have NVMe drives now (and a lot of them can silently corrupt if they run above 70C or so, which is their TJmax- my 970 Evos is, anyway). Also, the hot air from your GPU not only will hang around the board socket/NVMe/right above the card but also heat the heatsink directly, on the metal on the bottom of it currently, since there are no fins or fans on the bottom blowing up.
So, how the whole thing would work would be you have your heatsink oriented top to bottom, bottom fan on the heatsink blowing up, and top fan pulling hot air up. This will cool the GPU and NVMe better. You keep the fan on the back of the case the same, and it will exhaust heat from the VRMs and socket air through the back of the case via pulling it. Then, to cool the memory and socket area, as well as your NVMe drive further, you can use one of the G.skill memory coolers if you have one, blowing on the DRAM next to the tower. Or stick a spare 120mm fan over the memory if you have a way to power it, and use some zipties or something to rig it, so it is blowing down onto the memory and board to the right of the heatsink.
I haven't used a tower cooler for a very long time and use 240mm AIOs but have similar issues, especially since moving to the R9 3900x, with stagnant heat around the socket and above my GPUs- remember, OC'ed they can pull close to 750w a piece- even when they are sitting idle, they use 100w+ and are constantly exhausting heat up. So, I put a spare G.skill memory cooler blowing down on my memory, and a spare 140mm fan behind the socket, ziptied around some cords, blowing on it directly from behind. This reduced load temps on the CPU by roughly 10-15C depending on load, and caused my NVMe to idle around 35c instead of 50c (and under a full bench load with my system pulling 850-900w at the wall I'm sure my Ti's were heating that thing enough to the danger zone of 70c+ easily, but I never checked)
Anyway, some other people really swear by this cooler orientation method, though again it can make the CPU run hotter but the GPU and NVMe run cooler
.. I wouldn't change it though unless you have that spare fan and some way to have it blowing down onto your memory, which will also cool the socket, VRMs, NVMe etc and I would also do thorough before/after tests on all components but especially the CPU and GPU, while controlling for ambient temps if possible (set your heater to the same setting and use an indoor thermometer or any thermometer if you have one- since it's getting cold now you can open a window to cool the room down if needed, I've seen my ambient go from 67f all the way up to 77f in an hour or two of running Fire Strike Ultra, so I open a window to cool the room down if I am running a controlled test). Again, I would expect to see a decrease in GPU and NVMe temps, a possible decrease in DRAM temps if you have a fan blowing on it directly, but a possible small increase in CPU temperature due to more hot air from another high-watt component (the GPU) exhausting its hot air up through the sink after the change.
Imo, Linpack is pretty dang overkill for testing a consumer CPU, as the only loads that will ever even come close to generating that kind of demand or heat would be video encoding, folding on CPU or mining on CPU. No game and few CPU-only benchmarks will ever run that hot. If it's working for you, good, but I'm sure you wouldn't want to blow your setup up. Linpack is like the Furmark Fuzzy Donut of CPU benchmarks and really pretty unneeded. We aren't testing performance of supercomputers here, like top500.
I'd also avoid Prime95 in this day and age, given that it uses AVX by default. It's known for making Ryzen 3000 run so hot it will overheat/crash at anything more than 4.2GHz or so. Ironically, I found out SiliconLottery is using AVX Prime to test and bin Ryzen 3000 series chips, and they previously used ROG Realbench for 1hr to bin chips (the paper that came with my binned 4790k says as much). Anyway, this means they don't bin or sell any R9 3900x that run faster than 4.2GHz, but even the worst R9 3900x can easily do at least 4500/4450/4300/4250 1.425v manual CCX OC on the boards that have the option to. (Manual OCs top out at 43x all cores- CCX or Core Complex OC sets the ratio per 3 cores/6 threads across 4 Core Complexes, with 2 per chiplet. So, in that example I gave, the first 6 threads would run at 4500 (for gaming), second 6 at 4450, third 6 threads on the first CCX of the lower binned chiplet at 4300, and fourth 6 threads at 4250MHz. Since the 3900x runs very hot, and automatically assigns gaming loads that only use ~6-8 threads tops to the first CCX on chiplet 1, this method allows faster speeds for games while still enforcing Turbo clocks for 24-thread loads that won't cause thermal shutoff and will give increased performance compared to auto OC.)
Anyway, I don't use Prime95 or Linpack to test my rig. I'd recommend Realbench, HWBOT x265 looped or any high thread count x264 bench looped (these will crash pretty quickly if unstable, but generate far less heat), or something like that. Personally, I've never liked AIDA's stability test very much and don't use it myself but some people like it.
What I've been using is pretty simple- Cinebench R20 set to "realtime" process priority, run over and over. Yeah. If I'm unstable it basically crashes after 3 seconds or so and my rig resets (in the case of unstable DRAM) and if I am clocking too aggressively I'll trigger TJmax and overheat on the socket (surpass 95C) which will cause the machine to reboot and the bios to whine to me. If I'm stable on both memory OC and and not overheating, I can run it 5-10 times back to back totally fine, though my cooling has difficulty keeping up, even with the addition of the fans I mentioned. But this seems to work really well and if I can run R20 5 times without crashing back to back, the system never crashes under either Win10 or Linux, period. It also only takes me maybe 10 minutes of sitting down to check. Since my board/chip is essentially bleeding edge and has bugs/kinks in the BIOS to iron out over the next year, and I will definitely be updating my BIOS often to fix bugs/stability/memory OC issues/CPU overclocking being less than advertised, and we've gone through like 10 AGESA revisions already, I don't see much point in doing thorough overnight stability testing with Realbench, HWBOT x265, etc. yet because it will just have to be retested. Cinebench R20 in the manner I stated seems to verify my core and memory OC stability for gaming and general use well enough that I don't crash, so that's fine for now, and once we don't have any planned BIOS releases for a long time I will do more thorough testing...
For checking memory errors I use the built in test in Ryzen DRAM Calculator
Set like that. If I set it to use all free RAM, it complains and won't run (I use a ramcache program for my SSD, and a page file), so I tell it to run 9000MB to 3200% and this will make it run to around 105% on all threads. Since your chip is 12-thread you should be able to set it to 1600% or so and end up similarly. However, I'll warn that this program does not find all memory errors in all loads- I've seen someone complaining recently that they passed this to 32000% (or something, equating to over 1000% each thread, the percentage just seems to be amount of time to run), yet when rebooting would get random F9 memory error codes (= loading recovery module/setting bios to defaults, meaning they were NOT stable). Something like ROG Realbench for 1h which simulates heavy real-world use (and runs 3d apps, video encoding, web browsing etc. all at the same time) would probably work better.
Anyway, that's it/my suggestions or thoughts for now, hope you find it useful and it gives you some ideas to play with/consider.