AMD Ryzen 3000 Builders Thread

[Muadib]

3950x on LN2

5.2GHz Ryzen 9 3950X Benchmarks Blitzes the competition - OC3D

5.2GHz Ryzen 9 3950X Benchmarks Blitzes the competition AMD’s Ryzen 3rd Generation processors are coming, offering up to twelve cores at launch and up to 16 cores from September onwards, delivering compelling performance levels over a wide range of price points. In September, AMD plans to...

www.overclock3d.net

This stat is useless to consumers, but nice to see.

 

[Kenmitch]

This stat is useless to consumers, but nice to see.

I know. At least it's been spotted in the wild.

 

[DrMrLordX]

This stat is useless to consumers, but nice to see.

It's not entirely useless. It's interesting to see how scaling continues at higher and higher clocks. Also, I don't think der8auer was able to go that high with his 3900x samples.

 

[ClockHound]

Have my X470 Taichi board, 32GB micron e-rev memory, Inland 1TB NVME, some TBs of data drives... will re-purpose a case and D15.

What's left? Oh...a 3900X. As a very late adopter (X58 in 2014), being a tardy early adopter is trying my patience. If there isn't a new batch released by the weekend, drastic actions might be needed. Like another couple weeks of drastic waiting.

 

[VirtualLarry]

I was pleasantly surprised a few moments ago. Last week, I ordered another kit of GSKill Trident RGB DDR4-3600 "for AMD", SK Hynix chips, CAS 18-22-22-xx, 1T.

I was currently running the original kit @ 3600, FCLK 1800, and it was (mostly) stable.

I put in the second kit that I had received the other day. Beep-beep-beep, beep-beep-beep, beep-beep-beep, BEEP.

Well, BIOS reset. I tried 3400 speed, 3200 speed, then I realized, BIOS RAM count was only 16GB, it wasn't recognizing the second kit.

Paying more attention to installation, and lining things up correctly this time, I was able to get it to recognize all 32GB right off, and it's booted into Windows 10 64-bit Pro 1903 and mining on it, with 32GB (4x8GB @ 3600 / FCLK 1800). Amazing.

 

[lopri]

Update on MSI X470 Carbon: Rock solid DDR4-3600/CL16 with 64GB. Only needs 1.40V in BIOS. (HWinfo reports 1.408~1.424V)



PBO is enabled for the above LargeFFT run with 108W PPT, which is 20W higher than stock. I wanted to see how well the stock HSF handles the higher thermal load. I am back to using Wraith Prism cooler for this setup because it grants access to DIMM slots, unlike the Mugen heatsink which completely blocks the first slot. I do not like the Wraith's noise profile (read: loud), but I do like its cooling performance, especially how it keeps the memory sticks cool. With Mugen I observed the memory temp climbing above 65C.

 

[lopri]

Geekbench run with the above setup:

Micro-Star International Co., Ltd. MS-7B78 - Geekbench

Benchmark results for a Micro-Star International Co., Ltd. MS-7B78 with an AMD Ryzen 7 3700X processor.

browser.geekbench.com

If you do not mind the noise, the stock cooler for 3700X is a not a bad choice at all. The CPU performance based off of it, judging from the Geekbench scores submitted, is more than competitive, and as a bonus it cools the memory sticks as well. Sure, you will not get much overclock with it, but these chips do not overclock even with water cooling anyway.

 

[lopri]

I finally bought and installed a 3700x to replace my 1700. My overall impression is that it's a super nice chip. Everything is nice and snappy. I'm having the issue where I'm not hitting max boost even on a single core. I'm not surprised because it looks like the Asus version of BIOS 1.0.0.3AB for x370 was rushed. There are no less than 3 different spots where you can change the same PBO setting. Even with PBO enabled I've only seen 4.275 max. With PBO disabled the chip hits its 88w PPT limit and is maybe 50-100 mhz less on ST and MT than with PBO enabled. I'm optimistic AMD can fix this with a new BIOS.

Oh wow. That is the same crap I dealt with on Gigabyte BIOS.

 

[lopri]

Paying more attention to installation, and lining things up correctly this time, I was able to get it to recognize all 32GB right off, and it's booted into Windows 10 64-bit Pro 1903 and mining on it, with 32GB (4x8GB @ 3600 / FCLK 1800). Amazing.

What was wrong with it initially?

 

[VirtualLarry]

What was wrong with it initially?

I don't know, I must have somehow mis-aligned something when I put the second kit in, because it wasn't even detecting as 32GB in the BIOS, even when it reset the DRAM clock to 2133 as a failsafe. Once I removed and re-installed those two sticks, properly, then it detected as 32GB, and I was able to set DOCP to 3600 and 1800 FCLK.

 

[lopri]

What is the voltage for all 4 sticks with those kits at 3600? And the motherboard?

I am trying to find out how much more voltage is required typically for 4 DIMMs instead of 2 DIMMs.

 

[VirtualLarry]

Mobo is Asus B450-F ROG STRIX Gaming ATX mobo. R5 3600. GSkill DDR4-3600 Trident RGB "for AMD". Vdimm of 1.35V, Vsoc of 1.1V.

 

[lopri]

For all 4 sticks? That is excellent. Both my Gigabyte board and MSI board need more voltage for 2 4sticks in comparison to 2 sticks.

I spoke too soon about stability in my previous post. The system runs Prime95 fine but for whatever reason I am getting occasional stutters when playing online video. I am guessing Infinity Fabric is the suspect and trying to figure out what needs to be done.

 

[scannall]

For all 4 sticks? That is excellent. Both my Gigabyte board and MSI board need more voltage for 2 sticks in comparison to 2 sticks.

I spoke too soon about stability in my previous post. The system runs Prime95 fine but for whatever reason I am getting occasional stutters when playing online video. I am guessing Infinity Fabric is the suspect and trying to figure out what needs to be done.

Windows might be the cause.

 

[lopri]

I was thinking more voltages. I will test it out and report back. It is one of those things that happens unpredictably.

 

[lopri]

It might have been just a freak incident. I found a 12-hour long 4K video on YouTube and had it run full-screen while I was away, and when I came back after 10 hours or so it had lost about 18 frames out of 10K frames, which is insignificant. I will keep an eye on it.

 

[TheGiant]

It might have been just a freak incident. I found a 12-hour long 4K video on YouTube and had it run full-screen while I was away, and when I came back after 10 hours or so it had lost about 18 frames out of 10K frames, which is insignificant. I will keep an eye on it.

worse thing happen to me this morning
after 2 days of handbrake jobs with my 4k iphone movies my MB (MSI tomahawk b450) died
My guess is the power system on MB couldnt deliver 48h of sustained load
It is just the MB, got a replacement and the CPU (3900X) works
my system is pulling ~240W wall power, maybe that is too much for the b450 midrange board as sustained load for nearly 48 hours
but that never happened before with that 6600k- ofc it has just 110W of wall power...

 

[Despoiler]

worse thing happen to me this morning
after 2 days of handbrake jobs with my 4k iphone movies my MB (MSI tomahawk b450) died
My guess is the power system on MB couldnt deliver 48h of sustained load
It is just the MB, got a replacement and the CPU (3900X) works
my system is pulling ~240W wall power, maybe that is too much for the b450 midrange board as sustained load for nearly 48 hours
but that never happened before with that 6600k- ofc it has just 110W of wall power...

Check the back of the board where the VRMs are to see if there is any discoloration of the PCB. Many times the lower end boards skimp on the VRM cooling and the PCB itself gets too hot.

 

[DrMrLordX]

after 2 days of handbrake jobs with my 4k iphone movies my MB (MSI tomahawk b450) died

Ouch. What kind of VRM temps were you getting?

 

[TheGiant]

Check the back of the board where the VRMs are to see if there is any discoloration of the PCB. Many times the lower end boards skimp on the VRM cooling and the PCB itself gets too hot.

no traces of burning or overheating I checked before the replace
I got a fine deal with my HW shop - instant replace, morning it died and my assistant replaced it 2h after the MB died
strange is I can not reach the same memory timings, same board, same revision, same BIOS, its not stable..maybe the dying MB damaged the memory sticks...

Ouch. What kind of VRM temps were you getting?

high 80C, that shoudnt be a problem
I got a steady aiflow though my case
2x front 120mm, 2x back and 1 up
and you usually smell the burning problem

my guess it just died, we will see I got a big backlog to convert in handbrake

I will definitely look more closely to it

 

[thesmokingman]

worse thing happen to me this morning
after 2 days of handbrake jobs with my 4k iphone movies my MB (MSI tomahawk b450) died
My guess is the power system on MB couldnt deliver 48h of sustained load
It is just the MB, got a replacement and the CPU (3900X) works
my system is pulling ~240W wall power, maybe that is too much for the b450 midrange board as sustained load for nearly 48 hours
but that never happened before with that 6600k- ofc it has just 110W of wall power...

My x570 StrixE routinely does overnight encoding queues on a 3900x w/o fail. It runs about 250w at the wall too. On top of that it runs vpn and plex duties.

 

[TheGiant]

My x570 StrixE routinely does overnight encoding queues on a 3900x w/o fail. It runs about 250w at the wall too. On top of that it runs vpn and plex duties.

well as I said, we will see
the MB compatibility of new CPU especially with lower node (less V, more A) was always a problem
AMD definitely overdelivered with r3 lineup, just 3900X 12C may be too much for standard mid range board design with heavy load
with standard load/gaming/wtfother other load may be just 99,% perfect
AMD definitely did a N times better job than Intel

 

[DrMrLordX]

high 80C, that shoudnt be a problem

Generally not, I wouldn't think. Might have just been defective from the start. Once you made them work into the upper range of acceptable parameters . . . poof. If the VRM had been truly overwhelmed by the power draw, the system would have just throttled back the CPU to save the VRMs.

 

[Sunaiac]

I finally got my system working !

After a delay because the NVMe HDD wasn't coming.
After a long pause because of holidays,.
After another delay because the NVMe HDD was dead...

I built my system

It is still not watercooled, needed to test everything before going all plumber on it.

System specs :
- 3900X + stock cooler
- Taichi x470 bios 3.6 (not ABBA yet)
- 32GB 4x8 3600C15 gskill
- 1TB NVMe rocket sabrent
- 5700XT anniversary + stock cooler
- Phanteks P600s eclipse

Just had the time to install windows, and benchmark a little, only using cinebench R20 for now (so not a very good memory benchmark)
Here are my first results playing with 1usmus DRAM calc (not touching anything CPU side except -0.1v, just because I could):

STOCK + XMP : AIDA 64 read speed : not run, latency not run, CB R20 MC : 7227 SC 504
-0.1v + XMP : AIDA 64 read speed : 55714 MB/s, latency 69.8 ns, CB R20 MC : 7250 SC 496
-0.1v + 3533 DRAM calculator FAST : AIDA 64 read speed : 55686 MB/s, latency 67.3 ns, CB R20 MC : 7214 SC 506
-0.1v + 3600 DRAM calculator FAST: AIDA 64 read speed : 56519 MB/s, latency 66.7 ns, CB R20 MC : 7190 SC 505

That numbers averaged over the very representative amount of 1 run each time
Really just some quick toying.
Also installed and played street fighter V for two hours, nothing to report there, all fine.

Now I have to put in the waterloop, test if my CPU is 1900MHz FCLCK compliant, and start benching some visual studio compilation and a few games maybe ^^

What I aim for :
- no CPU core O/C other than the -0.1V and the gains brought by new AGESA
- AGESA 1.0.0.3 ABBA when available from ASROCK
- 3800/1900 Mem if CPU complies, else my current 3600/1800/CL14 with DRAM calc subs is fine
- no GPU O/C other than undervolting

so basically, nothing fancy, let the CPU auto overclocks work and call it a day.

 

[DrMrLordX]

@Sunaiac

Try some different LLC settings to see if you can improve your all-core boost clocks. I've gotten between 7300-7400 in CBR20 MT by using LLC0 and -.1v offset. Granted, that's on a different board and with the new AGESA version. Your scores are in line with what I got with 1.0.0.3ABB, mostly.