AMD Zen/Ryzen thread

Message

Post by **Chewi** » Fri Jun 02, 2017 8:56 pm

Naib wrote:Care to add yourself to this survey: Gentoo & Ryzen config & stability questionair (public) also: the auto-updating sheet of the results
Just trying to help pull together the pieces of info to help

Will do once I've seen whether this next change works. The last entry looks interesting, GCC 6 with native on almost the same board as me reports no problems. The main difference is they're using the beta BIOS with AGESA 1.0.0.6. I was going to try that next before I fixed the freezes.

Naib · Post by **Naib** » Fri Jun 02, 2017 9:03 pm

If possible that would be good, more datapoints the better (even adding a new entry with updates)

when I was on gcc-6.x there were some odd emerge affects (not segfaults as others have mentioned) more it would just stop (not quit stop, stop completely no /var/log/messages or dmesg entries...). Gcc-7.x has been smooth for me so there could be a gcc related bug (there is at least one confirmed issue with gcc-6.x and equally some recent ryzen compiles were VERY inefficient)

Personally I suspect timing ... the zen fabric has been shown to be very dependant on RAM speed & this was shown in windows with game benchmarking improving with RAM clock. BIOS's at the moment have been very conservative with RAM timings & equally AGESA has restricted maximum speeds...

Also NEddy suspected some rail stability and a recent AMD bug report did suggest disabling LLC on those motherboards that support it to help improve Vcore & such stability

wrc1944 · Post by **wrc1944** » Fri Jun 02, 2017 10:45 pm

On my two Gentoo AM4/Ryzen installs I was having serious problems for days- FWIW: details are at: ( https://forums.gentoo.org/viewtopic-t-1064308.html ).

Anyway, after openssl-1.1.0f was masked, and openssl-1.0.2k was replaced with openssl-1.0.2l , I was able to resolve my issues. It definitely was an interesting ordeal.
(Many thanks to Lars Wendler (Polynomial-C) https://gitweb.gentoo.org/repo/gentoo.g ... c1d0890ec1 )

Now I finally have two Gentoo nice full Kde,Xfce4, and Lxde ~amd64 Ryzen desktops both compiled with gcc-7.1.0-r1 on kernel-4.12.0-rc2, and currently with:

Code: Select all

CFLAGS="-march=native -O2 -pipe -fomit-frame-pointer -mno-fma4 -mno-tbm -mno-xop -mno-lwp"
MAKEOPTS="-j9"
ACCEPT_KEYWORDS="amd64 ~amd64"
CPU_FLAGS_X86="mmx sse sse2 sse3 sse4a sse4_1 sse4_2 ssse3 mmxext aes avx avx2"
INPUT_DEVICES="evdev"
VIDEO_CARDS="amdgpu fbdev"
ALSA_CARDS="snd_hda_intel"
RUBY_TARGETS="ruby21 ruby22 ruby23 ruby24"
PYTHON_TARGETS="python2_7 python3_4"

Both initial installs were done with CFLAGS="-march=k8 -mtune=generic -O2 -pipe", and then at first I was on gcc-5.4 and tried -march=bdver4 with the -mno-fma4 -mno-tbm -mno-xop -mno-lwp flags.

Had some success, got a reasonable desktop, but then had the openssl mess take me over, and for a while defeat me soundly. Finally at the same time the openssl fix occurred I decided to go with gcc-7 -march=native, and I'm not sorry. I did have to patch qtwebengine-5.7.1-r1 with 2 patches placed in /etc/patches/dev-qt/, which worked well. (Many thanks to Mark Wright on https://bugs.gentoo.org/show_bug.cgi?id=617896 )

The kdemultimedia dep libmp4v2-2.0.0 needs a gcc-7 patch also, but other than those 2 problems gcc-7 is doing nicely. Might be a few more random packages needing a patch with gcc-7 on some systems

EDIT: Forgot to mention- While I can run my ram trouble-free at 3200mhz when using other distros on this box, I did have to drop down to 2400 when doing these long installs and rebuilds/updating, or it would lock up after a few hours. At 2400, it hasn't happened again.. At only 65 watts on the R7 1700 cpu, I'm wondering if it isn't actually the ram overheating a little bit at 3200.

The G.Skill Flare ram is: Timing 14-14-14-34 Cas Latency 14 Voltage 1.35V by default

Post by **Maffblaster** » Sat Jun 03, 2017 4:04 am

Chewi wrote:I've recently joined this club.

Welcome to the cool kids club.

Post by **Chewi** » Sat Jun 03, 2017 9:03 am

Maffblaster wrote:
Chewi wrote:I've recently joined this club.
Welcome to the cool kids club.

Pfft, I was already in that club.

Building the kernel with bdver4 didn't pan out as I had a freeze overnight. I've left it with that but updated to the Gigabyte beta BIOS F6f with AGESA 1.0.0.6. If that doesn't work, I'll try GCC 7. If it still plays up, I'll have to start trying different march options to try and narrow down exactly which extension is breaking it.

Website · Post by **Zucca** » Sat Jun 03, 2017 9:49 am

I added a tip for Ryzen users to the Safe CFLAGS wiki page. That also includes the link to the questionair. Hopefully more people fill it.

RayDude · Post by **RayDude** » Sat Jun 03, 2017 11:05 pm

Sorry, first a little bragging about my good luck.

I'm running a stable Ryzen 1600 @3.8GHz, 1.4V core (I tried 4.0 and 3.9 and neither lived for very long) and @2933 (cas 17) DDR4 with a GTX1080. Emerges are rock stable and even with all the cuda cores blazing, the GPU at 80 C everything runs fine.

I think the difference is in the DRAM I bought. I suspect AMD has more issues with their DDR4 timing than they yet know. When I heard about the DDR4 issues I bought the memory that was most prized by PCMR / buildapc. I'm running these insane Galax DDR4-3600 Memory modules with default timings, but multiplied up to 2933. I haven't tried any higher because no one has gotten it to run higher, right? And most importantly, I'm working on this beast and stability is more important than speed.

The funny thing is: I'm running gcc 4.9.4 with -O2 -pipe. I've been advised by Neddy to go to 5.3, but I'm lazy and I haven't done it yet.

The one thing I really want to work is CPU temps. I got the it87 kernel module compiled, installed and running, but I can't get lm_sensors to work with it and the thread that said how to do it has been deleted. Can someone give me a pointer?

Post by **Chewi** » Sun Jun 04, 2017 7:29 am

RayDude wrote:I think the difference is in the DRAM I bought. I suspect AMD has more issues with their DDR4 timing than they yet know. When I heard about the DDR4 issues I bought the memory that was most prized by PCMR / buildapc. I'm running these insane Galax DDR4-3600 Memory modules with default timings, but multiplied up to 2933. I haven't tried any higher because no one has gotten it to run higher, right? And most importantly, I'm working on this beast and stability is more important than speed.

Maybe but my RAM was in the official compatibility list. You should be able to get more than 2933 now if you use an AGESA 1.0.0.6 BIOS. I only use 2667.

RayDude wrote:The funny thing is: I'm running gcc 4.9.4 with -O2 -pipe. I've been advised by Neddy to go to 5.3, but I'm lazy and I haven't done it yet.

That version is probably okay if you're not using -march.

RayDude wrote:The one thing I really want to work is CPU temps. I got the it87 kernel module compiled, installed and running, but I can't get lm_sensors to work with it and the thread that said how to do it has been deleted. Can someone give me a pointer?

You need it87 from git. It's not clear which temperature is which though and I never seem to get anything above about 50C on an air-cooled 1600X, which seems a little too good to be true.

I've now booted into Fedora 26 Alpha on a USB stick and chrooted into Gentoo. It's been for 12 hours so far. I'll give it another 12 hours before trying this kernel on my Gentoo system directly.

RayDude · Post by **RayDude** » Sun Jun 04, 2017 4:33 pm

Thanks. I have the it87 module installed and running, maybe I have to reboot to get the KDE thermal monitor to see it?

I used the kernel configuration from the sysrescue cd (5.0.0) to get my system stable. I couldn't boot on more than one CPU until I used it.

Here's the config I'm using: https://pastebin.com/Utf7qnn3

Yes, its bloated, but when I tried to change it, I lost stability. So clearly there are some intel drivers in there that are making things work better.

krinn · Post by **krinn** » Sun Jun 04, 2017 5:03 pm

RayDude wrote:I used the kernel configuration from the sysrescue cd (5.0.0) to get my system stable. I couldn't boot on more than one CPU until I used it.

That's quiet a different story than your previous post

RayDude · Post by **RayDude** » Sun Jun 04, 2017 11:32 pm

krinn wrote:
RayDude wrote:I used the kernel configuration from the sysrescue cd (5.0.0) to get my system stable. I couldn't boot on more than one CPU until I used it.
That's quiet a different story than your previous post

You're right. My bad. I completely forgot about it. It's one of those cases where you're panicking, thinking: "was it a bad idea to go bleeding edge?"

I was trying to boot various linuxen to see if I could find anything stable. When I booted sysrescue and it worked, I grabbed it's config, used it and then tried to tweak it. That was a bad idea. I stuck with it and then forgot how much trouble I had while I was playing with the new toy.

wrc1944 · Post by **wrc1944** » Mon Jun 05, 2017 4:21 am

My asrock-fatal1ty-x370-gaming-k4 along with some others is reviewed here: http://www.tweaktown.com/reviews/8135/a ... ndex4.html

On page 4 I saw in the circuit analysis the sensors chip is the nuvoton nct6779D, which has support in kermel 4.11.x and the 4.12-rc's. I enabled kernel support as a module, loaded it, and fired up my beloved Gkrellm , and sensors are back in my AM4 systems.

If you need sensors, it's probably worth looking at the other board's tweaktown's circuit analysis page for which chip your AM4 board uses. nct6779d works as good as it87 does on some AM3/AM3+ boards. For example, the msi krait has the nuvoTon NCT6795D, which probably will be compatible. The GIGABYTE AX370-Gaming 5 has ITE IT8792E, which I suppose is it87.
From make xconfig:

Nuvoton NCT6775F and compatibles (SENSORS_NCT6775)

CONFIG_SENSORS_NCT6775:

If you say yes here you get support for the hardware monitoring
functionality of the Nuvoton NCT6106D, NCT6775F, NCT6776F, NCT6779D,
NCT6791D, NCT6792D, NCT6793D, and compatible Super-I/O chips. This
driver replaces the w83627ehf driver for NCT6775F and NCT6776F.

This driver can also be built as a module. If so, the module
will be called nct6775.

Symbol: SENSORS_NCT6775 [=m]
Type : tristate
Prompt: Nuvoton NCT6775F and compatibles
Location:
-> Device Drivers
-> Hardware Monitoring support (HWMON [=y])
Defined at drivers/hwmon/Kconfig:1194
Depends on: HWMON [=y] && !PPC
Selects: HWMON_VID [=m]

mir3x · Post by **mir3x** » Mon Jun 05, 2017 6:18 pm

We all want to buy new Ryzen, question is which one

I found interesting benchamrks here:
https://www.xda-developers.com/xda-take ... inux-side/

Seems Ryzen 5 1600X is performing very well, in worst case just 20% slower than ryzen 7 1700X ( in john the ripper, c-ray, compile test).
Considering its 50% cheaper than 7 1700X, it's probably best offer now (Amd cut prices for ryzen 7 few days ago, so mb it not so good offer now), any thoughts ?
Maybe in next 3 years there will be next nice cheap ryzen for AM4 and now not worth to buy Ryzen 7 now, jsut Ryzen 5 1600X now?

Post by **NeddySeagoon** » Mon Jun 05, 2017 6:35 pm

mir3x,

mir3x wrote:We all want to buy new Ryzen, question is which one :D

Threadripper :)

More seriously, early adopters pay a premium for top end hardware.
If you can repurpose a lower spec CPU is a year or two, buy it now and upgrade later.
Buy the RAM for the CPU you plan to upgrade to, unless that will be repurposed with the CPU.
e.g. my bits get passed down the family, which is a good excuse to upgrade too.

krinn · Post by **krinn** » Mon Jun 05, 2017 7:54 pm

I must say i'm a bit surprise myself by ryzen, amd is plague with heat, but the 1600 is given at only 65W TDP?
And amd usually are cheaper, in france the ryzen 7 1800X is at 554€ while intel 7700K is at 409€
It's only france or do amd really have gone crazy? (i'm assuming the 7700K is better there, not only by the prize, but i could be wrong)
You can confirm lower tdp and way cooler cpu?

Website · Post by **Zucca** » Mon Jun 05, 2017 9:04 pm

NeddySeagoon wrote:Threadripper :)

More seriously, early adopters pay a premium for top end hardware.

... The two reasons at the moment why I still roll with FX-8350. I really look forward to Threadripper.

mir3x · Post by **mir3x** » Mon Jun 05, 2017 11:18 pm

It's only france or do amd really have gone crazy? (i'm assuming the 7700K is better there, not only by the prize, but i could be wrong)
You can confirm lower tdp and way cooler cpu?

Ryzen 5 1600 has 65 TDP,
1600X has 95 TDP and costs about 230 EU

hers compare i7700K vs 1600X
http://hwbench.com/cpus/amd-ryzen-5-160 ... e-i7-7700k

Tony0945 · Post by **Tony0945** » Mon Jun 05, 2017 11:50 pm

mir3x wrote:hers compare i7700K vs 1600X
http://hwbench.com/cpus/amd-ryzen-5-160 ... e-i7-7700k

Ah! But does i7 segfault or crash mysteriously? Yes, there are always teething troubles, but AMD really botched this launch and seems oblivious to problems.

I haven't owned an Intel since my first computer, a 486DX-100, but I've pretty much given up on AMD.

Ant P. · Post by **Ant P.** » Tue Jun 06, 2017 12:34 am

As a wise man once wrote, "never buy version 1 of anything". Maybe now people are beginning to understand that. Intel gets away with it because their high end desktop chips are just neutered Xeons from existing product lines - they get alpha tested on Ruby/Nodejs fans (the cloud).

I'll stick with my Phenom II for at least a few months more. Paid for 3 cores and been running stable with 4 for eight years now...

krinn · Post by **krinn** » Tue Jun 06, 2017 3:03 pm

mir3x wrote:hers compare i7700K vs 1600X
http://hwbench.com/cpus/amd-ryzen-5-160 ... e-i7-7700k

thank you, however
Max Case Temp: unknown for the 1600X

You guys with ryzen see a cooler cpu than what amd is use to? (my intel cpu is at 41°C doing nothing, and goes to 65°C when compiling)

Post by **NeddySeagoon** » Tue Jun 06, 2017 3:47 pm

Tony0945 wrote:... but AMD really botched this launch and seems oblivious to problems ...

I'm far more cynical. I don't think AMD know how to fix it yet. Maybe its not even all CPU problems.

Ant P. · Post by **Ant P.** » Tue Jun 06, 2017 6:33 pm

AMD seems to believe they've already fixed it, but the growing crowd of angry paying customers suggests otherwise.

tinea_pedis · Post by **tinea_pedis** » Tue Jun 06, 2017 6:59 pm

Assrock Taichi X370, 2.30 UEFI, 1700x at stock everything, 2x8G g.skill ripjaws v at 2133MHz. SMT disabled.
Migrated from FX-8350, on which everything was compiled with march=bdver2. Compiled gcc-4.8.5 with generic profile (no march, with -j8), and again to be safe and then everything@world.
Have not experienced segfaults before or after recompiling world. Kernel compile for 52 hours (forgot how many loops), no segfaults.

Post by **NeddySeagoon** » Tue Jun 06, 2017 7:56 pm

tinea_pedis,

gcc builds itself three times anyway.
First, with whatever random compiler you have installed, it need not be gcc. Just the C and C++ parts that it needs to build itself.
Next, it uses that gcc compiler to build its own C and C++ again.
Then it uses the gcc built with gcc to build it C and C++ parts and compares the output with the compiler it used to build this pass.
They should be identical. The build fails if they are not.
Lastly it builds the other languages.
Its the third build that gets installed on your system.

tinea_pedis · Post by **tinea_pedis** » Wed Jun 07, 2017 7:11 pm

This I learned not too long ago, but old habits die hard.. But this is lesser of many stupid practicies I've done in the past (in regards to gentoo, linux in general and hp-ux in the 90's).

AMD Zen/Ryzen thread

my kernel config

Re: my kernel config

Re: my kernel config