AMD Zen/Ryzen thread (part 2)

[Naib]

[krinn]

emerge is not a benchmark, so many things can alter the timing.
With my i7-4790K i have

[NeddySeagoon]

krinn,

Its perfectly clear that your i7-4790K is getting slower with age. :)
_________________
Regards,

NeddySeagoon

Computer users fall into two groups:-
those that do backups
those that have never had a hard drive fail.

[vaxbrat]

Pasted the CFLAGS wiki for experimental use with 5.4.0 from here in my make.conf since it appears that djvu doesn't compile with 6.4.0.

https://wiki.gentoo.org/wiki/Ryzen#GCC

However my 5.4.0. barks about not recognizing -mclzero so I took that one out.

[vaxbrat]

I got through my emerge -e system and am about ready to reboot into my 4.13.1 kernel to run an emerge -e world. I only had a handful of packages that were segfaulting, so for those, I decided to throttle back my job limit from -j16 to -j4. My theory was that natural scheduler affinity would be to put the jobs into just the cores and not threads on the first die. That appeared to work since none of my packages that failed under 6.4.0 (other than ones that needed 5.4.0) would fail with -j4. For my very last package which was webkit-gtk that failed with a segfault, I throttled back to -j8 since I wanted to see if affinity would spread things out amongst only cores and not threads between the two dies. That also appears to work, but that is only one data point. So when I do my emerge -e world, I will throttle back to -j8 only for the entire run of 1600 or so to see whether I have any failures at that level.

I only had a handful of packages out of the slightly more than 1100 that would only compile under gcc 5.4.0. None of those had segfaulting problems.

[Tony0945]

vaxbrat, that would seem to validate AMD's earlier advice to disable SMT.

[NeddySeagoon]

That's keeping a Ferrari just to drive to the corner shop.
_________________
Regards,

NeddySeagoon

Computer users fall into two groups:-
those that do backups
those that have never had a hard drive fail.

[Tony0945]

[vaxbrat]

I'm about to break the 1000 build mark on my 1600 build emerge and have yet to have a segfault happen by throttling back to -j8. The only times I stopped are a few times when I had to switch back to gcc 5 for ebuilds that can't compile with 6.4.0. It probably also helped that the system toolchain probably had all of its non-zen bulldozer instructions flushed out by this point. When this is done, I may flip back to -j16 to see whether/how many sefaults I might still get.

[vaxbrat]

That one finished okay and I found out my troubles with my amdgpu driver. It turns out the hdmi cable couldn't drive 4k res. When I swapped for a 4k displayport everything is happy happy.

I'm in the process of doing another 1600+ emerge -e world with -j16. If I hit a segfault, I'll try going back to -j12 with the theory that maybe trouble only breaks out when you have threads going across dies.

[mir3x]

I've sent mail to amd in monday, in tuesday i got rma from and resent it, in thursday i got resone that i can send. Im sending tomorrow to Netherlands.

I prepared my backup box.
I must say - one segfault per few days is nothing compared to that ubuntu on arm:

[Tony0945]

[vaxbrat]

I got back from work and found things had stopped at the first ebuild that wouldn't compile under 6.4.0. That was after about 700. I don't know why things would be working all of a sudden when the only thing different is getting the display card and display working okay.

[vaxbrat]

I just realized I had the cpu packaging sitting around and that has the serial number on it. Is there an AMD url to look that up and figure out week number? BTW things are still going just fine here without any segfaults.

TIA: My 1700x is PN YD17OXBCAEWOF sn 9GU8187O70072

[pjp]

Maybe this can help?

[mir3x]

[Tony0945]

Detailed explanation here: https://www.reddit.com/r/Amd/comments/6scnlg/ryzen_reading_your_production_batch_number/
It's unclear whether all of these numbers are on the box. All are on the chip. I think you have to take off your heatsink and re-apply Arctic Silver or whatever.

[mouacyk]

[Naib]

Well MSI have updated their bios for my mobo, agesa 1.0.0.6b & more RAM incompatibilities
the XMP profile finally has correct RAM timings so I didn't have to tweak the timing, only went for XMP-1 (2800) instead of the 3000 profile.

for the last 20min I have had 4 terminals open and each have had emerge mesa && emerge mesa && emerge mesa all four terminals have completed the 2nd emerge and the nct6775 sensor is indicating a temp (somewhere ...) of 53C

[Naib]

[Naib]

[thumper]

[vaxbrat]

Thought I had things under control but oh well. I fell back to -j12 and things seemed to behave after I got a segfault or two about 800 ebuilds in. We really should figure out if there's a way to tell build week number from the retail packaging serial number. I'm just beginning the process of retiring my fx-8350 and Kaveri boxes (when the apu version comes out), and I want to be able to go into MicroCenter and verify before I buy. I might end up building one threadripper just to have a seriously bad ass build and render server. It makes me cry thinking that work is still buying as new HP Z800 class workstations based on 6 generation old Sandy Bridge stuff and paying probably 3-4 times the price with the same amount of memory and a lot less disk.

[trippels]

[Naib]