Introducing the new AMD Ryzen

Jim1348
Jim1348
Joined: 19 Jan 06
Posts: 463
Credit: 257,957,147
RAC: 0

Christian Beer wrote:Just out

Christian Beer wrote:

Just out of curiosity I searched for "Ryzen" in the DB and here are the three other hosts that have this new CPU:

https://einsteinathome.org/host/12413792

https://einsteinathome.org/host/12502414

https://einsteinathome.org/host/12502620

Runtimes on windows seem to be between 27k and 38k seconds.

It is hard to tell on those machines, since we don't know the loading, but it would appear that on heavily-loaded cores the run time is at the high end of that range.  In that case, the Ryzen does a lot better (more so than the Intel chips) under Linux. 

I always had the impression that AMD did relatively better than Intel under Linux, but I have not had one since the Stone Age (before writing was developed), so I can't really compare them.

VinodK
VinodK
Joined: 31 Jan 17
Posts: 15
Credit: 246,751,087
RAC: 0

koschi wrote:All 16 threads

koschi wrote:
All 16 threads used, not exclusively Einstein though. Currently doing testing across multiple projects...

Don't forget to test with SMT disabled. 16 threads simultaneously might create a bottle neck elsewhere in the system. 

Keith Myers
Keith Myers
Joined: 11 Feb 11
Posts: 4,781
Credit: 17,814,814,381
RAC: 4,022,929

Vasishk-Taneya wrote:koschi

Vasishk-Taneya wrote:
koschi wrote:
All 16 threads used, not exclusively Einstein though. Currently doing testing across multiple projects...

Don't forget to test with SMT disabled. 16 threads simultaneously might create a bottle neck elsewhere in the system. 

I believe you should NOT disable SMT if running Linux with the latest 4.10 kernel.  That properly handles SMT and coreinfo shows it correct with that kernel.  SMT is a problem in Windows though and those users will have to wait on Microsoft to ship an update in the future to properly handle the new Ryzen architecture.

 

koschi
koschi
Joined: 17 Mar 05
Posts: 86
Credit: 1,664,597,555
RAC: 37,433

Sorry, I was away for a

Sorry, I was away for a week...

No way I'm going to disable SMT. On my old i7 3770 SMT yielded an extra 15-40% more throughput depending on the project.

I am already running kernel 4.10.2 from Ubuntu mainline repository, it shows the following core/cache hierarchy:

lscpu -e
CPU NODE SOCKET CORE L1d:L1i:L2:L3 ONLINE MAXMHZ    MINMHZ
0   0    0      0    0:0:0:0       yes    3700.0000 1550.0000
1   0    0      0    0:0:0:0       yes    3700.0000 1550.0000
2   0    0      1    1:1:1:0       yes    3700.0000 1550.0000
3   0    0      1    1:1:1:0       yes    3700.0000 1550.0000
4   0    0      2    2:2:2:0       yes    3700.0000 1550.0000
5   0    0      2    2:2:2:0       yes    3700.0000 1550.0000
6   0    0      3    3:3:3:0       yes    3700.0000 1550.0000
7   0    0      3    3:3:3:0       yes    3700.0000 1550.0000
8   0    0      4    4:4:4:1       yes    3700.0000 1550.0000
9   0    0      4    4:4:4:1       yes    3700.0000 1550.0000
10  0    0      5    5:5:5:1       yes    3700.0000 1550.0000
11  0    0      5    5:5:5:1       yes    3700.0000 1550.0000
12  0    0      6    6:6:6:1       yes    3700.0000 1550.0000
13  0    0      6    6:6:6:1       yes    3700.0000 1550.0000
14  0    0      7    7:7:7:1       yes    3700.0000 1550.0000
15  0    0      7    7:7:7:1       yes    3700.0000 1550.0000

 

Linux is already well aware, what is where. Disabling SMT might shorten the invidual run times, but not increase throughput.

 

edit: This WYSIWYG editor and or my Firefox displaying it are a nightmare, can't fix it, hence new post below...

koschi
koschi
Joined: 17 Mar 05
Posts: 86
Credit: 1,664,597,555
RAC: 37,433

Jim1348 wrote:Edit: Do you

Jim1348 wrote:

Edit: Do you have times for the Gravitational Wave search CV v1.00 (AVX)?  I normally do those on the CPU, and run the FGRP on a GPU.

 

I have these enabled, but am not receiving work for them, no idea why...

Christian Beer
Christian Beer
Joined: 9 Feb 05
Posts: 595
Credit: 135,167,370
RAC: 270,479

koschi wrote:Jim1348

koschi wrote:
Jim1348 wrote:

Edit: Do you have times for the Gravitational Wave search CV v1.00 (AVX)?  I normally do those on the CPU, and run the FGRP on a GPU.

 

I have these enabled, but am not receiving work for them, no idea why...

We are restricting the O1MD1CV search to specific Intel CPU models. I just added the Ryzen to this club and you should get work at your next request. I should have done that earlier, the benchmarks and tests I saw are just amazing.

koschi
koschi
Joined: 17 Mar 05
Posts: 86
Credit: 1,664,597,555
RAC: 37,433

Oki, thanks

Oki, thanks Christian!

Currently running some Phoronix benchmarks that will last another hour, will report back as soon as I have O1MD1CV results uploaded.

edit:

 

work fetch is ok now, got 15 WUs for the 16 threads (one was running FGRP in that moment)

Rob
Rob
Joined: 17 Apr 12
Posts: 8
Credit: 1,781,499
RAC: 0

Hello,I have a newly

Hello,

I have a newly acquired Ryzen 7 1700 and am running a few tasks right now. For any fellow Ryzen crunchers on Windows (10), you might want to check your "Power Options".
On Balanced mode, Windows currently overaggressively parks cores and condenses processes on fewer cores.

Example: I have 8 Einstein@Home tasks active right now on my CPU. If I use balanced, those tasks will be on 4 cores with their 8 threads. However, 4 more cores are being parked (not used). This should lead to weaker performance, as instead of using all 8 physical cores, Windows relies on 4 + SMT.
(Picture)

If I set the Power Options mode to High Performance, this changes. All 8 cores are awake and being utilized with tasks, resulting in 8C (16T) being used.
(Picture)

(All of this may not be all that relevant to "power crunchers" who will let BOINC run (close to) 16 tasks, anyway, because Windows should not get the idea of parking cores in that scenario anyway because it can't distribute 16 threads on less than 8 active cores)

AMD has announced Core Parking will be revamped with an update in April, so if you're using this version of Windows with Ryzen, you may want to manually check power settings if you want the best performance.

"In the near term, we recommend that games and other high-performance applications are complemented by the High Performance plan. By the first week of April, AMD intends to provide an update for AMD Ryzen™ processors that optimizes the power policy parameters of the Balanced plan to favor performance more consistent with the typical usage models of a desktop PC."

Also relevant to users of the 1700X/1800X (but not 1700) is, that currently there are intentionally wrongfully high reported CPU temps being reported.

"Specifically, the AMD Ryzen™ 7 1700X and 1800X carry a +20°C offset between the tCTL° (reported) temperature and the actual Tj° temperature. In the short term, users of the AMD Ryzen™ 1700X and 1800X can simply subtract 20°C to determine the true junction temperature of their processor. No arithmetic is required for the Ryzen 7 1700. "

Personally, less from a cruncher (I only do it sporadically) and more from an allrounder-POV.. I can recommend my CPU. The Ryzen 1700 has good power consumption, amazing multi-threaded applications performance (falling a little flat restricted to games) and, if one wants to, the ability to clock it to the levels of its bigger brothers. It does that with a good price, and a good and relatively quiet boxed cooler included (I have not tested long-term 100% CPU loads here!).

I have not yet tested undervolting (less power consumption, less heat, less noise, same performance) but I've heard that there are solid results with usually undervolting potential of 0.1 to 0.15V, resulting in 20W less power consumption at full load. [Source: Computerbase (German)]
I think the Ryzen family of CPUs may be pretty interesting for crunchers.

mmonnin
mmonnin
Joined: 29 May 16
Posts: 291
Credit: 3,232,287,015
RAC: 6,954

Depending on the project and

Depending on the project and the functions of the CPU that are shared between logical cores and physical cores, disabling SMT/HT can actually be a benefit. I've tested this on a different BOINC CPU project and gained a healthy performance boost with HT off on with a 2P 2760s. Those are Intel and it matters what parts are shared between logical cores. No point cramming a 2nd thread down a unit if a single thread can max it out. I would suggest at least to try it. Most things benefit, but not all.

koschi
koschi
Joined: 17 Mar 05
Posts: 86
Credit: 1,664,597,555
RAC: 37,433

First O1MD1CV WUs seem to run

First O1MD1CV WUs seem to run ~13hours (still not completed). I started at 3GHz in the morning, using auto vcore setting for the CPU which resulted in the BIOS showing 1.35-1.4v being given to the CPU. Unfortunately the monitoring chip on the ASUS PRIME X370-PRO isn't supported under Linux yet, so all sensor data I have is from BIOS only. Now the chip is running 3.6GHz-[at]-1.2v, cooling remains calm, the whole computer draws 140W running 15 Einstein AVX tasks plus one WCG.

 

edit:

I have done SMT vs. non-SMT testing on my i7 3770 (DDR3 1600, Linux) across various BOINC projects. I have grabbed a random WU out of the slots, then benchmarked it with 4 processes and 8 processes running in parallel. Results below:

WCG FAH
4 threads: 1344s
8 threads: 2237s
Throughput gain: 20%

POGS
4 threads: 2533s
8 threads: 4152s
Throughput gain: 22%

Universe Hspin2
4 threads:  9739s
8 threads: 12496s
Throughput gain: 55,8%

Milkyway
4 threads: 2316s
8 threads: 3736s
Throughput gain: 24%

Enigma
4 threads: 1639s
8 threads: 1011s
Throughput gain: 23,4%

TN-Grid gene vv
4 threads:  884s
8 threads: 1352s
Throughput gain: 30,8%

These speak for themselves, but I understand that they might not be achievable on all computers. On a multi socket computer with many cores, memory bandwidth might be an issue as well, while it wasn't for me.

I will be redoing the above tests on the R7 as well, once I have to time to babysit them...

 

edit:

few results completed (45-49k seconds), thats still 3-4h longer than the Haswell based Xeon E3-1220 v3 and E5-1650 v3 wingmen from ATLAS AEI Hannover
Their computer info page shows only 4 cores for the E3 and 6 for the E5 though. The first is a non-SMT CPU, while on the second SMT was most likely disabled (or ncpus=6 set in BOINC).

Keeping that in mind, the run times seem reasonable.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.