How efficient is your cruncher?

[Caring1]

I've been playing with a XEON system I've recently cobbled together and was looking to replace my current system in my specs with.
After running both for 24 hours on a power meter the 16 thread XEON looks to be my favoured choice at 2.7Kw for that period, compared to the 2.35Kw for the 4 thread Intel i5 system, running similar tasks etc on WCG.
I am waiting to see what Zen 2 brings to the table as I am keen to try one.

 

[phill]

Tomorrow I will update with all my crunchers Will be interesting to see how the compare and I'll get @blindfitter involved hopefully too

 

[thebluebumblebee]

thebluebumblebee | brick | i5-6500T| 1 SSD | 38W (4 threads) | 5,166 | 135.96 | Win 10 Pro | HP EliteDesk 800 35W G2 Desktop Mini PC

 

[FordGT90Concept]

https://boincstats.com/en/stats/-1/host/list/0/0/6707163cc325d1a837e345633e564fc4

No idea if I did my math right.

FordGT90Concept | Enermax Revolution 85+| Xeon E3-1230v3 + HD 5570 | 5 HDD + 2 SSD | 137w @ 100% (8 threads) | 4361 | 31.83 | Server 2012R2 |

 

[thebluebumblebee]

thebluebumblebee | Seasonic G-450 | Ryzen 7 1700 + GTX 460| 1 SSD | 135 watts | 20,761 |153.79 | Mint 18.1 XFCE | OET/HSTB only

This is based on the average of the last 22 days.
If I went headless and got rid of the parasitic 14 watts from the GTX 460, my PPW would jump to: 171.58!

 

[phill]

Well I'm not sure how accurate this list is going to be now but... Here's a list of my crunchers for the moment

Dell PSUs (dual) | 2 x L5640's + Onboard |5 1Tb SAS 7.2k drives | 275 watts | Mint 18.3 via running a VM| SCC only unless HST available
Corsair AX 1200 | 2 x X5650's + 7970 |1 SSD, 2 WD Velcoraptors| 310 watts | Mint 18.3 | SCC only unless HST available
Laptop | 1 x 6700HQ | 1 SSD | 65 watts | Windows 10 | SCC only unless HST available
EVGA 650w G+ | 1 x 2600k | 1 SSD | Est 150 watts | Windows 10 | SCC only unless HST available
EVGA 650w G+ | 1 x 6700k | 1 SSD | Est 140 watts | Linux Mint Mate 18.3 | SCC only unless HST available
Laptop Work | 1 x 4210M | 1 SSD | Est 45 watts | Windows 7 | SCC only unless HST available
EVGA 650w G+ | 1 x 2600k | 1 SSD | Est 150 watts | Linux Mint Mate 18.3 | SCC only unless HST available - Same machine as above, dual boot

If anyone would like any more info on any of these

EDIT - I forgot to mention, none of these machines run 24/7 so it's at best about 12 to 18 hours each, the R710 and the SR-2, they are just working when the solar is producing some good juice, otherwise, they are off and the more efficient crunchers are on and working away, like the laptop and the 2600k and 6700k..

 

[HammerON]

HammerON | Corsair AX1200 | i9 7980XE @ 3.4 GHz .98v + GTX 980Ti | 1 SSD | 367W @ 100% | 53,312 | 145 | Ubuntu 18.04

HammerON | EVGA SuperNova 650 G1 | AMD EPYC 7401P @ 2.8 GHz | 1 SSD |182W @ 100% | 22,046 | 121 | Ubuntu 16.04

HammerON | Thermaltake 550 | E5-2683 v3 @ 2.5 GHz + ATI 3650 | 1 SSD | 212W @ 100% | 26,088 | 123 | Ubuntu 16.04

 

[phill]

Just for clarity @HammerON , do you say thread = core?? Not thread = core + HT?? Curious as the AMD rig seems to be a little low on the points as it has the most threads....

 

[HammerON]

I was just following the example from the OP's list. All of them are crunching with HT so i9 7980XE = 36, AMD EPYC = 48 and the Xeon = 24.
I will just remove the threads part as it makes no sense to me...

The EPYC build has been a big disappointment, especially after I got the i9 7980XE build up and running. I am getting rid of the EPYC build and replacing it eventually with something else. I was thinking of a TR 2 possibly. I would like to see some crunching numbers on several TR 2 CPU's...

 

[phill]

Understood sir Thank you for clearing that up!!

 

[thebluebumblebee]

The EPYC build has been a big disappointment

What kernel are you using? When I upgraded my Mint 18.1 to 18.3, I had to go through a recovery boot (because of the blank screen issue with 18.3) to install a video driver but also saw that there was an AMD micro-code driver available that I also installed. Results? Psensor started to work correctly. I then upgraded the kernel to 4.15 and things really seemed to hum and I *think* that my average moved from 20K to 25K PPD, with OET. Then I got greedy and tried to install 4.18, after reading the TR2 reviews at phoronix.com, but I think I broke it. Now I'm planning on moving to openSUSE Tumbleweed, also based on what I saw at phoronix.com.

I've also wondered what the effect is of you only populating 2 channels of the 8 available.

Edit: uname -r for kernel version

 

[phill]

Well guys, I've been saying that I've been testing my 5960X for a while, so I've been looking at some results for you, so here goes

If anyone has any questions about the above, please let me know If there's anything I could/should have done differently please drop me a line

Overall though I am fairly happy with the results and considering the wattage loads are of a complete system so not just CPU wattages, for a near 3 year old CPU, I think it's doing ok I'll see if I can get some 4.80Ghz results in as well just for comparison sake.. Sadly I can get the CPU to boot into Windows at 5.00Ghz but it's not going to be stable enough for any WCG testing.. Plus the volts would be crazy high and I don't want to risk my CPU!

I'm running another test at the moment as well, I'm over half way through with it as I've got all of my crunchers whirring away at the moment, I'm trying to get a two day average of how they perform, so tomorrow after work I hope to be able to put up some results and numbers
Thanks for reading

 

[ArbitraryAffection]

Hi @phill where can i find benchmarks of how Well zen CPUs do on the various workloads used by WCG? I am interested in the perf/watt vs competing Intel solutions. Though I am sure the initial cost of investment will be significantly higher on the Intel, especially with 1700's at rock bottom prices (My two were £150 from amazon). But in general i am interested in the performance of the Zen core for WCG.

My technobabble:
If i had to guess, i would say, when working with enough independent, single-thread tasks to occupy all logical processors, 16 on a 1700, the Zen may fair better than say, a 6900K if at the same clocks. Because of the wider FPU engine in the Zen core (admitedly with narrower pipes). But in workloads not using over 128bits of data for SIMD in one go, two threads on a single Zen core will get, potentially, two 128bit pipes each, one ADD and one MUL. (since the core has 4x128bit, 2x add,2xmul). (Vs Broadwell having dual 256-bit pipes capable of both). Due to the nature of the SIMD operation two threads cannot independantly use one pipe so some of the FPU width on the intel arch (non Server architecture) is wasted when doing smaller operations.

Also interested in ISA used etc.

Anyone have any information? thx

 

[phill]

To be honest mate, unless you literally do the testing youreself, I'm not sure there's anything out there that would give you the data you'd need. Every system is setup differently and with that, some people might limit the WCG processes or only run for a few hours a day.

The testing I did was just on what wattage it was pulling with the system, I never really worried about much else. I mean Boinc does have it's own benchmark, but how reliable that is, I'd rather not put a number on it as I'm probably miles out..

What I can tell you is that I believe with my 5960X @ 4.20Ghz, it'll produce about half of what a Ryzen CPU can do, so whilst I would say Intel do suck they might be better off in certain work loads. Without testing things for hundreds of hours which I'm sure none of us have the time to do to get a decent reading, I just go along with, it makes whatever it makes and I'm happy with that.
Over time, your stats will build up and it will show you what you make, each day, over a week and a month (28 days) as averages. The charts I do for both WCG and FAH, can always be sent to you to see what I do each day for each of the units, WCG and FAH.

If there's anything else I can help with just ask, I'll do my best to find out for you

 

[hat]

What I can tell you is that I believe with my 5960X @ 4.20Ghz, it'll produce about half of what a Ryzen CPU can do

This bugs me. How can an 8/16 Intel chip only do half of what Ryzen can? It's also an 8/16 chip, with inferior IPC. How can Ryzen be so good at WCG specifically?

 

[phill]

I wish I knew @hat but that's what it does, about 11k a day and uses twice the power to do it lol But this is the problem with WCG, it's not consistent points or work units, they are all over the place and one day they can be through the roof and another day, rock bottom. I would love to know what changes...

 

[ArbitraryAffection]

This bugs me. How can an 8/16 Intel chip only do half of what Ryzen can? It's also an 8/16 chip, with inferior IPC. How can Ryzen be so good at WCG specifically?

To be fair Zen1 has similar if not better IPC to Haswell. Also a wider core and finer granularity in FPU engine. This potentially could allow it to gain significantly more performance with SMT in floating point heavy workloads.

 

[silentbogo]

This bugs me. How can an 8/16 Intel chip only do half of what Ryzen can? It's also an 8/16 chip, with inferior IPC. How can Ryzen be so good at WCG specifically?

To be fair Zen1 has similar if not better IPC to Haswell. Also a wider core and finer granularity in FPU engine. This potentially could allow it to gain significantly more performance with SMT in floating point heavy workloads.

Plus, in WCG there is no penalty for cross-CCX communication (if you remember Ryzen launch conundrum), since each thread/task runs on its own core independently.
Also, don't forget Spectre/Meltdown patches on the Intel side.
"Twice as fast" is overstatement, but "significantly faster" is quite realistic.

 

[ArbitraryAffection]

Plus, in WCG there is no penalty for cross-CCX communication (if you remember Ryzen launch conundrum), since each thread/task runs on its own core independently.
Also, don't forget Spectre/Meltdown patches on the Intel side.
"Twice as fast" is overstatement, but "significantly faster" is quite realistic.

Very good point. With 16 tasks essentially independent of each other, it scales very well to the modularity of the CCX design.

 

[hat]

Very good point. With 16 tasks essentially independent of each other, it scales very well to the modularity of the CCX design.

But if you have too many cores (high end TR, EPYC) it's a crapshoot, so they say.

Are those Smeltdown patches known for impacting WCG performance? In many tests, before/after results were largely the same. In some very specific workloads, performance drops... but I wouldn't think WCG would really be affected.

 

[ArbitraryAffection]

But if you have too many cores (high end TR, EPYC) it's a crapshoot, so they say.

Are those Smeltdown patches known for impacting WCG performance? In many tests, before/after results were largely the same. In some very specific workloads, performance drops... but I wouldn't think WCG would really be affected.

Perhaps somehting to do with the memory architecture and MCM nature of the High Core Count Zen parts? IDK about the patches though.

Spoiler: 1200 cruncher rig set up^^)

This is using salvaged parts haha. I don't really wanna put a 8 core in this board as the VRM is naked (no Heatsink). But it hsould be fine with a stock1700 maybe. I may get another one at some point. Till then the humble 1200 will do its part

Spoiler: The farm is complete

And so is the spaghetti, apparently O_O

These are all hooked up to one socket outlet, on that Surge tower. The one next to it is a powerline that goes to my router downstairs. Thats connected to the switch inbetween the two PCs there which hooks all the cams and rigs to the one ethernet on the switch~ Though the 1200 rig is using a wifi adapter as i didnt have a spare port on the switch. ~shrug~

This isn't dangerous is it? It's not using a lot of power and i'm not using dodgy cables or anything. Just anxious now though :/

Oh! I forgot to mention i got two really neat USB switches so i can use 2 mice, and 2 keyboards between 4 PC's and switch them with a button press. super handy^^

 

[hat]

You can always pick up one of those Kill-A-Watt devices to see how much power you're pulling from the wall. Plug your surge strip into it, and it will tell you how much you're pulling.

FWIW, I have both the rigs in my system specs (all 100% load, including the graphics cards, which pull 240w on their own), including my monitor, sound bar, a box fan, clock, and PS3 all connected to a single surge. It's not a problem.

I do want a Titanium power supply for my main rig though, for a number of reasons. The increased efficiency will help out a bit with this rig which is at 100% load all the time. It'll decrease power draw, and therefore heat, which will be nice on warmer days. On colder days, well, I'll just crank up my graphics cards some more (they're at 80% now) and get more heat. The power supply that's in there now really shouldn't be powering two 1070s mining and an overclocked 2600k running WCG 24/7 anyway...

 

[ArbitraryAffection]

You can always pick up one of those Kill-A-Watt devices to see how much power you're pulling from the wall. Plug your surge strip into it, and it will tell you how much you're pulling.

I have a wattage meter there but it's not set up correctly, it can do cost in £ over time too but i just use it for watts. 1200 is up and crunching. the 2 1700's, 1200 and 5350, plus 2 monitors and 4 IP cams is taking 377W at the wall right now. My main PC is probably pushing that up to around 450+, and the 200GE might mean overall use is near 500W mark. Gonna run it for a month and see what the bill's like, and see if its sustainable. Otherwise i might, try and take out some fans and RGB from my main PC and/or underclock the CPUs when crunching. But I think it hsould be fine..

 

[hat]

You can always try undervolting a little without underclocking. Not sure how much it would be worth, though, and then you've got to make sure it's still stable...

 

[ArbitraryAffection]

You can always try undervolting a little without underclocking. Not sure how much it would be worth, though, and then you've got to make sure it's still stable...

Could probably shave off 50 or so watts from the whole setup with a bit of voltage tweaking but not sure if im comfortable doing it, as you say it would need to be stability tested.