ES Xeon Discussion

[Jointer]

About the power needed - the power specs for the CPUs should be correct.
My 2x 2667 V4 consume around 300W in full CPU load, around 110W in idle. Specs is 130W i think, but remember, ES runs on lower frequencies.
Which is surprisingly good compared to 4770K, which idle at around 60-70W if i remember correctly.
The system have one watercooled TitanX Pascal so far and when that is also fully loaded, i am around 500W-550W max.
Measured by power meter on input side, PS is EVGA Supernova P2 750W.
The only (very) important thing is to get PS with two CPU power 8pin connectors.
So normal power supply is fine - you cannot overclock the CPUs, so do not need high power PS.
650-750W for one GPU, for two, i would go for 850-1200W depending on GPU models, to be safe.

And there was discussion about temperatures - i am at 45-48C max with Noctua NH-U12S at 800RPM in full load on these 130W CPUs. But i was around/over 50C before the TitanX was watercooled, because the hot air was partially going also into CPU fans.
Also my de-lid 4470K without OC and with that monster Noctua D14 have higher temps (55-60C) in full load and low fan RPM...
Quite surprising.

 

[Jointer]

I still expected it to be worse, because 4770K is 84W CPU.
Due to that TIM issues, the CPU was throttling even on stock speeds, so a done delid with CLU, which improved temperatures dramatically.
Anyway, i am very happy with my system, except it is still not my primary system, because i am not able to get the new Samsung 960 EVO anywhere.

 

[Jointer]

Well, i wonder how that 2690 V4 goes, i think the ES have the same frequency as 2667 V4, 2.6-3.1GHz or so and 14 cores. And in the retail specs, it is only 135W, which is hard to believe. None on Ebay right now, but i believe there were some yesterday.
This should be good upgrade... in few years, because 32 threads around 3GHz seems to be enough for my home use for some time.

BTW i should get two HP gen 9 servers with 2x 2687W V4 for work (no idea when), price somewhere around 8000 Euro or so (i do not pay for that, so do not care, lol), anyone want a Cinebench run on them?

 

[Toysrme]

Anyone running the es 2667 v4? What kinda R15? What kinda clocks are people seeing?


Sent from my iPhone using Tapatalk

 

[Jointer]

Anyone running the es 2667 v4? What kinda R15? What kinda clocks are people seeing?

ES Xeon Discussion

 

[Meh]

In terms of a workstation (for the science!) build that will part time as a gaming PC:

Scenario (assume GTX1080):
I have $1000 to spend on the CPU(s). There's a highly demanding game coming out in 2017 that I know will feature Vulkan and DX12 support. It's the only game I plan on playing, but I plan on playing the heck out of it for a long time.

6900k w/ 16 threads @ 3.2GHz > 3.7GHz
or
(2x) 2686 v3 QS QG7V w/ 72 threads @ 2.0GHz > 3.5GHz
or
(2x) 2685 v3 QS QGPK w/ 48 threads @ 2.6 > 3.3GHz

Thank you for your help and please pardon my ignorance. I've read most of this thread, but I'm still sifting through all of its implications.

 

[Jointer]

I guess it is Star Citizen.
I really doubt it would be able to use 16+ threads effectively, also in theory, if you have a game which max out 4 cores at 80% and you upgrade to 8 core, you will see 40% CPU usage and no FPS change. Because the app (game) just does not need more CPU power.
That is how it is with games, you are usually limited by performance by one thread, so go for CPUs with highest frequency.

What the dual CPU systems give you (or at least me) is ability to run more stuff at the same time.
For example you can run VMs, some apps and also game at the same time.

 

[Meh]

I guess it is Star Citizen.
I really doubt it would be able to use 16+ threads effectively, also in theory, if you have a game which max out 4 cores at 80% and you upgrade to 8 core, you will see 40% CPU usage and no FPS change. Because the app (game) just does not need more CPU power.
That is how it is with games, you are usually limited by performance by one thread, so go for CPUs with highest frequency.

What the dual CPU systems give you (or at least me) is ability to run more stuff at the same time.
For example you can run VMs, some apps and also game at the same time.

Good guess. Any thoughts on the QGPK for being able to handle both of these tasks? And you bring up a great point. Most of what I do requires a lot going on in the background and alt tabbing to check on it.

 

[Jointer]

Well, i am pretty able to run one low demand VM, have a browser open and game at the same time on 4770K.
Until i run out of 16GB memory.
CPU power seems not to be a big problem.
So really think if you really need dual CPU system for gaming, maybe one CPU would be enough.
I went with Asus Z10PE-D8 because i really liked the board and it was only 1/3 more money then workstation class 1x CPU board.
It could be better to go SLI, if that would be properly supported in Star Citizen.
I game at 2560x1600/60Hz and TitanX Maxwell is not able to run The Division with all settings at max. on 4770K.
But it does not look like the limit is the CPU.
I did not tested it on the new system yet (TX Pascal in it), the only game i tried on both was Rise of the Tomb Raider, had improvement from 70FPS to 112 or so.
I would recommend getting 1 CPU system with highest clocks and wait for nVidia 1080 Ti.

 

[Toysrme]

Oh god don't rely on star citizen to do shit. It's programming is god awful; even the thread scheduling is pitiful. The live game crashes with more than 32 cores because the thread scheduler is a joke. The PTU builds are not much better.





Please god as a Gold Ticket backer in star citizen and someone a QA lead once said was the most helpful backer to that game's development..
STOP GIVING THEM MONEY!?!



Furthermore, It's just Cryengine 3. It scales nearly mathematically in SLI per GPU. SLI was enabled in sept/oct of 2014 and works as well as you have CPU to feed the scaling.


Threads won't help get passed absurdly low clock rates on xeons. Even in new API's with the most experienced multithreading+ASYNC compute programmers. It's not a magic bullet.



40 threads at 2.3ghz is crushed by mediocre 4.5ghz on 4-8 threads even in the best cases.


All benchmarks are the average of 3 runs. In the two events (ROTR, Ashes) where very odd results were recorded, the best three of 5 runs were used.


Deus Ex: Mankind Divided (min,avg,peak)
dx11
2c 0.2, 19.5, 39.2
4c 29.4, 40.6, 67.3
8c 37.2, 54.3, 80.6
12c 31.8, 56.8, 78.4
16c 39.8, 57.1, 79.6
20c 39.0, 57.3, 81.3
40c 38.1, 55.5, 83.3

dx12
2c 0.7, 31.6, 52.9
4c 12.9, 55.0, 81.3
8c 14.0, 66.4, 103.4
12c 17.9, 66.8, 103.9
16c 14.7, 69.4, 105.3
20c 24.2, 68.4, 103.4
40c 21.1, 66.5, 99.5

Rise of the Tomb Raider (min geothermal valley, average)
2c 8.63, 47.91 (heavy object+texture pop in)
4c 21.46, 68.16 (object+texture pop in)
8c 30.61, 78.86
12c 29.77, 80.38
16c 31.90, 80.94
20c 33.90, 82.47
40c 32.44, 77.3

dx12
2c 6.1, 49.39 (heavy object+texture pop in)
4c 12.44, 91.88 (object+texture pop in)
8c 42.92, 119.31
12c 85.48, 122.29
16c 91.64, 122.60
20c 89.64, 121.65
40c 74.79, 120.87

Battlefield 1 (avg fps)
2c 17.7
4c 41.2
8c 48.5
12c 49.1
16c 49.1
20c 49.6
40c 47.8

dx12
2c 18.2
4c 40.8
8c 47.4
12c 47.8
16c 47.2
20c 47.3
40c 47.1

Ashes of the Singularity: Escalation
dx12 (AVG CPU FPS) REMEMBER this game caps at 16 CPU threads!!!
2c 0.9
4c 15.3
8c 27.1
12c 30.5
16c 34.4
20c 34.7
40c 33.8

Civ VI
2c 26.39
4c 25.58
8c 24.58
12c 24.19
16c 23.94
20c 23.57
40c 24.31

dx12
2c 25.38
4c 23.97
8c 23.96
12c 23.59
16c 22.94
20c 22.82
40c 23.59

Star Citizen (hacked spawning 5 connies in Vanduul swarm and allowing them to fight for 60 seconds)
min, max, avg
2c
4c
8c
12c
16c
20c
40c 21, 133, 77.095

(Not finished because star citizen can't run for 2 minutes without crashing)





^^ Testing I was doing in off hours before I sold those chips.







Definitely the sweet spot core wise is 12-20 threads, BUT the faster cpu design (Skylake) at only 8 threads is more than fast enough to power ahead of the threading a 4.5ghz/166.67 Haswell-e with 12 threads or 2.8ghz (main thread) with 2.3ghz 12-40 threads.

Clock rate & IPC matter much more than just having a "higher than normal" amount of threads do. You also see even in the best cases available that the extra threading does not increase speed, but simply reduces CPU load on the non-main core.


Sent from my iPhone using Tapatalk

 

[Toysrme]

The problem with new API's in the handful of titles that work well with them:
1) Minimum frame rates are consistently inconsistent rubbish
2) Visual effects are muted. dx12/vulkan(Doom) do not support the top graphic effects dx11 does.


Sent from my iPhone using Tapatalk

 

[Jointer]

Great, finally someone who can bring a bit of reality to gaming on these systems.
I was not able to find any comparison like this, so thank you very much, Toysrme!
BTW these tests were done by switching cores off on some (same) Xeon CPU?

Nice numbers on the Tomb Raider, i think i tested in DX11 only, because until today, there was no game which would be noticeably faster in DX12, so i did not even try.

And also, i think the lowered performance with 20+ cores could be simply caused by having too much cores - Windows scheduler might "spread the load threads" between all cores to spread the load evenly across available CPUs, which cause inefficiency by migrating the needed data between caches and even sockets.
I even experienced something like this once, i had a "single threaded" app of some sort, which was loading all cores evenly. When i adjusted affinity to run only on one non-HT core, the performance increased slightly, at the "expense" of one core being loaded at 100% all the time.

EDIT: i tried quick benchmark in Tomb Raider (the only game i have on that system) at 1600x1200, with TX Pascal - DX11 118FPS, DX12 136 FPS, hooray.
The GPU runs at 2100MHz.

 

[Toysrme]

Yes, ROTR initially didn't do well in dx12, but they continued working on it and it responds pretty well. Seemingly will put work across however many threads you have/allow.


Deus Ex: Mankind Divided is the multi threading poster child. It'll spawn threads


Sent from my iPhone using Tapatalk

 

[Meh]

So under the assumption a game perfectly uses all threads available (in hypothetical fairy land):

4 GHz on 16 threads = 2 GHz on 32 threads
That is to say, the game may use 16 threads @ 40% each or 32 threads @ 80% each

Could this then be expressed as each machine having 64 units of bandwidth? How and when do boost/turbo come into play in this equation?

As is, performance would theoretically be equal; Except computers tend to do best with fewer components with more power...

I would be leaning toward a 6850k for 40 lane SLI availability (just in case they don't totally phase it out) if it weren't for the incredible pricing on these ES/QS Xeons and the fact I still use this as a workstation 95% of the time.

 

[Toysrme]

Please remember that there will always be more important things happening on the main thread that are required to finish before advancing; no matter if it's a new or old API. New API's just allow more distribution of rendering tasks.

As it stands now turbo boost differs between all bins. What I found was that in gaming; the workload is generally still light and all turbo clocks were hit. So the main core would be at max turbo, the others would be at the full load turbo.
For the ES e5-2698v4 (no OC) that base 2.0ghz;
I would have 2.79x ghz main thread and 2.2x ghz all other threads. (Something along those lines)


Where as cinebench r15 would crush everything and downclock bins down to 1.8-1.9ghz.




For gaming, even with the three good dx12 and single Vulkan title; A 4 core Skylake i5 will CRUSH a high thread count Xeon.



Sent from my iPhone using Tapatalk

 

[Toysrme]

No, actually Anandtech is with me on this one. Intel Haswell-EP Xeon 12 Core Review: E5-2650L V3 and E5-2690 V3
http://images.anandtech.com/doci/8679/12C Xeon Frequency Reponse.png?_ga=1.255507464.1297422707.1472075143




They're all the same CPU silicon designs. All of the CPU's are binned for voltage leakage and frequency response ontop of simple failures with portions needing disabling.

As Anandtech showed (and anyone that's also had to add minimum power consumption to complete an exact task) has seen. Despite being the same designed silicon; you get different frequencies across all the different chips running the same workloads on the same number of cores.





Oh, and you're confusing what is ancient GPU bound benchmarks for showing *anything* about CPU performance. That's a big failure on your part bro.

 

[pesceman3]

Anyone have experience with the E0 2690 v4 QH2M? Not sure I would want to try such an early ES, but you never know.

 

[Patriot]

No, actually Anandtech is with me on this one.
Oh, and you're confusing what is ancient GPU bound benchmarks for showing *anything* about CPU performance. That's a big failure on your part bro.

No need to be patronising... or dismiss data without cause.
Both of you need to be more civil. This forum is a search for knowledge not a fisticuffs match.

Games like any workload can favor different designs on an individual basis.
Anandtech had good data, as Toys pointed out most games are gpu limited, This however does not make the data invalid. If you click on minimum frames you will see a bit more interesting side of things.

Games do currently favor fewer higher clocked cores. Some are multithreaded friendly and the hit is not as bad. Some are completely GPU bound and it matters little to none what cpu you have.
You should build your computer based on your most important workload, and remember, jack of all, master of none.

I currently game on a Xeon 1680v3, yes in some games a 4790k is faster. No it does not stomp all over it, nor are the frames missed... Once you hit 120+ fps it really does not matter.

 

[Patrick]

@Toysrme @Nanotech et. al. Opinions are great. Just want to remind folks to keep it civil as @Patriot stated

 

[Toysrme]

I'm wondering at what point letting someone know they are wrong became "uncivil" @Patrick @Patriot because that's a personal/societal/culture problem if they can't handle it.

I've not been offended by anything said to me and would be shocked to know someone couldn't handle being told they're wrong and why...

Added in GTA-V and after 4 days of trying; finished the Star Citizen benchmark.