For FFMPEG (x265) are 4 x 4650 v2 > 2 x 2680 v2?

[Cutha]

Hi,

I am struggling to find enough information to make an educated choice as there is not many bench marks for quad socket systems. As x265 encoding is NUMA aware, are 4 x E5-4650 v2's better for encoding x265 then 2 x 2680 v2's? When I looked at the PassMark listing for Multi CPU's I was a bit confused. I know there is some overhead cost for 2P and 4P systems but the numbers seem crazy low.

I have a bit of an opportunity to change things up if I want and after this I might be stuck for several years with what I have. I have been mulling over this for months and I need to pull the trigger or just forget about it.

On the PassMark website I couldn't find scores for the E5-4650 v2 so I guessed based on the 2670v1, 2670v2 and the 4650v1 and figure it's about ~14,000.

On ebay.ca right now, the E5-4650 v2's are about $300CAD. I had considered going with a Ryzan (TR) but the benchmarks I saw they seem hindered in x265 encoding.

Thanks for your time,
Seamus

 

[StammesOpfer]

Passmark is skewed by single thread performance. So if you have something that can use all the cores/threads then they will scale better than passmark would indicate. The fact that we are talking video encoding and passmark scores leads me to ask if this is for a Plex Server (that tends to be when these two subjects meet)? Have you looked into something that supports hardware transcoding instead of doing it all in CPU?

Might look at the nvidia quadro p2000 since it does not have a maximum number of streams.
Helpful transcode matrix for nvidia cards. Video Encode and Decode GPU Support Matrix

 

[Cutha]

Thanks for the reply.

Yes, primarily Plex. Yes I have looked into hardware transcoding. Currently it's on a box with 2 x 2680 v2. I was excited about the hardware encoding in Plex until I realized that it has the same issues as all (as far as I know) hardware encoding, much larger files for the same quality. My upload bandwidth is not great so increasing the size is not an option. Our cable provider just moved to DOCSIS 3.0 but didn't increase upload speed. I was holding my breath for two years for this upgrade...

I did some testing with x265 and it worked out great with my friends on the client side. I just wish Plex would provide an option to encode to x265 for remote sessions. At one point I started converting my videos to x265.

While trying to learn about all of this I did see many posts about x264 NOT being NUMA aware and performing horribly. I couldn't seem to find any updates on this. Do you happen to know anything about this?

Are 4 x E5-4650 v2's better for encoding x265 then 2 x 2680 v2's? I will try to look again for reviews that might have this info.

Thanks,
Seamus

 

[StammesOpfer]

Well the short answer is it depends. As you already discussed the NUMA awareness. If you are running multiple simultaneous encodes you should be able to run 4 encodes 1 tied to each socket. To me I don't think it would be worth it though to move to the size and power consumption of a 4p system especially since you don't need it to be real time. You just have it automatically run through you entire library. Heck you might even be better off adding a second system and split the load between 2 computers and you can use the second one for something else or sell it once you get through your backlog of content.

 

[Cutha]

Thanks for your reply. I was hoping more for an enthusiastic YES its worth it, quad socket all the way! It would help justify my urge to try it out.

I actually have 4 SM dual socket servers with E5-2670's in them. I will see what I can figure out about automating the process between them.

Would a single E5-2600 series CPU work in a quad socket system? I tried to look it up but google told me that I am asking a dumb question and it has no answers for me, at least the way I asked it.

 

[WANg]

Thanks for your reply. I was hoping more for an enthusiastic YES its worth it, quad socket all the way! It would help justify my urge to try it out.

I actually have 4 SM dual socket servers with E5-2670's in them. I will see what I can figure out about automating the process between them.

Would a single E5-2600 series CPU work in a quad socket system? I tried to look it up but google told me that I am asking a dumb question and it has no answers for me, at least the way I asked it.

The Sandy-EPs are dual sockets only. Just look at the Intel ARK.

 

[wildpig1234]

Would a single E5-2600 series CPU work in a quad socket system? I tried to look it up but google told me that I am asking a dumb question and it has no answers for me, at least the way I asked it.

e5-16xx, e5-26xx, e5-46xx are all interchangeable so yes, a single e5-26xx cpu will work in a 2011 MB whether that MB is single or dual or quad socket.

 

[Cutha]

e5-16xx, e5-26xx, e5-46xx are all interchangeable so yes, a single e5-26xx cpu will work in a 2011 MB whether that MB is single or dual or quad socket.

Thank you. I could kiss you but I think we would all be more comfortable with a "like" instead.

 

[wildpig1234]

i would kill for an affordable 2011 MB and four decent clock e5-46xx cpu ... they are still hugely expensive compared to e5-26xx most of the time....once a while you do come across a seller with reduced price on the cpu because seems like not a lot of people actually search for them so they either go for low bid or the seller has to sell low due to lower demand. the MB is however always expensive... very expensive. and you need a custom case..

 

[wildpig1234]

hmm.. actually not that expensive:

Supermicro 4U 848A-R1K62B / X9QRI-F+ / Barebone Server / 24x Trays / Rails | eBay

I would definitely get this compared to just MB since this has PSU and the custom case. it is big and heavy though... The cheapest option would be 4x e5-4650v1 cpu at $50-60 each giving 2.9ghz x 32 cores or theoretical multicore of 92.8ghz.

One epyc 7601 is 2.7ghz x 32 cores= 86ghz.....lol...and only use 165w compared to 130w x4

This compared to dual 2697 v2 at 72ghz. you could do 4x e5-4657L v2 for total of 129ghz but each cpu is $500 ;(

The thing though is that i already have a dual e5-2696 v3 setup with total multicore of 100.8. so i probably won't buy into this. But if you need like something for massive amount of cheap ddr3 ram, this 4 socket v2 thing might be good..

 

[Cutha]

Thanks for posting that. I came across the same setup as you posted except they ship to Canada and that is actually what prompted me to start this thread.

I have been going around in circles over this. I was going to buy a 4U 24 or 36 bay Supermicro server anyway and one of the 4 socket systems caught my eye. I actually had considered a new setup with a higher core count. My CDN $ don't go as far as your USD bucks, perhaps that plus shipping costs to Canada make up the difference in the E5-4650v1. The prices I listed here are what I could find on ebay that are located in NA and the seller had quantity to fill a board (2 or 4 CPU's).

I figured if I did this table up it would point me in a clear direction. And I think it has, the E5-4620v2 has 50% more processing power than my quickest server and as the cost for the Supermicro system is not much different between the 2P and 4P servers by the time it lands on my door. Except of course I need to buy 4x E5-4620v2 CPU's. If I went with the Epyc 7601, with just the cost of the CPU, it would take almost 24 years before the energy savings paid for itself. I obviously didn't account for energy consumption of RAM and chipset.

I couldn't figure out how to get a cheap system for the E7-2890v2. Nice specs and cheap...

I have more than blown my computer budget this year, I just ordered 100 3TB HGST drives, 6 DC 3700 400GB SSD's, and 4x Ryzen 1800x w/MB and RAM (one each for me and my boys). If I order the 4P system it will be without CPU's for now and use a single E5-2600v2 CPU in it.

Looking at this I might switch a couple of my E5-2670v1 servers to E5-2620v2's. They are cheap and the reduction in electrical load might keep me under the threshold of my UPS.

Well this took me an embarrassingly long time to put together and now I am tired as hell and I likely made mistakes but I am just going to post it as it is.

Thinking on this, I guess it would make much more sense to have at least a rough number for IPC differences between gens and different uArchs instead of a total Ghz.

Name        Cores  Ghz  GhzTrb Sckt tGhz  $/CPU    tot$    $/Ghz    W/CPU  tPower W/Ghz   $/Yr    Yrs for Epyc payback

E5-4650 v1    8    2.7   2.9   4    92.8   $150    $600    $6.47    130    520    5.60    $484    13.9           
E5-4650 v2    10   2.4   2.7   4    108    $270    $1.1k   $10.00   95     380    3.52    $353    21.0           
E5-4640 v2    10   2.2   2.5   4    100    $450    $1.8k   $18.00   95     380    3.80    $353    17.2           
E5-4627 v2    8    3.3   3.5   4    112    $450    $1.8k   $16.07   130    520    4.64    $484    10.1           
E5-4620 v2    8    2.6   2.8   4    89.6   $145    $580    $6.47    95     380    4.24    $353    23.7           
E5-4657Lv2    12   2.4   2.7   4    129.6  $3.5k   $14k    $108.02  115    460    3.55    $428    0.0           
                                                               
Epyc 7601     32   2.2   2.7   1    86.4   $5k     $5k     $57.87   180    180    2.08    $167    0.0           
                                                               
E5-2670       8    2.6   3     2    48     $100    $200    $4.17    115    230    4.79    $214    102.3           
E5-2628Lv2    8    1.9   2.1   2    33.6   $100    $200    $5.95    70     140    4.17    $130    #DIV/0!           
E5-2630Lv2    6    2.4   2.6   2    31.2   $75     $150    $4.81    60     120    3.85    $112    #DIV/0!           
E5-2620 v2    6    2.1   2.4   2    28.8   $45     $90     $3.13    80     160    5.56    $149    #DIV/0!           
E5-2670 v2    10   2.5   2.9   2    58     $195    $390    $6.72    115    230    3.97    $214    98.3           
E5-2680 v2    10   2.8   3.1   2    62     $235    $470    $7.58    115    230    3.71    $214    96.6           
E5-2690 v2    10   3     3.3   2    66     $375    $750    $11.36   130    260    3.94    $242    56.8           
E5-2697 v2    12   2.7   3     2    72     $430    $860    $11.94   130    260    3.61    $242    55.3           
                                                               
E7-4860 v2    12   2.6   3     4    144    $1.2    $4.8k   $33.33   130    520    3.61    $484    0.6           
E7-4870 v2    15   2.3   2.7   4    162    $300    $1.2k   $7.41    130    520    3.21    $484    12.0           
E7-4880 v2    15   2.5   2.9   4    174    $510    $2k     $11.72   130    520    2.99    $484    9.3           
E7-4890 v2    15   2.8   3.2   4    192    $588    $2.3k   $12.25   155    620    3.23    $577    6.5           
E7-8857 v2    12   3     3.4   4    163.2  $390    $1.6k   $9.56    130    520    3.19    $484    10.9           
E7-8880 v2    15   2.5   2.9   4    174    $325    $1.3k   $7.47    130    520    2.99    $484    11.7           
E7-8891 v2    10   3.2   3.5   4    140    $450    $1.8k   $12.86   155    620    4.43    $577    7.8           
                                                              
E7-2850 v2    8    2.3   2.6   2    41.6   $850    $1.7k   $40.87   105    210    5.05    $195    116.6           
E7-2870 v2    15   2.3   2.7   2    81     NA      #VALUE! #VALUE!  130    260    3.21    $242               
E7-2880 v2    15   2.5   2.9   2    87     $390    $780    $8.97    130    260    2.99    $242    56.4           
E7-2890 v2    15   2.8   3.2   2    96     $130    $260    $2.71    155    310    3.23    $288    39.1

 

[wildpig1234]

e5-2640 v2 is 95w vs e5-4650 v1 130W tdp..... i guess you save a little on power but youre paying nearly twice as much per cpu. they are about the same speed i think....

yeah i looked at this before a while back.... the cpu entry price for E7 are very attractive... rock bottom price for each cpu given these have 15 core. Its too bad that the MB are very rarely avail by itself so you have to get a whole system which are all well over $3K and are all very big and heavy.... i think if i could just find a MB by itself i could just get that and leave it out on a desk....lol

 

[Cutha]

e5-2640 v2 is 95w vs e5-4650 v1 130W tdp..... i guess you save a little on power but youre paying nearly twice twice as much per CPU.

Twice as much on eBay? Do you mind sharing your search? Sometimes I am blind.

i think if i could just find a MB by itself i could just get that and leave it out on a desk....lol

I have had the exact same thought!

 

[wildpig1234]

sorry my mistake. i mean e5-4620 v2 compared to 4650 v1....lol

 

[wildpig1234]

There is not really a way except to get the whole system for these E7.....it looks like the MB aside from being all nonstandard, they all use weird things like memory and i/o riser cards and very odd oversized board.... by the time you can make a custom case for it, the time and effort would be about as expensive as just getting the whole system ;(

Here;s link to a supermicro E7 4 socket:

Supermicro | Products | Motherboards | Xeon® Boards | X10QBI.
Supermicro | Products | SuperServer | 4U | 4048B-TRFT

you can download the manual and see the details.

 

[wildpig1234]

Here's a relatively "cheaper" E7 v2:

Fujitsu Primergy RX4770 M1 No Memory / No CPU / No Hard Drives CTO Rack Server | eBay

Cisco UCSC-C460-M4 C460 M4 Server Chassis w/o CPU / RAM - Fully Tested | eBay

Dell R920 40-Core Server 4x E7-4830 v2 2.2GHz 64GB-8 | eBay

I am still looking... but have not been able to find a complete barebone system for anywhere even close to $1500 yet....

 

[msg7086]

Also wanna add that x265 does get slow when using too many threads, same as x264 but with higher threshold. First with --wpp, then with --frame-threads, should work well up to 40 threads. After that, --pmode might be used to saturate a bit more cores, and from there --pme which helps almost nothing.

Streaming in hevc does help on encoding efficiency, but the compatibility is the problem. Otherwise you can just encode everything beforehand and enjoy direct streaming without any re-encoding, which, also hurts the quality a lot.

Either CPUs are v2 which means they do not support AVX2, and that's a huge disadvantage. Some test shows a speed difference of 30% on x265 encoding. A E5v3 or an i7-4770 would be more suitable for this task. Ryzen should also do the job very well considering the price point.

And 4 sockets is IMHO really not a good choice. You would feel huge pain managing NUMA nodes for each process. I have a few 2xE5 nodes for x265 encoding, and I'm already tired of managing CPU affinity again and again.