Plex server CPU w/ dedicated GPU

[Bluntwrapped]

Yo guys, I’m looking into purchasing either a Xeon v3 or v4 cpu as part of a server build.

I’m looking to have a fileserver (NAS) with Plex and dedicated GPU for transcoding, a Windows based VM with dedicated GPU for AE and other rendering projects as well as a Linux based VM for PiHole, OpenVPN and perhaps a few other clients.

I have the board - x99 WS/IPMI
RAM - 2x32gb Samsung ECC DDR4 2400MHz
And a current cheap Xeon E5-1603 V3 for testing.

The 1603 V3 only supports up to 1866MHz ram, and while both sticks run fine at 1866MHz, it’s with the stock CL 17 timings (meant for 2400MHz) as I haven’t been able to find 1866MHz timings from a similar kit.

Basically, I’m looking for a CPU with enough cores and grunt to cover all of the above without compromising too much, + leave some flexibility for the future. Now I know only the V4’s (aside from the unlocked 16xx V3 series) support 2400MHz DDR4, I’m leaning towards one of the V4 CPU’s.

Having never used Plex, what I’ve read is it can now make use of GPU transcoding (maybe still in beta stage) GTX cards are limited to 2 streams without modding, and Quadro cards (P2000) and above are limitless.

Many seem to now use and recommend P2000’s after doing some googling.

What I’m trying to figure out is, if a dedicated GPU is handling Plex transcoding, does that also mean fewer CPU cores will be required, compared to transcoding with CPU only?

Hope this makes sense. If anyone has any experience, advice or feedback I’d greatly appreciate it

Cheers

 

[Avalon]

Do more reading.

PiHole uses bugger all resources, it’ll run on a Pi. OpenVPN will be relatively light assuming you use hardware acceleration. Be warned, ASUS X99 boards have a habit of over-volting chips.

HW transcoding requires PlexPass, you either pay monthly, annually or buy lifetime. If the GPU is doing the work, the CPU only has to do audio or sub transcoding which uses a lot less resources. The majority of people using P2000’s are likely doing so because they didn’t read and followed YouTube guides like sheep.

Plex 101:

1. Do not transcode 4K in software (CPU). Do not transcode H265. Never transcode HDR (until time mapping is fixed).

2. HW Decoding is only functional under Windows. It can be enabled in other OS’ but with variable results.

3. Plex only outputs H264 when it transcodes.

Unlimited Quadro transcodes means they are not artificially limited. It still has a limit, the quoted figures are 17x1080/10 to 720/4 transcodes or 16x1080/10 to 1080/8 at H264.

Here is where that becomes relevant. Local clients will generally direct play (unless you suck at buying clients, connecting them to your server or curating media). So where’s the need for a load of transcodes? If it’s remote, what uplink speed do you have and what quality will you be running? A standard 80/20 FTTC connection can only do 2x1080/8 transcodes, so how many concurrent transcodes do you need?

GPU wise the P2000 is arguably the worst value option available, it’s popularity is down to people making poor decisions because of hype. They change hands for £260-350 at present which is significantly up on what they cost 12-18 months back used. A 1050 Ti 4GB goes for half that or less and gives the same results with a driver tweak. If you want a Quadro then the M2000 is 1/4 the price and does 11/13 streams respectively. The caveat with the 960/M2000 series is they get H265 decode (under Windows only at present officially), but they can’t do HDR to SDR transcodes (FFMPeg tone mapping is broken anyway). So if paying 4x the price for a feature you can’t use sounds like a great idea, then feel free to buy a P2000.

CPU wise I like the v3’s for value, the ES chips can be stupidly cheap for what they are (check the stepping). 2K of CPU Mark is suggested for each software (CPU) 1080 transcode, so if you tell us how many transcodes you plan on having running g at the same time it’s easier to suggest a suitable CPU.

 

[Bluntwrapped]

Do more reading.

PiHole uses bugger all resources, it’ll run on a Pi. OpenVPN will be relatively light assuming you use hardware acceleration. Be warned, ASUS X99 boards have a habit of over-volting chips.

HW transcoding requires PlexPass, you either pay monthly, annually or buy lifetime. If the GPU is doing the work, the CPU only has to do audio or sub transcoding which uses a lot less resources. The majority of people using P2000’s are likely doing so because they didn’t read and followed YouTube guides like sheep.

Plex 101:

1. Do not transcode 4K in software (CPU). Do not transcode H265. Never transcode HDR (until time mapping is fixed).

2. HW Decoding is only functional under Windows. It can be enabled in other OS’ but with variable results.

3. Plex only outputs H264 when it transcodes.

Unlimited Quadro transcodes means they are not artificially limited. It still has a limit, the quoted figures are 17x1080/10 to 720/4 transcodes or 16x1080/10 to 1080/8 at H264.

Here is where that becomes relevant. Local clients will generally direct play (unless you suck at buying clients, connecting them to your server or curating media). So where’s the need for a load of transcodes? If it’s remote, what uplink speed do you have and what quality will you be running? A standard 80/20 FTTC connection can only do 2x1080/8 transcodes, so how many concurrent transcodes do you need?

GPU wise the P2000 is arguably the worst value option available, it’s popularity is down to people making poor decisions because of hype. They change hands for £260-350 at present which is significantly up on what they cost 12-18 months back used. A 1050 Ti 4GB goes for half that or less and gives the same results with a driver tweak. If you want a Quadro then the M2000 is 1/4 the price and does 11/13 streams respectively. The caveat with the 960/M2000 series is they get H265 decode (under Windows only at present officially), but they can’t do HDR to SDR transcodes (FFMPeg tone mapping is broken anyway). So if paying 4x the price for a feature you can’t use sounds like a great idea, then feel free to buy a P2000.

CPU wise I like the v3’s for value, the ES chips can be stupidly cheap for what they are (check the stepping). 2K of CPU Mark is suggested for each software (CPU) 1080 transcode, so if you tell us how many transcodes you plan on having running g at the same time it’s easier to suggest a suitable CPU.

First of all thanks for taking the time to write and reply some interesting stuff there, I’ll be sure to do as much reading in the coming weeks as possible.

Do you have a reference or link when you mentioned x99 over-volting? I updated the bios to the latest version 4001 (released only a few weeks ago) and wonder if this is something which has since been fixed.

I appreciate some of the points you raised, especially regarding tone mapping. I was using madVR with MPC-HC previously, but since I sold I sold my GTX1080 I haven’t been able to use it since. Doesn’t seem to work with my iGPU. HDR content is back to looking undersaturated and washed out.

From what I read last night, the latest Plex beta may have fixed or improved this?

I was looking at the 1050ti too so thanks for confirming. Again, I’ll do more reading.

I have BT ultrafast fibre (upgraded just over a month ago) I’m getting 300/40 on average.

I’m not entirely sure how many concurrent transcodes I’ll have as my plan is to give my friends and family access. (lets say 8 at a guess) my mum and dad travel a fair bit so my idea is to give them both access via their mobile devices.

From what I can see the V3’s have a few attractive qualities.

1. The availability and price
2. The option to run all cores at the single core boost frequency.

Still at work, supposed to be working so I’d better get back for now. I’ll re-read everything you’ve said later, just incase I’ve missed something.

 

[Bluntwrapped]

Deleted

 

[Avalon]

Google ‘x99 CPU Kill’ or similar, it’s something you probably want to keep an eye on, from memory a number of potential workarounds existed. You get a lot of V3 for your money, also i7 5820k/5960x/6800K etc. prices are getting quite reasonable if single core performance and overclocking is more important to your workload. For example I paid £25.99 delivered for a 2620v3, it's not going to set any records for performance, its only 2.4Ghz with a 3.2Ghz turbo, but that’s 10K of CPU Mark for £25.99. You are right that with a little work, it’s possible to use the all core turbo hack on v3's, your 16xx can however be overclocked.

Plex beta is aka PlexPass and while tone mapping fixes upstream are being worked on (Plex uses FFMPeg for this), I wouldn’t hold your breath. Plex has a nasty habit of focusing on other things in preference to historic fixes. Either way if you have an HDR capable screen, Plex still only outputs SDR H264 1080 when transcoding, tone mapping will come with a performance hit, it’s logically a better option in GPU than CPU, especially as most HDR tends to be H265. Have a look at https://www.elpamsoft.com/?p=Plex-Hardware-Transcoding for indications of what each card can do, paying P2000 money when a 1050 is just as capable is just silly. I paid £85 for a M2000 simply because it would fit nicely in an ITX build and my 980GTX didn't (297mm), i'd have been just as happy with a 960 that did, but the price difference was minor. It's worth remembering that if you have media in a suitable format that doesn't require transcoding, then the requirements to run Plex are very minimal (Pi2 onwards realistically).

With 40 up, you’re looking at something like 5x1080/8 transcodes max or say that’s 10K of CPU Mark as a guide if all 5 users transcode a 1080 H264 source and it’s done in software (CPU), I tend to see about 20-30% of my remote users online at peak times, more at Christmas etc. With a GPU doing the heavy lifting, you have a much lower CPU requirement.

 

[Bluntwrapped]

Google ‘x99 CPU Kill’ or similar, it’s something you probably want to keep an eye on, from memory a number of potential workarounds existed. You get a lot of V3 for your money, also i7 5820k/5960x/6800K etc. prices are getting quite reasonable if single core performance and overclocking is more important to your workload. For example I paid £25.99 delivered for a 2620v3, it's not going to set any records for performance, its only 2.4Ghz with a 3.2Ghz turbo, but that’s 10K of CPU Mark for £25.99. You are right that with a little work, it’s possible to use the all core turbo hack on v3's, your 16xx can however be overclocked.

Plex beta is aka PlexPass and while tone mapping fixes upstream are being worked on (Plex uses FFMPeg for this), I wouldn’t hold your breath. Plex has a nasty habit of focusing on other things in preference to historic fixes. Either way if you have an HDR capable screen, Plex still only outputs SDR H264 1080 when transcoding, tone mapping will come with a performance hit, it’s logically a better option in GPU than CPU, especially as most HDR tends to be H265. Have a look at https://www.elpamsoft.com/?p=Plex-Hardware-Transcoding for indications of what each card can do, paying P2000 money when a 1050 is just as capable is just silly. I paid £85 for a M2000 simply because it would fit nicely in an ITX build and my 980GTX didn't (297mm), i'd have been just as happy with a 960 that did, but the price difference was minor. It's worth remembering that if you have media in a suitable format that doesn't require transcoding, then the requirements to run Plex are very minimal (Pi2 onwards realistically).

With 40 up, you’re looking at something like 5x1080/8 transcodes max or say that’s 10K of CPU Mark as a guide if all 5 users transcode a 1080 H264 source and it’s done in software (CPU), I tend to see about 20-30% of my remote users online at peak times, more at Christmas etc. With a GPU doing the heavy lifting, you have a much lower CPU requirement.

Had a quick read last night about the X99 issue. Since ASUS flat out denied there was a problem (despite many dead CPU’s) the outcome and whatever was causing the problem is inconclusive. A few mention bios updates, and not having a problem since so here’s hoping, years later the maturity of the BIOS means no further issues.

My dad has been using a 6800k on a ASUS WS for years without any issues whatsoever. Kinda wonder if people were attempting to put silly volts through their chips after realising the HEDT line don’t overclock as well as the desktop counterparts.

One thing for sure is this 1603v3 is a piece of crap.. (putting it politely) 4 cores, no HT, no boost but 140w TDP.. In terms of silicon lottery it must be down there as one of, if not the worst..

The thing it has been good for is it’s made me realise the VRM heatsink heats up quickly, so whatever CPU I do go for I’ll make sure to have active cooling on it.

The link you shared above is brilliant I’ve bookmarked it thanks.

I suppose going back to my first question, if the GPU does most of the heavy lifting, how many cores and how high does the boost need to be to not be a bottleneck while streaming?

 

[Avalon]

TDP is a largely irrelevant measure of power usage unless you plan on running a CPU flat out 24/7 or have thermal constraints, but yea, 1603 v3 wasn’t really that great.

I ran a 2630L v3 for a while, I could get 6 or so concurrent transcodes out of it due to the way Plex handles transcoding buffers, I never needed more and it had a number of other dockers with moderate CPU/IO workloads running at the same time. I’d suggest setting a budget and having a look at what’s available in it. 2630L/2630’s are cheap and plentiful, that said for a little more you can get much more for your money.

 

[BigBoy]

I run a plexpass (windows native) with sonarr/Lidarr/Radarr/Heimdall/pihole/airsonic/booksonic/laseylibrarian and a couple of others running in docker. Sabnzb and deluge run Windows native too.

I have 7 external clients but never more than three concurrent external streams and one internal.

CPU is an E3 1260v2 with 24gb of ddr3 and 12tb of storage over 4 disks. Windows and transcoding is one on an SSD for the sake of speed. GPU is a Quadro K600 (able to do 2x1080p streams) then the CPU's onboard GPU picks up the slack.

CPU usage hardly ever goes above 60% and ram usage sits around 8-10gb

It's a low power system at idle 100w ish peak is around 200w.

Spare hardware is a good place to start with Plex servers but x99 gear is power hungry so don't expect it to be quiet or low power!

It all depends on what you want it to do and how much money you want to throw at it.

 

[Avalon]

I run a plexpass (windows native) with sonarr/Lidarr/Radarr/Heimdall/pihole/airsonic/booksonic/laseylibrarian and a couple of others running in docker. Sabnzb and deluge run Windows native too.

I have 7 external clients but never more than three concurrent external streams and one internal.

CPU is an E3 1260v2 with 24gb of ddr3 and 12tb of storage over 4 disks. Windows and transcoding is one on an SSD for the sake of speed. GPU is a Quadro K600 (able to do 2x1080p streams) then the CPU's onboard GPU picks up the slack.

CPU usage hardly ever goes above 60% and ram usage sits around 8-10gb

It's a low power system at idle 100w ish peak is around 200w.

Spare hardware is a good place to start with Plex servers but x99 gear is power hungry so don't expect it to be quiet or low power!

It all depends on what you want it to do and how much money you want to throw at it.

Where on earth are you getting your power numbers from? My 6800k/32GB (x99) averages 107w with multiple transcodes going and multiple VM’s running with pretty much the same as you, and that includes an old Netgear 24 port managed switch and two media streamers on the same socket due to a Y splitter, it’s also in my lounge with the side off and the loudest thing is the PSU fan (1200w of inefficient overkill because it was spare).

 

[BigBoy]

Where on earth are you getting your power numbers from? My 6800k/32GB (x99) averages 107w with multiple transcodes going and multiple VM’s running with pretty much the same as you, and that includes an old Netgear 24 port managed switch and two media streamers on the same socket due to a Y splitter, it’s also in my lounge with the side off and the loudest thing is the PSU fan (1200w of inefficient overkill because it was spare).

a cheap power socket wattage reader from ebay. probably not the best thing to go by

 

[Bluntwrapped]

I’m thinking about one of the L Xeon’s.

55-75w should be good, if I can get one cheap enough.

 

[Avalon]

I’m thinking about one of the L Xeon’s.

55-75w should be good, if I can get one cheap enough.

Unless you plan on running the CPU at 100% 24/7 or are operating them in a thermally constrained environment (eg 1U/power/cooling limited), then please stop looking at meaningless TDP numbers as an indication of real world power usage. It's the maximum TDP of the chip, not it's actual usage. Idle power consumption across a given architecture rarely differs significantly thanks to intel's power gating E.g my 6800k idles a few watts more than the 2630L v3 it replaced despite one being 55w TDP and the other being 140w TDP (one also general finishes CPU intensive workloads quicker and goes back to idle). I could drop a 2620 v3 in with similar results.

 

[krooton]

Yeah, I have an old 95w 2700k in my UnRaid box on 24/7, drawing a pittance, it is only ever pushed if transcoding, which is pretty rare.

 

[Bluntwrapped]

TDP is a largely irrelevant measure of power usage unless you plan on running a CPU flat out 24/7 or have thermal constraints, but yea, 1603 v3 wasn’t really that great.

I ran a 2630L v3 for a while, I could get 6 or so concurrent transcodes out of it due to the way Plex handles transcoding buffers, I never needed more and it had a number of other dockers with moderate CPU/IO workloads running at the same time. I’d suggest setting a budget and having a look at what’s available in it. 2630L/2630’s are cheap and plentiful, that said for a little more you can get much more for your money.

Yep that’s a fair point man. I was thinking overall consumption idle/flat out, but as you said, a higher clocked chip will get the job done quicker and may require overall less power.

Much to consider. The V3’s are plentiful and cheap while the V4’s have limited availability and are expensive.

I was thinking the L’s are more attractive long term with their lower power consumption, especially when compared to the likes of Ryzen but I take your point.

I’ll have another look and see what the options are.

 

[phill]

Honestly, I'm not sure how much difference a dedicated gpu makes with Plex.. Intel quick sync is required and makes a huge difference. I run a i7 6700 that can comfortably handle 4 hw transcodes, in addition to running blueiris with 7 cameras, which conveniently also utilises quick sync.

 

[Avalon]

It’s all relative to intended use. For a low power system focused on Plex where most things are direct played, with occasional transcodes, a low end CPU + iGPU set-up is ideal. A current gen i3 easily packs enough physical cores/grunt to do software transcoding and the iGPU to do the heavy lifting (obviously with PlexPass). In a mixed environment where you perhaps running more virtualised systems and/or dockers with reasonable workloads, then more cores give a lot more flexibility/pack a bigger punch, chuck in a modest Nvidia GPU with a patched driver or a cheap Quadro and it’ll handle anything you throw at it, including limited 4K with caveats.