Build an encoding RIG today

[Rev Jim Jones]

If you were going to buld a new encoding rig from the ground up today, what parts would you use?

Still go with Ryzen 1700 or the 8700K ? The 8700K overclocks to 5GHz easily so those 6 cores could probably outdo Ryzens 8 cores.

Handbrake scales out to 6 cores - right? So is a 6 core over clocked to 5GHz better then 8 cores at say around 3.8 to 3.9 GHz ?

Would you set M/2 drives to read/write encodes too? What about writing encodes to a RAM drive to ease wear and tear on an M.2 SSD drive? Maybe no need, maybe just write to a spinner? What advantage, if any, would having an M.2 drive/s (maybe Raid 0 M.2's?) have over just reading/writing encodes to a spinner?

How much RAM is optimal above what the OS needs ? Any advantage to having more RAM for an encoding specialised RIG ?

So, in short.....how would you build an encoding RIG today using whats available today?

[sophisticles]

If money is no object AND by encoding you mean strictly encoding and not encoding+filters, then I would pick up a dual Xeon Gold setup, or at least something that supports AVX-512.

http://x265.org/x265-receives-significant-boost-intel-xeon-scalable-processor-family/

The x265 people claim a 67% per core speed boost from AVX-512 for x265 and 40% boost for x264.

For confirmation we have this test:

https://us.hardware.info/reviews/7585/7/intel-core-i9-7980xe--7960x-review-intel-back-...-x265-and-flac

In this review using x265 via Stax Rip, a 7820X, 8C/16T, 3.6Ghz base clock speed, 4.3 Ghz turbo single core speed, actually beats a Ryzen 1920X, 12C/24T, 3.5Ghz base clock, 4Ghz max turbo and the 7820X has a gimped AVX-512 unit.

On the other hand, if you're using a codec that does not make use of AVX-512 and you are using filtering, we can look at this review:

http://www.guru3d.com/articles_pages/intel_core_i7_8700k_processor_review,14.html

As you can see, encoding to XAVC S Long, 4k, with 2 filters, no gpu acceleration, with audio, there is little significant difference in processor performance between a Ryzen 1600 and a Ryzen 1800X and a few processors in between, including Intel 8C/16T cpu's.

I've said this before and I'll say it again, unless you either A) get a very good deal on a new rig or B) your current pc dies and you have no choice but to build a new system, I would wait.

2018 is poised, in theory, to bring 2 very disruptive technologies to the desktop, AVX-512 and Optane DIMM, the latter which may redefine the way we look at ram and disk drives.

Optane DIMMS promise to bring some unheard of performance gains in many applications and in an odd way AMD has confirmed the wisdom of Intel's investment in Optane DIMMS by releasing this:

https://www.theverge.com/circuitbreaker/2016/7/26/12285568/amd-radeon-pro-ssg-graphics-card-ssd

http://www.pcgamer.com/amd-radeon-pro-ssg-pairs-vega-with-2tb-of-memory/

Supposedly, with this card, you can scrub through a time-line with 8k video and even render it out at up to a claimed 90 fps!

Then we have NVIDIA's Volta card, which we should see gaming cards based on this gpu in early 2018 and one has to believe that Volta will feature an improved NVENC unit, maybe even with hardware vp9 encoding and I know from using a GTX1050 on Ubuntu with ShotCut, if you encode to NVENC and use some good filtering techniques, like a bit or color correction, maybe a bit of sharpening and/or denoising, you have a hard time telling the difference between the hardware encoder and x264/x265.

[Rev Jim Jones]

WOW - lots of good info there. Going to take some time to get through all that.

I wonder how long it will take for these new technologies you talk of to actually be available

[sophisticles]

WOW - lots of good info there. Going to take some time to get through all that.

I wonder how long it will take for these new technologies you talk of to actually be available

AVX-512 is already here, though only the more expensive high end Intel cpu's feature full speed AVX-512 units; as for Optane DIMMS:

https://www.extremetech.com/computing/259155-micron-launches-new-nand-based-dimms-inte...s-optane-dimms

https://itpeernetwork.intel.com/new-breakthrough-persistent-memory-first-public-demo/

https://techreport.com/news/31932/optane-dimms-and-companion-cpus-will-arrive-in-2018

https://wccftech.com/intel-cascade-lake-xeon-scalable-platform-optane-dimm-2018/

Without a doubt, Intel Optane DIMMs have the potential to reshape how we think about computer architecture in general.

https://wccftech.com/intel-roadmap-leak-cascade-lake-x-hedt-coffee-lake-s-b360-2018/

Everything points to 2018 when we can expect to see HEDT parts that support both AVX-512 and Optane DIMMs BUT...a lot depends on AMDs Ryzen refresh scheduled to be released in February 2018, if AMD releases a lackluster "update" with minor clock speed improvements and some minor refinements in power consumption I would think Intel would be happy to compete with it's current lineup and push back release of it's premium technology until 2019, but IF AMD decides to drop a bomb on Intel and perhaps release a 10C/20T or 12C/24T main stream cpu OR brings Threadripper down in price, Intel may decide to move up their release.

A lot depends on how February goes for AMD and what NVIDIA's Volta based gaming cards look like.

[Rev Jim Jones]

*



*****WARNING - WALL of TEXT AHEAD*****


Again - more great info.

AVX512 - so are you saying that x265 and x264 are already coded to take advantage of the AVX512 instruction set?

This would suggest yes ..."Our initial results with the latest build of x265 show a 67% average per-core gain for encoding using HEVC Main profile, and a 50% average gain with Main10 profile across different presets."

Yet this, the way it is worded implies its a work in progress now ...."As we enhance x265 to take advantage of the new technologies that these new processors bring to the light, including AVX-512, we expect that users of x265 will love the benefits that they see with these new Xeons."

Perhaps the AVX512 instruction set coding has been done for x264 and x265 and is in a test build and not a general release yet? Or is it already a main stream thing and programs like Staxrip are already using it?

Then the next VERY interesting part....."This even extends to the Core i9 (Skylake-X) consumer processor family, which are based on the same Purley architecture."

This says that Skylake X has the same under laying arcitecture as the Xeons - the Skylake X family is essentially a Xeon.

So it is absolutely worth buying into Skylake X for encoding. By encoding, I mean re-encoding BD25/50's down to small files for other devices usage.

So I have been doing a deep dive into which CPU to get into for this new build. I want to stay M-iTX. I would like to use a small case - an N1 or a DAN A4 or possibly the new Ghost S1 MkII - all SFF. I have looked at the S4 mini but it seems TOO small. Reason for the small case is so I can move the system easily between my home in town and my house by the lake when I go up there. Also nice to have it sitting on my desk rather then on the floor in all the dust that builds up up down there.

So with that back ground out of the way......been looking at CPU's. Cant argue against the value of Ryzen. But since a TR is out of the question for me due to size, it would be an R7 1700 with an OC on it. But that brings heat issues in a small case. However after extensive SFF forum reading it is being done.

So R7 1700 - can OC to around 3.8GHz on all 8 cores. Needs 32GB of fast RAM because of Ryzens infinity fabric - OC the RAM and the CPU goes faster. No onboard APU, which means a graphics card has to go in the single x16 Pcie slot on the mother board. Heat build up not too bad and lower overall wattage's on that OC. All in all - seems doable in a SFF case on air cooling with the right fan setup (there are a myriad of fans out there for this)

M-iTX boards for R7 have been sparse. The best itx main board seems to be the Asus ROG Strix X370 - it has 2 x M.2 slots and 4 Sata III connections onboard.

The other way to go is x299 and Skylake - just expensive to do it this way. One advantage is the ASROCK x299 mITX board - its way HOT! 64GB Quad Channel RAM, 3 x M.2, USB 3.1 Gen 2 and supports bifurcation on the x16 PCIE slot - if one needs such a split on that x16 slot.

But which Skylake X CPU? Seems the 12 core 7920x is the sweet spot. 12 cores with a 2.9GHz base clock and 140wTDP with 40 PCIe lanes wired to the CPU - price has come down to around $1080 for that CPU.

Now thanks to your heads up - it also has AVX512 support - it was always there, I just didn't realize the importance of it and the implications to encoding.

For a case for the x299 board/Skylake CPU I would need water cooling and a radiator - the Ghost S1 allows for this (its on kickstarter NOW - ends in 12 days time)

So all things considered it looks to me the x299 ASROCK mITX and a Skylake x CPU is the way to go.

Are there any implications to HDD vs SSD for encoding? In a mini case with the ASROCK x299 board it has 3 x M.2 slots to allow all SSD drive infrastructure - how important/unimportant is that?

If i go the ASROCK x299 board (most likely) I plan to use 1 of the M.2 slots to connect a 10Gbps Pcie card via an M.2 to Pcie adapter - so I can move big BD50 files from the NAS to the RIG fast for encoding.

The NAS idea I have is to get a QNAP TVS 1282T3 that has TB3 and 10Gbe onboard as the repository for the BD50's - so I move them on and off the QNAP to the m-itx setup for encoding then move the BD50 back from the m-itx to the NAS for its longer term storage once I have an encoded version and move to the encoded files onto the QNAP as well for further use as needed.

So with all that, its complex to figure this all out.

[Rev Jim Jones]

So, digging into the AVX512 instruction set.

Found this ;

Important changes available in Xeon Scalable Processor Family “Skylake-SP” CPUs include:
Up to 28 processor cores per socket (with options for 4-, 6-, 8-, 10-, 12-, 14-, 16-, 18-, 20-, 24-, and 26-cores)
Floating Point and Integer Instruction performance improvements:
New AVX-512 instructions double performance
(up to 16 double-precision FLOPS per cycle per AVX-512 FMA unit)
Up to two AVX-512 FMA units per CPU core (depends upon CPU SKU)

A new server platform (formerly codenamed “Purley”) to support this new CPU product family

Intel Xeon Gold – Recommended for most HPC workloads
The best balance of performance and price. In particular, the 6100-series models should be preferred over the 5100-series models, because they have twice the number of AVX-512 units


Re Skylake X SKU's.

This SKU shows that the CPU has the AVX 512 extensions

https://ark.intel.com/products/126699/Intel-Core-i9-7980XE-Extreme-Edition-Processor-2...up-to-4_20-GHz

As does this one

https://ark.intel.com/products/126695/Intel-Core-i9-7940X-X-series-Processor-19_25M-Ca...up-to-4_30-GHz

As does this one

https://ark.intel.com/products/126240/Intel-Core-i9-7920X-X-series-Processor-16_50M-Ca...up-to-4_30-GHz

As does this one

https://ark.intel.com/products/123613/Intel-Core-i9-7900X-X-series-Processor-13_75M-Ca...up-to-4_30-GHz

As does this one

https://ark.intel.com/products/123767/Intel-Core-i7-7820X-X-series-Processor-11M-Cache...up-to-4_30-GHz


what they dont say on those product sheets is how many FMA units the CPU's have and thats the crux of it. The full fledged Xeons have 2 FMA units

*******************
Edit - found this on Extremetech

The Core i9-7900X (10-core) and above, including the Core i9-7980XE, have two 512-bit AVX-512 ports, while the 8-core and six-core parts have a single port for FMA-512. This means the higher end CPUs can support much higher throughput (64 single-precision or 32 double-precision operations per cycle, compared with 32 SP/16 DP operations on the 7800X and 7820X).

*******************

Cannon Lake will be the first consumer CPU's with the AVX512 instruction extensions....

"Cannon Lake CPUs are the first consumer CPUs to include the AVX-512 instruction set."

Seems as though Coffee Lake CPU's do not have the AVX512 extensions

"Before Cannon Lake's launch, Intel launched another 14 nm process refinement with the codename Coffee Lake"


So yeah.....if you wait for the CannonLake refresh you can get the AVX512 instruction extensions on a consumer CPU. Looks like right now the only option to get the AVX512 instructions is to go with Xeon Gold as Sophisticles mentions or go with Skylake X CPU's.

Brilliant heads up Sophisticles...thanks for this.

The deep dive gets deeper.........CES is just around the corner. I think it may be prudent to hold till then at least. *sigh* ...goes back to waiting.

[sophisticles]

The only problem with the Intel cpu's is that they are so expensive, the 7920X you mentioned is $1000 for a 12C/24T processor where as the 12C/24T Threadripper costs just $500 and a 16C/32T Threadripper 1950X has dropped down to $700:

http://www.microcenter.com/product/483132/Ryzen_Threadripper_1950X_34_GHz_16_Core_TR4_...oxed_Processor

For $300 less you get 4 more cores that support 8 extra threads compared to the Intel cpu. On the other hand, it's been proven that as thread count increases the quality of an encode, as done via x264 and x265, decreases, so in theory using a 1950X will give lower quality encodes than using a 7920X or a 1920X; of course, thanks to the higher thread count you could compensate by using more aggressive encoding settings.

I think the wait and see approach is probably the wisest.

[Rev Jim Jones]

So the TL: DR version is;

Ryzen 7 1700 on Asus ROG x370 with 32GB Ram with 1 x M.2 SSD and 1 x Sata3 SSD (or HDD) to save money - but gets the job done - just a little slower.
Skylake X 7900x and up on the ASROCK X299 with 64GB RAM and 3 x M.2 SSD- leaves the wallet smoking but gets the job done fastest and can go all the way to 18 cores on M-iTX if needed or you are wealthy.

Not looked too deeply into the Coffee Lake side of things since its only 6 cores. Baseline in my considerations are 8 cores and up.



Yes agree TR is a much better value proposition - but I cant fit a TR on a M-iTX board. If i could I would most likely go for TR.

Need the Small form factor for my situation. The only 3 options so far are Ryzen 7 1700 on the Asus ROG x370 board or Skylake X on x299 ASROCK M-iTX board or the 8700K Coffee Lake 6 core CPU.

- issues on the approaches are;

Ryzen 7 and Asus ROG x370 board ;

- only 32GB RAM thats Dual channel, only 2 x M.2 slots - one of which i need to connect an M.2 to Pcie adapter to so I can hook up a 10Gbe card so I can get the big files back and forth to the rig from the NAS quickly, leaving only 1 M.2 SSD for doing the encoding on which means I then have to connect a Sata3 Drive (either SSD or HDD) so that OS and source files are one drive and be able to write encodes to the other drive. Not enough RAM for a RAM drive (a big enough one anyway)

Skylake X and ASROCK x299 board ;

- 64GB Quad channel RAM, 3 x M.2 slots one of which i need to connect an M.2 to Pcie adapter to so I can hook up a 10Gbe card so I can get the big files back and forth to the rig from the NAS quickly, leaving 2 x M.2 SSD - one for OS, one M.2 to write BD50 files from the NAS too and can use the huge RAM capacity to make a RAM drive to write the encodes out too then when finished send back over the 10Gbe wire to the NAS storage. By writing the encodes to the RAM drive it saves the wear and tear of writing constant small chunks of data to the M.2 SSD.

and then the 900lb Gorilla in the room - Skylake X is essentially a Xeon in sheeps clothing - 10 core Skylake X and up gets 2 x FMA units per core for AVX 512 extensions.

So at the end of the day, money aside, the X299 ASROCK board is far more versatile, Skylake X is the only way to get more then 8 cores onto the M-iTX form factor and the only way to get AVX 512 extensions using a M-iTX form factor. The ASROCK x299 M-iTX board is the only Small form factor (SFF)board that has 64GB RAM much less Quad channel.

It really looks like those AVX512 instructions set extensions are important and will only get more so as x265 becomes more mainstream as it will over time (as 4K becomes more a of a staple thing in our life x265 HAS to become the go too for encoding...eventually)

[Rev Jim Jones]

hahahaha

So reading through threads here I came across this on the X265 thread...... it discusses the AVX512 thing. Seems its not implemented at all yet given the last response on the 15th of December. It pays to read and read before posting.

19th Nov 2017 09:24 #752
vertigo220 vertigo220 is offline
Member
Nov 2016
Search PM
Does anyone know if AVX-512 will be added anytime in the near future (next several months or couple years)? I'm getting ready to build a new computer and trying to decide between Intel and AMD, and one of the factors I'm looking at is that Intel has AVX-512 instructions and AMD does not. If it's going to be a couple years at least before it's added to x265, it probably won't matter much, but if there's a good chance of it being implemented in the next year or so, it would be a good reason to go with Intel.



15th Dec 2017 12:32 #753
LigH.de LigH.de is offline
Member
Aug 2013
Central Germany
Search PM
The sources have been changed to use NASM instead of YASM (which seems to be more or less obsolete, feature-wise; altering YASM to make it support AVX-512 may be too much efforts, their developers stated). A first step is done...

[KarMa]

There's the other thing to think about. The new "AV1" standard coming from the Alliance for Open Media (consisting of companies like Google, Netflix, Mozilla, AMD, Nvidia, Intel, Facebook, and so on). It seems to be nearing standardization where the decoder requirements are set. (within the month probably). Firefox even has a AV1 decoder already built in. But available encoders for AV1 will be limited to probably one for awhile, provided for free by the group I assume, with very unoptimized code. Which will be slowly optimized over years. So I'm not sure how AVX-512 will play into this.

Handbrake scales out to 6 cores - right? So is a 6 core over clocked to 5GHz better then 8 cores at say around 3.8 to 3.9 GHz ?

x264 can pretty much use all the cores you throw at it (forget the actual limit), and tend to believe x265 is the same way. Handbrake uses both x264 and x265. There are efficiency losses as you add cores but even a 16 core encoding would result in negligible efficiency losses for 1080p/4k content.

https://forum.doom9.org/showthread.php?p=1213185#post1213185

You can always just run two encodings at once, with half of the cores doing one job and the other half doing the other.

[Rev Jim Jones]

Thanks Karma. Not heard of AV1. More reading ahead.

[sophisticles]

hahahaha

So reading through threads here I came across this on the X265 thread...... it discusses the AVX512 thing. Seems its not implemented at all yet given the last response on the 15th of December. It pays to read and read before posting.

19th Nov 2017 09:24 #752
vertigo220 vertigo220 is offline
Member
Nov 2016
Search PM
Does anyone know if AVX-512 will be added anytime in the near future (next several months or couple years)? I'm getting ready to build a new computer and trying to decide between Intel and AMD, and one of the factors I'm looking at is that Intel has AVX-512 instructions and AMD does not. If it's going to be a couple years at least before it's added to x265, it probably won't matter much, but if there's a good chance of it being implemented in the next year or so, it would be a good reason to go with Intel.



15th Dec 2017 12:32 #753
LigH.de LigH.de is offline
Member
Aug 2013
Central Germany
Search PM
The sources have been changed to use NASM instead of YASM (which seems to be more or less obsolete, feature-wise; altering YASM to make it support AVX-512 may be too much efforts, their developers stated). A first step is done...

The link I provided straight from x265.org makes it pretty clear that AVX-512 is supported:

Our initial results with the latest build of x265 show a 67% average per-core gain for encoding using HEVC Main profile, and a 50% average gain with Main10 profile across different presets.

This was written back in July 11 and it clearly says that those results are with the latest build of x265, I fail to see how this can be misinterpreted.

Then it says:

Give them a spin, and let us know what you think!

How would they expect you to try it if it wasn't already implemented?

[Rev Jim Jones]

Well, its hard to know what is what from a few posts.

What I am wondering and asked but got no reply on, was if AVX512 was implemented in an x264/265 stable release that is widely available and being used in the various programs or if it was in a beta or something and was still being tested.

If AVX512 is indeed implemented in a full general release version thats out there already then it reinforces that the Skylake X is the way to go right now - high performance always costs money - no matter if its a fast CPU, a fast car or a beautiful blonde young mistress that is high performance - performance costs.

What I am debating is, Skylake X, Ryzen 7 or Coffee Lake 8700K. They all have pro's and con's. The only real con I can think of with the Skylake X and X299 approach is money.

[Rev Jim Jones]

......so i ordered the Ghost S1 SFF case on kickstarter.

Case gets delivered in April, so will wait out till then for the CPU/itx board.

Ryzen 2 is rumored to be coming. Lets see what that brings.

[stonesfan99]

If money is no issue, Threadripper 1950X. If money is an issue, Ryzen 1700X or 1800X. I went with the 1800X in my rig because it was less money for more cores/threads than the i7-8700k which had just come out at the time.

[usually_quiet]

......so i ordered the Ghost S1 SFF case on kickstarter.

Case gets delivered in April, so will wait out till then for the CPU/itx board.

Ryzen 2 is rumored to be coming. Lets see what that brings.

I just looked at that case. It seems well designed. It is smart to wait. When selecting parts, you will need to pay attention to any restrictions for component dimensions and TDPs listed in the instructions.