AMD Ryzen Gen 2 Set For Q2 2018

[The Stilt]

One can say x265 is the important thing today and h264 less so. Time works for skl here.

But take a TR 1950 at 900usd and compare it to a similar priced skl x on x265 encoding and whats the difference then?
Straight of the head i would guess most of the advantage of those wide vectors is then gone due to price disadvantage.

I actually tested the IPC on Ryzen, Coffee Lake and Skylake-X in X264 and X265 few days ago, using the most recent binaries.

X264: 5.55fps (Ryzen), 6.14fps (CFL), 6.42fps (SKLX)
X265: 2.47fps (Ryzen), 3.40fps (CFL), 3.50fps (SKLX)

0.156.2905 build for X264 and 2.7+2 for X265.

Medium-preset, 1080P, Main (8-bit) and CRF 17.0 for both encoders.
X264 has some minor AVX-512 implementations in it at this point.

 

[The Stilt]

Well I've not checked lately, but when TR was released, it perfomed pretty well in video encoding tasks. SKL-X was only faster with x265 codec if I remember correctly. About Blender it peformed pretty well too, dont know about music production though

Both of the most common video encoders (X264 and X265) are NUMA aware.
Normal Blender rendering generally doesn't involve the physics. Blender ships with Bullet physics engine and it is extremely latency sensitive.

 

[Markfw]

I actually tested the IPC on Ryzen, Coffee Lake and Skylake-X in X264 and X265 few days ago, using the most recent binaries.

X264: 5.55fps (Ryzen), 6.14fps (CFL), 6.42fps (SKLX)
X265: 2.47fps (Ryzen), 3.40fps (CFL), 3.50fps (SKLX)

0.156.2905 build for X264 and 2.7+2 for X265.

Medium-preset, 1080P, Main (8-bit) and CRF 17.0 for both encoders.
X264 has some minor AVX-512 implementations in it at this point.

When you say Ryzen, is that 1800x/1700x or Threadripper ? And which skylake and coffeelake CPU's were used ?

 

[The Stilt]

When you say Ryzen, is that 1800x/1700x or Threadripper ? And which skylake and coffeelake CPU's were used ?

Ryzen DT SKU.
SKUs don't really matter for IPC measurements, but R7 1700, 8700K and 7960X were used.
All clocked to 3.8GHz.

 

[HurleyBird]

X264: 5.55fps (Ryzen), 6.14fps (CFL), 6.42fps (SKLX)
X265: 2.47fps (Ryzen), 3.40fps (CFL), 3.50fps (SKLX)

Memory speeds?

 

[The Stilt]

Memory speeds?

2666MHz for all (14-14-14-32).

 

[Markfw]

Memory speeds?

Yes, Ryzen needs at least 3200cl14 to work well, not sure on Intel, but their reasonable preferred speed should be used.

 

[MarkPost]

@The Stilt: We know Intel CPUs are faster clock to clock, thread to thread. Thats the reason AMD offers more cores at a similar price, to compensate that (and clock freq) disadvantages.

EDIT: for example, my 1700X is faster than my 7800X with x264, and slower with x265

 

[The Stilt]

Yes, Ryzen needs at least 3200cl14 to work well, not sure on Intel, but their reasonable preferred speed should be used.

Makes no difference in video encoding, or in single threaded workloads (like the IPC test) thou.

 

[Markfw]

Makes no difference in video encoding, or in single threaded workloads (like the IPC test) thou.

In every test I have seen it sure does, it reduces latency a LOT

 

[The Stilt]

In every test I have seen it sure does, it reduces latency a LOT

Generally it has a huge effect, but it makes no difference in video encoding.

 

[yummycandy]

Does anyone of you have a clue, when the NDA ends?

 

[epsilon84]

Yes, Ryzen needs at least 3200cl14 to work well, not sure on Intel, but their reasonable preferred speed should be used.

At least 3200 CL14? From what I've seen that is the optimal standard, rather than the 'minimum'. Many current AM4 motherboards struggle to run RAM faster than 3200 CL14 (some *may* get 3466 CL14 but that is the exception, not the norm) Hopefully Ryzen 2000 chips will have less issues with faster memory.

If you are going to attempt to not disadvantage AMD and run CL14 3200, then in terms of fairness you will also need to run DDR4 4000+ for Intel as that is the respective max RAM frequency for both platforms.

 

[The Stilt]

my 1700X is faster than my 7800X with x264, and slower with x265

This is expected.
R7 1700X should be around 9% faster when both are running at stock.

 

[Topweasel]

Behind yes, but if it means AMD have another unit they can charge a higher price for then it makes financial sense for them.

[There is a world beyond the epeen contest...]

But the point was/is. If they had a faster CPU that they could charge more for, it would make more sense to offer that when they could charge the highest premium. That would be before the mythical 8c coffee lake.

 

[krumme]

I actually tested the IPC on Ryzen, Coffee Lake and Skylake-X in X264 and X265 few days ago, using the most recent binaries.

X264: 5.55fps (Ryzen), 6.14fps (CFL), 6.42fps (SKLX)
X265: 2.47fps (Ryzen), 3.40fps (CFL), 3.50fps (SKLX)

0.156.2905 build for X264 and 2.7+2 for X265.

Medium-preset, 1080P, Main (8-bit) and CRF 17.0 for both encoders.
X264 has some minor AVX-512 implementations in it at this point.

Okey so with the new binaries a 1950x would edge out a similar priced 7900x for x265 encoding as i can tell?

Do we have any efficiency comparisons for x265 load? Or similar wide vector loads? I think thats the interesting part because it can give us hints to when amd adapts 256 wide fpu.

 

[Atari2600]

But the point was/is. If they had a faster CPU that they could charge more for, it would make more sense to offer that when they could charge the highest premium. That would be before the mythical 8c coffee lake.

Yes... and...?

I said if they see that a minor respin down the line would give them enough additional performance that they have a CPU they can charge more for at that point then they are making room for it now in the lineup.

Obviously in an ideal world they'd have that performance now. But its not always an ideal world.

 

[Topweasel]

But that wasn't the point I brought that up earlier that they could be saving it for if they can push the process a little bit or build a stock of cherry dies. But the point some people were pushing for was that it would be an ace for AMD to pull out after the 8c cfl. If AMD had it, they would release it now to maximize ASP.

 

[FlanK3r]

Both of the most common video encoders (X264 and X265) are NUMA aware.
Normal Blender rendering generally doesn't involve the physics. Blender ships with Bullet physics engine and it is extremely latency sensitive.

Do you think, it is possible add 256-bit vectros in new Zen core (2019/2020) ?

 

[The Stilt]

Okey so with the new binaries a 1950x would edge out a similar priced 7900x for x265 encoding as i can tell?

Do we have any efficiency comparisons for x265 load? Or similar wide vector loads? I think thats the interesting part because it can give us hints to when amd adapts 256 wide fpu.

Most likely only at 4K resolutions.
At 1080P X265 cannot use more than 10-12 cores efficiently due to it larger default CTU size (64), despite the theoretical limit is 16.875 cores.
With X264 the default CTU is half the size, meaning it can use more cores efficiently.

Currently X265 has no AVX512 code in it, despite it is certainly a work in progress (assembler was replaced with NASM a while ago).
X264 contains a very modest amount of code in AVX512 and it provides < 5% performance improvement over AVX2. VP9 supports AVX512 as well, however the performance of the encoder appears to be so poor (for multiple reasons) that
I don't think it is relevant how much or little AVX512 improves the performance.

Do you think, it is possible add 256-bit vectros in new Zen core (2019/2020) ?

Personally I don't believe that Zen2 will increase the width of the execution resources.
I guess it will ultimately depend on the strategy: Is it possible (or viable) this time around to make separate server and desktop designs, and if it isn't does the server design need to have wider resources or not.

 

[dnavas]

Personally I don't believe that Zen2 will increase the width of the execution resources.
I guess it will ultimately depend on the strategy: Is it possible (or viable) this time around to make separate server and desktop designs, and if it isn't does the server design need to have wider resources or not.

I hope that's not the case! As you note, there are definitely benefits to it being there. I was kind of hoping for a half-wide AVX512 unit. :sigh:

 

[EXCellR8]

Looks like I'll be skipping Zen+ for either a C. Peak HEDT or Matisse build next year... much too early to retire the 1000 series chips as the initial build was a labor of love and both machines are dialed in so nicely now. I also want to see where we are with memory in a year or two and upgrade the whole platform if necessary. I typically don't buy into the refresh or optimization generations of CPUs.

 

[krumme]

Anyone wondered if the change of ceo at GF perhaps is connected to lacking 7nm execution?

 

[scannall]

Anyone wondered if the change of ceo at GF perhaps is connected to lacking 7nm execution?

The new CEO there is an old IBM guy, and it would make sense since they now have so many IBM engineers and IP.

 

[krumme]

The new CEO there is an old IBM guy, and it would make sense since they now have so many IBM engineers and IP.

Okey makes good sense then. Also connection to ibm and knowledge of it is vital forward.