Blackmagic Forum

Hi.

AMD Ryzen Threadripper 2990WX 32-Core / 64 Threads are on preorder at Newegg or Amazon for $1,799.99. Release Date: 8/13/2018. Limit 1 per customer:

https://www.newegg.com/Product/Product. ... 6819113541

And Amazon.com:

https://www.amazon.com/dp/B07G25SD1P?tag=anandtech01-20

AMD normally makes all new CPU's available worldwide on the launch day through its distributor network.


Expect some more Refresh X399 Motherboards to be announced the coming days.

Here is a link to a 16 min long YouTube interview with AMD's Jim Anderson where he show a new AIR Cooler for Threadripper 2. It also fit the old Threadripper 1:



Here is more information on it. It must be purchased separately:

https://www.anandtech.com/show/12996/am ... megacooler

The new Threadripper 2990WX with 32-Core / 64 Threads have a TDP of 250 watt. If you already have a ASUS Threadripper generation 1 motherboard designed for a TDP of 180 Watt can you soon buy a cooling enhancement kit for Threadripper Generation 2, which holds a basic expansion heat sink, and a fan. Here is a link:

https://www.guru3d.com/news-story/asus- ... -gen2.html

But you will have to update you motherboards BIOS first.

Regards Carsten.

It will be interesting to see how well they do vs similarly priced Intel chips.

I'm personally holding off till next year when gen 3 comes out on the 7nm architecture.

Dan Sherman wrote: It will be interesting to see how well they do vs similarly priced Intel chips.

Hi.

Here is a Rendering Benchmark: Cinebench 15 MultiThreaded together with prices for both Intel and AMD previous CPU's:

https://www.anandtech.com/show/13123/am ... 32-cores/5

I am sure more will follow in the coming days.

Regards Carsten.

Dan Sherman wrote:It will be interesting to see how well they do vs similarly priced Intel chips.

Intel has a small edge in single threaded performance at a given clock speed and a small clock speed advantage, but at a comparable price you're looking at 18 vs 32 cores, both with SMT.

So for now AMD is quite far ahead at $1800...

Carsten Sellberg wrote:

Dan Sherman wrote: It will be interesting to see how well they do vs similarly priced Intel chips.

Hi.

Here is a Rendering Benchmark: Cinebench 15 MultiThreaded together with prices for both Intel and AMD previous CPU's:

https://www.anandtech.com/show/13123/am ... 32-cores/5

I am sure more will follow in the coming days.

Regards Carsten.

Cinebench scales almost linearly with core count though. The frequency vs. core count trade off is a lot more complex in an application like resolve and is even dependent on the source material.

Dan Sherman wrote:Cinebench scales almost linearly with core count though. The frequency vs. core count trade off is a lot more complex in an application like resolve and is even dependent on the source material.

Indeed. It'd be fun to have one, and I'd probably see improvement for transcoding/exporting, but none of the source material I have would benefit at all for actual work. If I had 6K/8K RED, then I'd be chomping at the bit.

There is a Video with the 2990WX that includes a Resolve benchmark but it seems to not perform that well.
https://www dot youtube.com/watch?v=xZpN_wnHlz4

Thanks Andreas.



The DaVinci Resolve 15 part is from 3:07 to 3:47.

As I hear the soundtrack: ' a simple 4K/60 video converted to a QuickTime YouTube upload using default settings. For good measure, anti-camera shake is applied'

Regards Carsten.

Carsten Sellberg wrote:Thanks Andreas.

The DaVinci Resolve 15 part is from 3:07 to 3:47.

As I hear the soundtrack: ' a simple 4K/60 video converted to a QuickTime YouTube upload using default settings. For good measure, anti-camera shake is applied'

Regards Carsten.

It was still the fastest, but I'm betting it wasn't fully leveraging all the processing power available to it.

Love these reviews were people use apps which they have no clue about.
Then they got results which again don't have clue how to explain
Above video is a very good example.

Indeed, Resolve has parts of it's pipeline that aren't multithreaded that well or are using GPU exclusively. You'd need to know how the software works (and BMD isn't forward with that information) to know how to benchmark with it.

Threadripper/EPYC are more interesting in their PCIe lane count. That's more NVMe and GPU capacity, which will make a huge difference in performance. When you can have 24TB of your local cache running at >40GB/s, that's where these systems shine.

It appears that if you're using highly threaded applications which are not constrained by memory bandwidth then the 2990WX performs pretty well. However, on a per-core basis, the 2990WX is severely constrained in its memory access because half of the cores have no direct access to system RAM.

This video does a good job of explaining the situation and showing how the 2950WX actually outperforms the 2990WX on quite a number of applications due to this memory bandwidth issue.

Andrew Kolakowski wrote:Love these reviews were people use apps which they have no clue about.
Then they got results which again don't have clue how to explain
Above video is a very good example.

And he pronounces it Davinsee. *facepalm*

RCModelReviews wrote:This video does a good job of explaining the situation and showing how the 2950WX actually outperforms the 2990WX on quite a number of applications due to this memory bandwidth issue.

I've watched a *lot* of TR2 videos yesterday and this one from Hardware Unboxed is definitely above average. You could tell a lot of the articles and videos were very rushed to get content out for the embargo release but the quality is quite dubious.

I'm sure we can expect much better content out in the coming weeks.

I am sure Matt Bach from Puget Systems is working on an article for Resolve now. I'm not interested in h264 export times, I want to see 8K decode performance!

Yep, h264 exports which use poor encoders or GPU based ones. Great way of CPU performance measure

Jason Tackaberry wrote:

RCModelReviews wrote:This video does a good job of explaining the situation and showing how the 2950WX actually outperforms the 2990WX on quite a number of applications due to this memory bandwidth issue.

I've watched a *lot* of TR2 videos yesterday and this one from Hardware Unboxed is definitely above average. You could tell a lot of the articles and videos were very rushed to get content out for the embargo release but the quality is quite dubious.

I'm sure we can expect much better content out in the coming weeks.

We are leaving in era of yotubers- there are suddenly 1000s of experts in every area. Sad times of 1mln videos out there with just 100 of them being decent.
Some videos are useful but every one needs a good filtering

Hi.

From the middle of the page in this link:

https://hexus.net/tech/reviews/cpu/1206 ... wx/?page=3

'AMD may argue that HandBrake is the wrong sort of test for a 32-core processor. Instead, the company says, users of professional apps such as DaVinci Resolve will see a clear benefit from 16 cores to 32. We're currently evaluating the benchmark.'

Can anybody confirm that the Threadripper 2990wx scale well in Resolve?

Regards Carsten.

Decoding of RED should scale- there were graphs showing 28 (if I remember well) cores running at close to 100%. Decoding ProRes, DNxHR etc should also scale well, but you won't even need so much processing power even for 8K.
32 is a lot of cores and some processes simply won't scale anymore. For example x264 "good" scaling saturates around 16 cores (for HD). After this you have to many small pieces (which frame is divided to) and also quality starts to suffer. You better of splitting encode into 2 (specially when it's long one) and running 2 instances. At these cores numbers you are far from linear scaling though (against 1 core).
As I always repeat- I would always take 24 cores at 5GHz over 32 cores at 3.8GHz.

Hi.

Many reviewers saw strange benchmark result under the test of the new Threadripper 2990WX. The current thinking alludes this to lower memory bandwidth on the Threadripper 2990WX, and only few blamed Windows for it. They think Windows 10 Pro is unable to recognize the strange NUMA-layout (2 nodes with memory and 2 nodes without) of the Threadripper 2990WX. So they run some Windows 10 Pro to Linux comparisons.

Here is first a link with an explanation of NUMA Non-uniform memory access – Understand it & its usefulness with Windows Server 2012:

https://blogs.technet.microsoft.com/pra ... rver-2012/

And here a link to The Windows 10 vs. Linux Performance On AMD Threadripper 2990WX:

https://www.phoronix.com/scan.php?page= ... dows&num=1

As you can see is the Linux performance much higher than the Windows 10 Pro performance.

So now, we must all wait for Microsoft. And according to AMD, Microsoft has not committed to a timeline for updating its scheduler.

Regards Carsten.

Yes, tests on Linux show sometimes huge performance gain, eg .30% (for the same app).