Ryzen 3000 & Navi Megathread | Level One Techs

AMD has a hard cap of $50. I was expecting to see trading open this morning at $32 or $33.

I’d not be surprised to see $35-38 though.

That’s definitely true.

1 Like

Oooh. I didn’t know that.

I mean, it’s not a rule, but just a trend.

Most people don’t believe they can pass $50

1 Like

Wanted to bring attention to this article about memory scaling on Zen2. Turns out what you really want is DDR4-3600 with tight timings. Can save a bunch of money there.

relative-performance-cpu

@wendell said something like that in the video too. IIRC the memory clock is coupled to the infinity fabric speed up to 3600 and decoupled at anything higher. That decoupling introduces latency which makes higher speed memory effectively pointless. At least that was my very basic understanding.

Speaking of that: Where was that Ryzen memory speed calculator again?

This?

1 Like

YES! Thanks. :grinning:

1 Like

I am really interested in the Navi HEVC decode / h.264 encode performance.

As a Plex user, we are used to Nvidia cards for HW acceleration transcoding for those devices that can’t play HEVC. Nvidia cards like the Quadro P2000 or GTX 1060 (once delimited) can do around 18 - 20 simultaneous real time streams of 10mbps 1080p HEVC 8 bit to (10mbps I think) 1080p h.264. Those are mainly limited by their low VRAM. Would be nice if AMD could bring some competition to the overpriced Quadros.

@wendell Any chance you could test something like this with ffmpeg once it gets fixed? (since Plex uses ffmpeg)

I think EposVox is working on this exact thing. Some initial info about that is in the megathread on his channel, did you see that? up to 6 streams before it starts to slow down

2 Likes

he’s doing something…

3 Likes

Yeah saw that and directly came over here. ffmpeg + navi seem to have teething issues that is why I suggest a revisit of this some time in the future.

5 streams is what the RX 580 could do while Vega 64 could do ~15. That all was on Win10 since Plex hadn’t enabled AMD acceleration on Linux due to old ffmpeg build used.

Seems like EposVox was using 8mbps source videos? Also, he is doing the opposite, transcoding h.264 to HEVC.

1 Like

because we are struggling a bit on the windows site with this, and it is new territory for me, can you document a bit Vega64 ~15 streams? is that a test you are doing? is there a script or some such we can use to test?

These would need couple timings from one speed to see where it goes like that, I dont know how that concludes you need 3600 with …tight timings?

Seems hard to find information on how the RX 5700 and RX 5700XT stacks against the RTX 2060/S and 2070/S since some sources place the RX 5700XT close to the RTX 2070S but some place it significantly lower…granted it’s $100 cheaper.

I wish I could but those were not my tests. Tests were done through Plex, not bare ffmpeg.

Here is the reddit thread relevant to Vega performance: https://www.reddit.com/r/PleX/comments/9q2lw7/amd_gpu_transcoding/

And slothtechtv’s contribution to the AMD Plex tests: https://www.youtube.com/watch?v=aXt06PgEOAU&feature=youtu.be

Basically fire up Plex and start some transcode playbacks.

One way of calculating maximum theoretical simultaneous streams is calculating decoding FPS and encoding FPS, taking the lowest one of those FPS numbers, then measuring VRAM usage per transcode and finding out if the limiting factor is the dec/enc process or the VRAM. Really quite convoluted and theoretical. I don’t know how you could measure FPS enc/dec performance but Nvidia has done it:

Page 9 for NVENC:
https://developer.download.nvidia.com/designworks/video-codec-sdk/secure/7.1/01/NVDEC_DA-06209-001_v08.pdf?-0ei-lpY62Nz3Sqptll9QW-_Nnm7TzLRVyKDzTdyldvH7wI8ETPFsjnCwzuuDHlxk5-jCtIgacO-uCGlLuicHK3gvnC2H9vHASDGG3d5knjC11JCOqXZuUUffRvbMrvXdR-JvbI2seJn5wupRwgyv2gkojrzqx7_klrEBIxV-l06MmM

Page 6 for NVDEC:

Here are some estimates for Nvidia cards based on Nvidia’s documentation quoted above, testing and some guesswork:

https://www.elpamsoft.com/?p=Plex-Hardware-Transcoding
Maybe their methodology might be of help.

Plex on Windows supports AMD HW accel because of the Microsoft Media Foundation codecs. I think Linux AMD HW accel is still not supported and NVDEC is also not officially supported yet but there is a patch.

That’s a lot easier said than done. It’s also not really worth it for them. Even when they had better GPUs for less money no one bought them anyway. They had no cash for a long time and burned almost all of it on Zen. It is gonna take time. 2020-2021 likely

Amd stock price is massively manipulated. People are actively trying to keep it down. Not to mention there is a lot more than consumer parts driving share price. Desktop is small potatoes. Also the whole market is down today too anyway

It is still over 40 bucks since i bought some at 20, so i’d say it’s doing okay

$32…

1 Like

Those will also not be overclockable. I think they’ll only allow overclocking on XT SKUs.

Intel clocks are based on non-AVX workloads. It’s very interesting that Zen 2 doesn’t require an offset to remain stable… Very handy for people using FFmpeg.

My 4960x isn’t stable at 4.4Ghz for FFmpeg on Linux, but is stable with a manual offset to 4.3… Glad to hear Zen 2 requires no offset work. But I will be fully convinced when FFmpeg can remain stable on Linux with a PBO offset.