Inspiring

Inconsistent export times (Mercury Playback CUDA slower than Software Only)

Forum|Forum|1 year ago
November 25, 2024
11 replies
3643 views

UPDATE - This thread was originally posted in the discussions section and then merged with a newer bug report post on the 14th of February 2025.

OVERVIEW (14th Feb '25)

I’m having an issue with very inconsistent export times, particularly when the Mercury Playback Engine is set to CUDA.

What's interesting is that when the Mercury Playback Engine is set to Software Only, my test timeline takes around 3:45-4:30 minutes to export. However, when CUDA is selected, the export can take up to 19 minutes. That said, it can also take 7 minutes (typically right after I restart Premiere).

MORE INFO/ TROUBLESHOOTING STEPS IN THREAD BELOW....

Original post (25th Nov '24)
I’m hoping someone can clear up a bit of confusion I’ve been having around using CUDA vs Software Encoding in Premiere Pro/Media Encoder.

Everything I’ve read says that if you’re on Windows and have an Nvidia GPU, you should always select CUDA under the Render Engine for better performance. But after noticing some surprisingly long export times, I decided to do a bit of testing, and the results have been completely the opposite.

In my tests, using Software Only for encoding was literally 10x faster than CUDA. I tested this on the same sequence, exporting both through Premiere Pro directly and through Media Encoder. When set to Software Only, the export would finish in about 1 minute. With CUDA, the same sequence would take over 10 minutes. This pattern held across multiple tests, including longer sequences, as well as sequences with a mix of codecs, effects, After Effects comps etc...

No matter the sequence, Software Encoding is consistently coming out way faster than CUDA.

I tested this behavior in both Premiere/Media Encoder 24 and the 25 Beta. Still, the results stayed the same.

It's also worth noting, that when I'm set to Software Only, during the export my CPU is seeing 100% utilisation, but my GPU is also getting 20-45% utilisation.

So, is there a bug with CUDA or hardware acceleration I might have missed? Have things progressed to a point where Software Encoding is now the better option? If the answer is that Software Encoding is the better option now, that's fine. But, given how consistently CUDA is still recommended, I'm just a bit confused.

If anyone can shed some light on this, I’d really appreciate it!

Cheers!

For reference, my specs are:

Adobe Premiere Pro version number: v24.6.3 / v.25.2.0
Operating system - Windows 11 Pro [23H2]
System Info: CPU, GPU, RAM, HD:
- CPU - i9-13900K
- GPU - Nvidia 3070Ti
- GPU driver (PC) - Studio (32.0.15.6590)
- RAM - 96 GB (48GB x2 - Corsair Vengeance DDR5 5600MHz CL40)
- Hard Drives
  - C:Drive = Sabrent NVMe Gen4 SSD (2TB)
  - Media Cache = Crucial P3 Plus Gen4 NVME SSD (500GB)
  - Proxy = Samsung 990 Pro NVMe SSD (2TB)
  - Media = SanDisk Extreme Portable SSD - USB-C (2TB)
Video format:
- Mixed, based on different projects. Normally I'll be editing using transcodes of the rushes, either DNxHD36 or ProRes422 Proxy. However, I will frequently have projects where I pull in h.264 assets.Rushes themselves are typically XAVC based, ProRes, or H.264

J

jermcdee

Participant

Any new info on this? I've had this issue for a while. I'm on an Intel based Imac Pro. Mac OS Sonoma 14.7 Radeon Pro Vega 64 16GB with 32GB RAM.

I try to always export using "Previews" - always to a ProRes codec of some kind. I usually have to switch from GPU to Software to speed up the export of a fully rendered timeline. Now i see in the latest versioin of Premiere they've REMOVED the option of Software Encoding. Is this an issue on newer Macs?

J

jamie_macAuthor

Inspiring

Hi @mayjain7130546

Thanks for the update.

Sadly, having run some initial tests in v25.2, I'm still having the issue. In fact, it seems to be worse now.

Main difference I noticed, was that after switching to Software Only (via startup menu), the render now takes roughly twice as long as in v24.6.4.
The CUDA export (v25 default setting) on the other hand took the longest it ever has.

These are the results of my initial export test, using my 17 minute reference timeline:

Default (CUDA)
10:30 min // 7:10 min // 27:30 min // 17:10 min

Software Only
8:30 min // 8:28 min (vs. 4:30 min in v24)

I need to do some more tests to try and confirm what the new the baseline is, but if this is going to be the default going forward, it sounds like it's going to become a much bigger problem for me!

M

mayjain7130546

Participating Frequently

Hi @jamie_mac and everyone,

Hope you're doing well! Apologies for not yet sharing any conclusive update with you based on the data you shared.

I'm continuing to discuss this with my team, but the latest decision has been to remove software rendering, except for an option to enable it via a launch setting.

This change has been in the beta since January. You can find the announcement and discussion here: Adobe Community Discussion

Thanks for your patience! I’ll be sure to follow up on this.

Best,
Mayjain

J

jamie_macAuthor

Inspiring

Thank you @Kevin-Monahan

And thanks for merging the two threads. Fingers crossed something can be done to figure this out!

Kevin-Monahan

Community Manager

Hi @jamie_mac,

I think we need to call @mayjain7130546 back to the thread to see the results of any testing that was supposed to take place. I hope that they stop by soon. Sorry for the frustration.

Thanks,
Kevin

Kevin Monahan - Sr. Community and Engagement Strategist – Adobe Pro Video and Audio

Kevin-Monahan

Community Manager

Hi @jamie_mac l,

Sure. I hope your post gets more traction here. I appreciate the information and screenshots you shared. Very helpful. I hope a product team member or Adobe Expert will step in shortly. Sorry for the issue.

Thanks,

Kevin

Kevin Monahan - Sr. Community and Engagement Strategist – Adobe Pro Video and Audio

J

jamie_macAuthor

Inspiring

UPDATE - Bug Report merged with original discussion thread

Hi,

I’m having an issue with very inconsistent export times, particularly when the Mercury Playback Engine is set to CUDA.

What's interesting is that when the Mercury Playback Engine is set to Software Only, my test timeline takes around 3:45-4:30 minutes to export. However, when CUDA is selected, the export can take up to 19 minutes. That said, it can also take 7 minutes (typically right after I restart Premiere).

I originally made a post about this in the ‘Discussions’ section, as I wasn’t sure if there was something I was missing, or if the previous wisdom of 'if you have Nvidia, always enable CUDA' had changed. However, from what others have said, as well as the Adobe Support Chat person who remoted into my PC to troubleshoot, this seems to be a bug.

I had some initial interest from some Adobe Staff on my previous post who said they were going to go off and do some testing, but I’ve not heard back for a couple of months.

I’m honestly just a bit sick and tired of being on a deadline and having to play the ‘should I cancel and restart the export’ game.

I would absolutely understand if CUDA wasn't any faster than Software based on whether or not the source/ export codec was able to benefit from hardware acceleration. But for CUDA to be consistently slower just doesn't make sense.

Here’s a link to my original thread which has some more details.

https://community.adobe.com/t5/premiere-pro-discussions/cuda-vs-software-only-render-engine-cuda-is-slower-than-software-only/m-p/15129466#M551316

The general overview is this:

Issue – With Mercury Playback Engine set to CUDA, export times are unpredictable and often take significantly longer than when Software Only is selected.
Adobe Premiere Pro version number: v24.6.4 / v.25.2.0 (I use v24 daily, but have tested in v25)
Operating system - Windows 11 Pro [23H2]
System Info: CPU, GPU, RAM, HD:

CPU - i9-13900K
GPU - Nvidia 3070Ti
GPU driver (Nvidia) - Studio (32.0.15.6590) / Studio (32.0.15.7216)
iGPU driver (Intel) - 32.0.101.6557
RAM - 96 GB (48GB x2 - Corsair Vengeance DDR5 5600MHz CL40)
Hard Drives
- C:Drive = Sabrent NVMe Gen4 SSD (2TB)
- Media Cache = Samsung 990 Pro NVMe SSD (2TB)
- Media = Samsung 990 Pro NVME SSD (4TB)
- Export = Samsung 850 Pro NVME SSD (256GB)
PSU - EVGA 750w

Video format:

Mixed, based on different projects. Normally I'll be editing using transcodes of the rushes, either DNxHD36 or ProRes422 Proxy. However, I will frequently have projects where I pull in h.264 assets. Rushes themselves are typically XAVC based, ProRes, or H.264

Workflow details: Mixed depending on the project. Sometimes this will include Multicams, Essential Graphics, or Dynamic Linked AE comps. However, the test export that I've been using is just footage, with no effects applied.
Steps to reproduce - (Very important!) I wish I could give you some, but honestly I can't find a pattern. In my test timeline, Software Only (Mercury Playback) is always faster (approx 4min). But what I tend to find is that the CUDA (Mercury Playback) version will be fastest immediatley after I restart Premiere (7:30min). This time then climbs with subsequent exports/ time. Basically, it seems to be an issue that comes about after I've been using Premiere for a while.

Expected result - Predictable export times
Actual result - Varied export times

FAQ

Hardware decoding settings are set to both Intel and NVIDIA
Exporting to H.264 (with Hardware Acceleration enabled)
Windows 11 Graphics Preferences (for Premiere/ After Effects/ Media Encoder) set to 'High Performance'
Footage being used in test export is 4:2:0 (8 bit). However, similar results have been seen with other codecs (for example XAVC 4:2:2 (10 bit))
I have tried reinstalling Nvidia driver (via DDU), rolling back to 556.12 and updating to latest
GPU temp is around 45c during export (maxes out at 55c)

I'm sure there will have been bits that I've not mentioned, so any questions, feel free to ask!

---UPDATE---

Funnily enough - I was writing this while I had an export going in the background. That 18min timeline took 24min to export with CUDA (Mercury Playback) enabled, but after I switched to Software Only, it took 5:40min.

That's exactly the sort of thing I mean!

J

jamie_macAuthor

Inspiring

Hey,

So I just got off with one of the Adobe customer service chat people who remoted into my computer and tried a few things. Sadly, I don't think we're much closer to a solution, but figured I'd update things here.

Main thing was rolling back my GPU driver to 556.12 (specifically the Game Ready version). Apparently there is an issue with the 560 series of drivers, so they recommended 556.12 as a stable option.

I asked about Game vs Studio drivers and they said that Game was more optimised for now.

After Restarting Premiere and starting a quick test export of my baseline test, things looked like they were off to a promising start. But ultimately, it ended up taking about 8 minutes (for reference Software only does it in about 4 min).

I then repeated the test immediately afterwards and the time climbed to 12 minutes, which is consistent with what I've found up until this point.

Sadly, the support rep didn't have a good answer for what was going on there. They were mainly focused on getting the GPU being utilised, which they were keen to point out was peaking at up to 100% "Utilisation" in task master during the export.

Unfortunately, as I confirmed via a screenshot from last week where I was getting 98%, lack of 'Utilisation' doesn't seem to be an indicator of the problem.

One thing that they did suggest was that the reason for the fast Software vs CUDA export time was all down to the high amount of system RAM that I have (96GB).
Essentially, what they were saying was that on simpler projects, where the RAM doesn't get filled, Software Only (i.e. CPU & RAM) will out perform the GPU. If the RAM get's overwhelmed, then the GPU would be a better option.

So according to them, the answer to the question of "if I have an Nvidia GPU, should I set the Render Engine to CUDA", is... it depends.

I didn't get a clear idea of how to recognise when one would be better than the other. Which is obviously frustrating as it means I'm still stuck where I am, seconding guessing whether or not to cancel exports and restart them in the alternative renderer.

Ultimately, they said that they weren't sure what else they could do as it would be up to the development team to "analyse the algorithm".

So, in short. There is apparently a known issue with the 560 series of GPU drivers, but rolling back didn't change anything for me.

There also doesn't seem to be any sign that this issue is going away anytime soon.

_____

For reference, I'm using:

- Premiere Pro (v24.6.4)

- Media Encoder (v24.6.4)

J

jamie_macAuthor

Inspiring

Are there any moderators who are able to move this thread to 'Bugs' rather than 'Discussions'?

I'm hoping that it might stand a better chance of getting resolved that way!

J

jamie_macAuthor

Inspiring

Hey there,
I've not had any updates on this for a while and am still experiencing issues.

Just the other day, I waited 19 min for a simple 17 long video to export using CUDA. This seemed crazy to me so I switched to Software Only and the export time dropped down to 4 min.

Not having a reliable idea of export times is quite frustrating and has already landed me in hot water on a few tight turnaround projects.

Given the responses so far, it seems like this isn't expected behaviour, so I'm guessing it would be a better to move this discussion to bugs. Does anybody know if I can do this? Or shall I just start a new thread over there and link to this one?

Thanks!

A

Alex367466572b5i

Known Participant

@jamie_mac Show a screenshot of the task manager when using CUDA, Performance tab

J

jamie_macAuthor

Inspiring

Things continue to be frustratingly inconsistent.

I've just done a test export with the exact sequence that I was referring to in my previous post and the CUDA version now exported in a time of 7:30, 10:45 and 11:45 minutes in repeated tests. The Software export in contrast took 4:55min

Here's the export logs from today:

01/27/2025 10:49:55 AM : Queue Started


 - Source File: C:\Users\Jamie\AppData\Local\Temp\ProjectFile.prproj
 - Output File: E:\ProjectFile_(Software).mp4
 - Preset Used: Screener_1080p_(w/Proxy+OutputTitle)
 - Video: 1920x1080 (1.0), 25 fps, Progressive, 203 (75% HLG, 58% PQ), Hardware Encoding, Nvidia Codec, 00:17:19:03
 - Audio: AAC, 320 kbps, 48 kHz, Stereo
 - Bitrate: VBR, 1 pass, Target 1.00 Mbps
 - Encoding Time: 00:04:53
01/27/2025 10:54:48 AM : File Successfully Encoded


01/27/2025 10:54:48 AM : Queue Stopped



01/27/2025 11:10:04 AM : Queue Started


 - Source File: C:\Users\Jamie\AppData\Local\Temp\ProjectFile.prproj
 - Output File: E:\ProjectFile_(CUDA).mp4
 - Preset Used: Screener_1080p_(w/Proxy+OutputTitle)
 - Video: 1920x1080 (1.0), 25 fps, Progressive, 203 (75% HLG, 58% PQ), Hardware Encoding, Nvidia Codec, 00:17:19:03
 - Audio: AAC, 320 kbps, 48 kHz, Stereo
 - Bitrate: VBR, 1 pass, Target 1.00 Mbps
 - Encoding Time: 00:11:48
01/27/2025 11:21:53 AM : File Successfully Encoded


01/27/2025 11:21:53 AM : Queue Stopped

Here's the logs from last time, just to show how inconsistent things can be!

01/20/2025 11:45:33 AM : Queue Started


 - Source File: C:\Users\Jamie\AppData\Local\Temp\ProjectFile.prproj
 - Output File: E:\ProjectFile_(CUT01)_(SOFTWARE).mp4
 - Preset Used: Screener_1080p_(w/Proxy+OutputTitle)
 - Video: 1920x1080 (1.0), 25 fps, Progressive, 203 (75% HLG, 58% PQ), Hardware Encoding, Nvidia Codec, 00:17:19:03
 - Audio: AAC, 320 kbps, 48 kHz, Stereo
 - Bitrate: VBR, 1 pass, Target 1.00 Mbps
 - Encoding Time: 00:03:56
01/20/2025 11:49:30 AM : File Successfully Encoded


01/20/2025 11:49:30 AM : Queue Stopped


01/20/2025 11:50:14 AM : Queue Started


 - Source File: C:\Users\Jamie\AppData\Local\Temp\ProjectFile.prproj
 - Output File: E:\ProjectFile_(CUT01)_(CUDA).mp4
 - Preset Used: Screener_1080p_(w/Proxy+OutputTitle)
 - Video: 1920x1080 (1.0), 25 fps, Progressive, 203 (75% HLG, 58% PQ), Hardware Encoding, Nvidia Codec, 00:17:19:03
 - Audio: AAC, 320 kbps, 48 kHz, Stereo
 - Bitrate: VBR, 1 pass, Target 1.00 Mbps
 - Encoding Time: 00:18:54
01/20/2025 12:09:09 PM : File Successfully Encoded

Here's a screenshot of task manager during today's CUDA export

And here's the Software Only export

Thanks!

M

MovieMutant

Participant

I'm having the same issue. Render time using Hardware Acceleration is significantly slower than Software Only.

CPU - i9-12900K
GPU - Nvidia 4070Ti
GPU driver (PC) - Studio (566.14)
RAM - 64 GB (16GB x4 - Corsair Vengeance DDR5 6000MHz CL36)

J

jamie_macAuthor

Inspiring

Good to know that I'm not going crazy and there's somebody else out there who's as confused as me!

FYI - @mayjain7130546 has gone off to do some testing at their end to see if they can give any insights.

M

mayjain7130546

Participating Frequently

Hi @jamie_mac,

Thank you for providing the details. Could you confirm if this significant difference is observed on other devices as well? In our local testing, we haven’t been able to reproduce this issue.

To investigate further, could you please share a sample project (including media files) where this difference occurs when switching between CUDA and Software-Only renderers in File > Project Settings > General > Renderer?

We’ll test the project locally and share our findings with you.

Looking forward to your response.

Best regards,
Mayjain

J

jamie_macAuthor

Inspiring

Hi @mayjain7130546 - thanks for taking a look at it for me. Do you have an email I can send things to?

M

mayjain7130546

Participating Frequently

Thanks. I've sent you a private message. Please check and let me know if you have any questions or need further details."

Show more replies

Sign up

To post, reply, or follow discussions, please sign in with your Adobe ID.

Sign in to Adobe Community

To post, reply, or follow discussions, please sign in with your Adobe ID.

Scanning file for viruses.

This file cannot be downloaded