Closed Bug 1417241 Opened 8 years ago Closed 7 years ago

FF 57.0: high GPU usage on Youtube VP9 videos on AMD RX 570 driver 17.11.1 WHQL

Categories

(Core :: Audio/Video: Playback, defect, P2)

Product:
Core
Component:
Audio/Video: Playback
Version:
57 Branch
Type:
defect

Tracking

()

Status:
RESOLVED DUPLICATE of bug 1461268
Tracking Flags:
Tracking Status
platform-rel --- ?
firefox57 --- affected

People

(Reporter: spulci, Unassigned)

References

Details

(Whiteboard: [platform-rel-AMD])

Attachments

(1 file)

ff_57_system_details.txt
9.87 KB, text/plain
Details
Attached file ff_57_system_details.txt —
User Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:57.0) Gecko/20100101 Firefox/57.0 Build ID: 20171112125346 Steps to reproduce: 1. Use FF 57.0 64 bit on Windows 10 1709 (build) 16299.64 2. Check in the about:config that new VP9 Amd codec is enabled (by default) and hardware acceleration enabled (by default too) 3. Use an AMD Radeon Rx570 based card with a fresh install of WHQL 17.11.1 driver 4. Open a software that display GPU usage (like vendor Trixx software from Sapphire in my case) 4. Open a VP9 FHD coded Video on Youtube (only VP9 videos are affected by this issue) 5. Compare results with a FHD avc1/mp4 youtube video Actual results: You can see high cpu clock while playing videos. I can observe the GPU clock using my video card vendor software (Trixx for Sapphire Nitro cards). Before starting video play the clock speed is 300mhz. While playing VP9 video it jumps to 1209mhz. Also GPU memory clock jump from 300mhz to 1750mhz. My UPS register a wattage increase of 30-35 watts during VP9 video playing. Now open a non VP9 video source like an avc1 one on step 5: you will see GPU clock @300-310mzh, GPU memory clock @300mhz. Power usage increase about 10-15 watt during video play Expected results: Videos on step 4 and 5 are both hardware accelerated. The new VP9 codec is power hungry and throttles gpu and its memory to high clock speeds with abnormal power consumption if compared to avc1 videos (even they are both same resolution FHD). Optimized AMD VP9 coded should use far low resources and power as stated in FF 57.0 changelog
Blocks: 1332234
status-firefox57: --- → affected
Component: Untriaged → Audio/Video: Playback
Product: Firefox → Core
(In reply to Simone Pulcini from comment #0) > You can see high cpu clock while playing videos. I can observe the GPU clock > using my video card vendor software (Trixx for Sapphire Nitro cards). Before So you are seeing high CPU clock or high GPU Clock? Or both? > starting video play the clock speed is 300mhz. While playing VP9 video it > jumps to 1209mhz. Also GPU memory clock jump from 300mhz to 1750mhz. What is jumping to 1209MHz? CPU or GPU? Can you provide a link to a H264 only youtube video? Or the steps you are using to switch between H264 and VP9 in youtube. When playing a video, can you copy the overlay content of "stats for nerds" (can be obtained by right clicking on the youtube player) Also, please install https://addons.mozilla.org/en-US/firefox/addon/devtools-media-panel/ and follow the instructions on how to install and go to the media devtools tab. press Refresh Now button, and copy the content here. Thank you
Flags: needinfo?(spulci)
Priority: -- → P3
(In reply to Jean-Yves Avenard [:jya] from comment #1) > > You can see high cpu clock while playing videos. I can observe the GPU clock > > using my video card vendor software (Trixx for Sapphire Nitro cards). Before > > So you are seeing high CPU clock or high GPU Clock? Or both? Sorry my typo! This bug is only related to GPU (Graphic Processing Unit). During the bug CPU is showing a normal behaviour (not affected) > > starting video play the clock speed is 300mhz. While playing VP9 video it > > jumps to 1209mhz. Also GPU memory clock jump from 300mhz to 1750mhz. > > > What is jumping to 1209MHz? CPU or GPU? GPU. All clock speed are GPU related > Can you provide a link to a H264 only youtube video? Or the steps you are > using to switch between H264 and VP9 in youtube. Here you are: two sample videos find on youtube with details a. VP9 (248) / opus (251) youtube video (issue affected): https://youtu.be/9h31Ez_zF7g Video ID / CPN 9h31Ez_zF7g / VixPjn-jyQjVRcRR Viewport 2560x1440 Current / Optimal Res 1920x1080@30 / 1920x1080@30 Codecs vp9 (248) / opus (251) Host r6---sn-hpa7znsl b. avc1.640028 (137) / mp4a.40.2 (140) youtube video (not affected by the issue): https://youtu.be/rVYrFmHUrMU Video ID / CPN rVYrFmHUrMU / EXiw4ex0RURNkaRZ Viewport 1280x720 Current / Optimal Res 1920x1080@12 / 1920x1080@12 Codecs avc1.640028 (137) / mp4a.40.2 (140) Host r2---sn-hpa7znse Both a. and b. are GPU decoded (hardware acceleration enabled). > Also, please install > https://addons.mozilla.org/en-US/firefox/addon/devtools-media-panel/ and > follow the instructions on how to install and go to the media devtools tab. > press Refresh Now button, and copy the content here. During a. playback: Media Info : [ 0 : { url : "https://www.youtube.com/watch?v=9h31Ez_zF7g&feature=youtu.be" mediaElements : [ 0 : { currentSrc : "blob:https://www.youtube.com/1dea2a9b-87e9-412a-8456-ee2321db159b" currentTime : 10.496291 readyState : 4 videoPlaybackQuality : { ratio : "100%" droppedVideoFrames : 0 corruptedVideoFrames : 0 } bufferedRanges : [ 0 : { start : 0 end : 39.381 } ] mozMediaSourceObject : [ 0 : { sourceBuffers : [ 0 : { start : 0 end : 39.381 } ] } 1 : { sourceBuffers : [ 0 : { start : 0 end : 48.047 } ] } ] debugInfo : { Container Type : "MediaSource" Audio Decoder(audio/opus) : "opus audio decoder" Audio Frames Decoded : "627" Audio State : "ni=0 no=0 wp=0 demuxr=0 demuxq=0 decoder=0 tt=-1,0 tths=-1 in=627 out=627 qs=0 pending=0 wfd=0 eos=0 ds=0 wfk=0 sid=0" Video Decoder(video/vp9, 1920x1080 @ 29,98) : "amd vp9 hardware video decoder (remote)" Hardware Video Decoding : "enabled" Video Frames Decoded : "319 (skipped=0)" Video State : "ni=0 no=0 wp=0 demuxr=0 demuxq=0 decoder=0 tt=-1,0 tths=-1 in=320 out=319 qs=1 pending:0 wfd=0 eos=0 ds=0 wfk=0 sid=1" Dumping Data for Demuxer : "122b63ba800" Dumping Audio Track Buffer(audio/webm) : "mLastAudioTime=12,541000" Audio Track Buffer Details : "NumSamples=1969 Size=1174849 Evictable=371548 NextGetSampleIndex=627 NextInsertionIndex=1969" Audio Track Buffered : "ranges=[(0,000000, 39,381000)]" Dumping Video Track Buffer(video/webm) : "mLastVideoTime=10,677000" Video Track Buffer Details : "NumSamples=1440 Size=6203286 Evictable=1223045 NextGetSampleIndex=320 NextInsertionIndex=1440" Video Track Buffered : "ranges=[(0,000000, 48,047000)]" MediaDecoder State : "channels=2 rate=48000 hasAudio=1 hasVideo=1 mPlayState=PLAYING mdsm=122c258e000" MediaDecoderStateMachine State : "GetMediaTime=10536479 GetClock=10538250 mMediaSink=122c2566a00 state=DECODING mPlayState=3 mSentFirstFrameLoadedEvent=1 IsPlaying=1 mAudioStatus=idle mVideoStatus=idle mDecodedAudioEndTime=12534500 mDecodedVideoEndTime=10644000mAudioCompleted=0 mVideoCompleted=0mIsPrerolling=0" VideoSink Status : "IsStarted=1 IsPlaying=1 VideoQueue(finished=0 size=4) mVideoFrameEndTime=10545000 mHasVideo=1 mVideoSinkEndRequest.Exists()=0 mEndPromiseHolder.IsEmpty()=0" } } ] } ] [ { "url": "https://www.youtube.com/watch?v=9h31Ez_zF7g&feature=youtu.be", "mediaElements": [ { "currentSrc": "blob:https://www.youtube.com/1dea2a9b-87e9-412a-8456-ee2321db159b", "currentTime": 10.496291, "readyState": 4, "videoPlaybackQuality": { "ratio": "100%", "droppedVideoFrames": 0, "corruptedVideoFrames": 0 }, "bufferedRanges": [ { "start": 0, "end": 39.381 } ], "mozMediaSourceObject": [ { "sourceBuffers": [ { "start": 0, "end": 39.381 } ] }, { "sourceBuffers": [ { "start": 0, "end": 48.047 } ] } ], "debugInfo": { "Container Type": "MediaSource", "Audio Decoder(audio/opus)": "opus audio decoder", "Audio Frames Decoded": "627", "Audio State": "ni=0 no=0 wp=0 demuxr=0 demuxq=0 decoder=0 tt=-1,0 tths=-1 in=627 out=627 qs=0 pending=0 wfd=0 eos=0 ds=0 wfk=0 sid=0", "Video Decoder(video/vp9, 1920x1080 @ 29,98)": "amd vp9 hardware video decoder (remote)", "Hardware Video Decoding": "enabled", "Video Frames Decoded": "319 (skipped=0)", "Video State": "ni=0 no=0 wp=0 demuxr=0 demuxq=0 decoder=0 tt=-1,0 tths=-1 in=320 out=319 qs=1 pending:0 wfd=0 eos=0 ds=0 wfk=0 sid=1", "Dumping Data for Demuxer": "122b63ba800", "Dumping Audio Track Buffer(audio/webm)": "mLastAudioTime=12,541000", "Audio Track Buffer Details": "NumSamples=1969 Size=1174849 Evictable=371548 NextGetSampleIndex=627 NextInsertionIndex=1969", "Audio Track Buffered": "ranges=[(0,000000, 39,381000)]", "Dumping Video Track Buffer(video/webm)": "mLastVideoTime=10,677000", "Video Track Buffer Details": "NumSamples=1440 Size=6203286 Evictable=1223045 NextGetSampleIndex=320 NextInsertionIndex=1440", "Video Track Buffered": "ranges=[(0,000000, 48,047000)]", "MediaDecoder State": "channels=2 rate=48000 hasAudio=1 hasVideo=1 mPlayState=PLAYING mdsm=122c258e000", "MediaDecoderStateMachine State": "GetMediaTime=10536479 GetClock=10538250 mMediaSink=122c2566a00 state=DECODING mPlayState=3 mSentFirstFrameLoadedEvent=1 IsPlaying=1 mAudioStatus=idle mVideoStatus=idle mDecodedAudioEndTime=12534500 mDecodedVideoEndTime=10644000mAudioCompleted=0 mVideoCompleted=0mIsPrerolling=0", "VideoSink Status": "IsStarted=1 IsPlaying=1 VideoQueue(finished=0 size=4) mVideoFrameEndTime=10545000 mHasVideo=1 mVideoSinkEndRequest.Exists()=0 mEndPromiseHolder.IsEmpty()=0" } } ] } ]
Flags: needinfo?(spulci)
Same issue with recently released AMD Radeon driver 17.11.2 (non WHQL)
Issue persists with Amd Radeon driver 17.11.4
I am having a similar GPU core clock experience with a R9 390 and the latest drivers (And older drivers for that matter) on Win10 with 58.0b8
Actual scenario with 17.12.1 Adrenalin WHQL driver is worse than before!! I suggest to temporarily disable the new AMD VP9 Firefox codec (I'm under 57.0.2) to avoid GPU overheat. Infact with this new driver release, watching a VP9 video with Firefox makes the GPU stable @1300mhz, even after closing youtube video or exiting Firefox totally (I double checked that no Firefox processes are still resident in memory with task manager after closing). GDDR5 ram underclock itself @300mhz after closing video correctly, but the GPU will stay at high clock until you reboot your system!!! I made several checks with the two videos linked on this bug page. Only VP9 videos are affected by this new worse scenario. Drivers behave correctly using a 3d game: closing the game will underclock GPU and GDDR5 Ram. As well as other browsers like Opera that shows VP9 videos with a stable 300mhz clock during playback for both GPU and GDDR5. This new issue makes the GPU always power hungry so I hope that Mozilla will take soon under inspection (even with a fast remote fix, disabling the new codec by default while fixing those bugs). I'm available to do more tests for developers on my side.
Does AMD Radeon Settings show 1300mhz clocks after you close the video? FWIW I haven't been able to reproduce the high clock bug with any driver/card combo I've tried.
(In reply to kyle.plumadore from comment #7) > Does AMD Radeon Settings show 1300mhz clocks after you close the video? To detect GPU clock and GDDR5 clock I'm using the vendor software. In my case I'm using Sapphire Trixx 6.4.0 utility. Same datas are showed in the global WATTMAN tab inside the AMD Radeon Settings. I can share a link to a dropbox image of the wattman screenshot before and after playing a VP9 video if needed > FWIW I haven't been able to reproduce the high clock bug with any > driver/card combo I've tried. I really don't know what else to try. I can give you information on my motherboard configuration as well if needed. I suspect a bug in the new optimized VP9 codec for AMD GPUs. No other browsers show this issue with VP9 on my pc
Here a link to a PNG with WATTMAN opened after watching a VP9 Video with Firefox 57.0.2 64bit on a AMD Ryzen 7 1700 with a Sapphire RX570 4GB Nitro Card (Windows 10 1709 build 16299.125). Firefox has been closed before capturing the screenshot and no GPU load is applied (just normal Windows desktop in 2k resolution): https://www.dropbox.com/s/ghleyy5kmtgkd5m/gpu_load.PNG?dl=0 Gpu stays at 100% overheating. I had to reboot to obtain normal behaviour. This scenario affects just Firefox
Good news (?) I was able to find a configuration that shows this issue. From what I can tell, it's limited to certain brands of RX 470/480 cards. We'll look into it and hopefully have it fixed in an upcoming driver release.
(In reply to kyle.plumadore from comment #10) > Good news (?) I was able to find a configuration that shows this issue. From > what I can tell, it's limited to certain brands of RX 470/480 cards. We'll > look into it and hopefully have it fixed in an upcoming driver release. Don't worry and take the time you need @AMD to fix it! BTW, you cite 470/480 card...I've got a 570. I suppose this bug is Polaris based issue, right? And what about Firefox? Is this bug due to particular Firefox behaviour with the new VP9 optimized AMD codec? So no one @Mozilla need to write a fix for this? Are you referring to both scenario: high Gpu Power States during playback and 100% gpu usage even after closing playback and Firefox? Thanks for your support. Let me know if you want one more test before public release. Happy to be of help if needed
(In reply to kyle.plumadore from comment #10) > Good news (?) I was able to find a configuration that shows this issue. From > what I can tell, it's limited to certain brands of RX 470/480 cards. We'll > look into it and hopefully have it fixed in an upcoming driver release. I've raised this as a P2 because we're probably going to need blacklist this specific driver version. Can you give us a list of affected device IDs?
Flags: needinfo?(kyle.plumadore)
Priority: P3 → P2
Adrenalin 17.12.2...still no luck. During playback same bug with a slightly increased power consumption :( Closing the playback and/or Firefox causes the same GPU full throttle 100% @1300mhz while this time the GDDR5 ram behaves correctly going down to 300mhz after playback (but jumping to high frequencies while watching videos).
Sorry for the slow reply - I've been on vacation for the past week.. (In reply to Simone Pulcini from comment #13) > Adrenalin 17.12.2...still no luck. During playback same bug with a slightly > increased power consumption :( Closing the playback and/or Firefox causes > the same GPU full throttle 100% @1300mhz while this time the GDDR5 ram > behaves correctly going down to 300mhz after playback (but jumping to high > frequencies while watching videos). We're still debugging the root cause of this issue - it's unfortunately not as straight forward as it may seem. I'll try to reply here when it's fixed. For your earlier comment, it is indeed 470/480/570/580 series cards that show the issue, but it doesn't appear that all flavours have the issue. (In reply to Anthony Jones (:kentuckyfriedtakahe, :k17e) from comment #12) > (In reply to kyle.plumadore from comment #10) > > Good news (?) I was able to find a configuration that shows this issue. From > > what I can tell, it's limited to certain brands of RX 470/480 cards. We'll > > look into it and hopefully have it fixed in an upcoming driver release. > > I've raised this as a P2 because we're probably going to need blacklist this > specific driver version. Can you give us a list of affected device IDs? I don't have a list of impacted devices (not all 470/480/570/580 boards are impacted, but I'm not exactly sure which are). The following deviceIDs should cover all 470/480/570/580 boards: 0x67C0 .. 0x67FF 0x6980 .. 0x699F
Flags: needinfo?(kyle.plumadore)
One more thing - all my comments are in relation to the high gpu clocks after firefox is closed. Due to the nature of the amd vp9 decoder on current hardware, it's expected for the clocks to raise during vp9 playback. Despite the potentially higher power usage over h.264, it's important to have the vp9 decoder available for high resolution content as youtube is moving away from providing 4k h.264.. Also end users will use less internet bandwidth by using vp9 over h.264 for the same video quality.
(In reply to kyle.plumadore from comment #15) > One more thing - all my comments are in relation to the high gpu clocks > after firefox is closed. Due to the nature of the amd vp9 decoder on current > hardware, it's expected for the clocks to raise during vp9 playback. I see...well I'm surprised. I'm now watching a VP9 video (confirmed by nerd statistics on the video itself) on Youtube using Opera 49.0.2725.64 on the same hardware. It plays with GPU and GDDR5 around 300mhz with a power increase around 5-10 watt (compared to 40 watt and more watt of Firefox 57). Can I assert that Opera VP9 decoder is far more optimized on RX 570 than the Firefox one?
No need to assume, you can be certain that the VP9 AMD decoder isn't used with VP9 content with Opera, but is instead using the libvpx software decoder. So of course, GPU usage will be low, but CPU usage will be higher. To disable the AMD VP9 HW decoder, go to about:config and search for the preference media.wmf.amd.vp9.enabled then set it to false. No need to restart Firefox, next time you play a youtube video it will be using the VP9 software decoder. If power usage is your main concern on your desktop, then I suggest you disable VP9 altogether and only use h264. It's much cheaper to decoder h264 than vp9 both in software and hardware. The downside of course is that for identical bit rates, h264 will give you a lower quality. Set media.webm.enabled to false in about:support to be guaranteed to never use vp9 with YouTube.
(In reply to Jean-Yves Avenard [:jya] from comment #17) > No need to assume, you can be certain that the VP9 AMD decoder isn't used > with VP9 content with Opera, but is instead using the libvpx software > decoder. So of course, GPU usage will be low, but CPU usage will be higher. Thanks for explaining this difference. Btw CPU usage during VP9 playback under Opera is about 3.5% on my Ryzen system. I'm not concerned about power consumption on my desktop system (while I'm on laptop ones). I found it unusual to see this high increase while watching videos. That's all :)
platform-rel: --- → ?
Whiteboard: [platform-rel-AMD]
(In reply to Anthony Jones (:kentuckyfriedtakahe, :k17e) from comment #12) > I've raised this as a P2 because we're probably going to need blacklist this > specific driver version. Can you give us a list of affected device IDs? Still no blacklist with Firefox 58.
Seems fixed today...I'm actually watching several youtube VP9 videos on Firefox 58 with Adreanlin 18.1.1 (as well as the one i posted above on comment 2) . I double checked that media.wmf.amd.vp9.enabled = true and I can say that 1. Stopping or pausing a VP9 video on youtube makes my Sapphire Nitro RX 570 to reduce power state to correct clock: 300mhz for GDDR5 and 300mhz clock for the GPU (higher clocks during playback confirm that the video is GPU accelerated) 2. Power consumption during playback is bit reduced on my system (a few watt less but better than ever) So let's wait for AMD to confirm this. I can't understand what is changed on my system comparing it to a few days ago when the issue was still there. I will keep this issue "under control" in the next few days
Thats likely the new drivers... There's been no change on this code on Firefox side
(In reply to Jean-Yves Avenard [:jya] from comment #21) > Thats likely the new drivers... There's been no change on this code on > Firefox side Right...I agree with you. But even with driver 18.1.1 the issue 1 on comment 20 isn't totally fixed: there are some kind of scenario with Youtube and VP9 where the bug is still present. It's not easy to understand how to reproduce the scenario this time for me but when the bug is triggered it will avoid the GPU to reach lower p-states while stopping playback until a system reboot. :-(
We haven't intentionally resolved this issue in a public driver (yet), but there is some randomness to reproducing it. As a temp workaround without needing to reboot the system, you can kill the amddvr.exe process. You may also prevent it from starting when windows loads to workaround the issue entirely (although you will lose any dvr funcitonality)
same problem here on my rx 570, i fixed it by setting: media.hardware-video-decoding.enabled to false in about:config GPU would normally go to 1375 core clock @ 40+ watt & stay there even after closing Firefox, disabling it makes it sit at 10 watt, 300 on the core. seems to be an issue with 60fps video-playback only
Firefox 59.0.2 64 bit with Radeon Adrenalin 18.3.4: issue seems to be fixed. Double improvement: 1. Bug about high cpu after playback has vanished since this release. I can play VP9 videos: stopping the playback implies GPU to reach its lower pstate in terms of GPU and GDDR5 clock (around 300mhz) 2. During 1080p VP9 playback I can see a greatly improved power consumption: my system reaches from 60 to 85 watts during playback (far less then 105 watt from previous release). Waiting for Kyle Plumadore acknowledge. Meanwhile thanks to all @Mozilla and @AMD.
Adrenalin 18.4.1: the bug is back again (while it was fixed at least on my side with 18.3.4)...and now it's even worst than before: the workaround in comment 23 does not fix the issue with this release. This time, when you start a Youtube VP9 video the GPU goes to full clock speed...stopping the playback or closing the tab has no effect...but closing Firefox make the GPU to come back to lower P-State again.
I only encounter this problem when playing 720p60-fps or 1080p60-fps videos, any other 30fps video seems to not cause any issues. FIX: set media.hardware-video-decoding.enabled to false in about:config a reboot is required for changes to take effect this works with 17.2.1, 18.2.1, 18.1.1, 18.2.1, etc. I am currently using 18.2.1, since that is the latest (non-beta) driver
(In reply to hidde from comment #27) > I only encounter this problem when playing 720p60-fps or 1080p60-fps videos, > any other 30fps video seems to not cause any issues. > > FIX: set media.hardware-video-decoding.enabled to false in about:config > a reboot is required for changes to take effect > > this works with 17.2.1, 18.2.1, 18.1.1, 18.2.1, etc. > I am currently using 18.2.1, since that is the latest (non-beta) driver This is a pretty poor solution as it would disable hardware acceleration completely. There's also no need to reboot after changing the preference. Instead set media.wmf.amd.vp9.enabled to false. No need to restart. Just reload the page
RE comment #28 Thanks for the suggestion, I will try it out in a few hours & report back here. I just do not want my GPU to pull another 30~50 watts seemingly unnecessary, even after I close Firefox.
RE comment #28 your approach did not work, 60fps video playback still uses 63 watts after closing Firefox, 40watts I will continue to keep hardware acceleration disabled for now, since it plays using 11 watt without any issues CPU is at 20watts, idle Firefox closed its at 14watts if you where wondering
Depends on: 1461268
Fixed by bug 1461268
Status: UNCONFIRMED → RESOLVED
Closed: 7 years ago
Resolution: --- → DUPLICATE
You need to log in before you can comment on or make changes to this bug.

Attachment

General

Creator:
Created:
Updated:
Size: