AMD Graphic card

[Iceshaker Hotshot]

Hello I am considering switching to AMD. Does anyone already have experience with the new AMD 22.7.1 driver. does it also increase the FPS is SL? 

[Henri Beauchamp]

AMD OpenGL drivers have always been slow.

Even the new improved Windows drivers (which bring Windows drivers on par with Linux ones for AMD cards) will perform far less good than an equivalent (same price and hardware levels) NVIDIA card with OpenGL.

The difference is enormous (like 50 to 100% more fps with NVIDIA).

[Kathrine Jansma]

The newer AMD driver is a little faster then the old one. But NVIDIA is still significantly faster.

In addition there are some reports of instabilities or excessive memory usage with the 22.7.1 driver.

Like AMD removed a bunch of OpenGL extensions in the update, which may cause issues.

)

[Jenna Huntsman]

FWIW, on my setup - (I normally run Linux, but on Win 11 for testing purposes). AMD Driver 22.8.1

Am seeing avg. FPS of ~80 in a loaded scene (Belli ground level, facing South West)

Firestorm 6.6.3 (67375) Aug  8 2022 14:47:02 (64bit / SSE2) (Firestorm-Betax64) with Havok support
Release Notes

You are at 203.9, 245.0, 52.4 in Crooked Posse located at simhost-09d9fd0ddccf428d9.agni
SLURL: http://maps.secondlife.com/secondlife/Crooked%20Posse/204/245/52
(global coordinates 270,796.0, 249,845.0, 52.4)
Second Life Server 2022-07-07.573176
Retrieving...

CPU: AMD Ryzen 7 3800X 8-Core Processor              (3900 MHz)
Memory: 32681 MB
Concurrency: 16
OS Version: Microsoft Windows 11 64-bit (Build 22000.856)
Graphics Card Vendor: ATI Technologies Inc.
Graphics Card: AMD Radeon RX 6900 XT
Graphics Card Memory: 32452 MB

Windows Graphics Driver Version: 31.0.12019.5007
OpenGL Version: 4.6.0 Core Profile Context 22.8.1.220810

RestrainedLove API: RLV v3.4.3 / RLVa v2.4.2.67375
libcurl Version: libcurl/7.54.1 OpenSSL/1.1.1l zlib/1.2.11.zlib-ng nghttp2/1.40.0
J2C Decoder Version: KDU v8.2
Audio Driver Version: FMOD Studio 2.02.07
Dullahan: 1.12.3.202111032221
  CEF: 91.1.21+g9dd45fe+chromium-91.0.4472.114
  Chromium: 91.0.4472.114
LibVLC Version: 3.0.16
Voice Server Version: Vivox 4.10.0000.32327

Settings mode: Firestorm
Viewer Skin: Firestorm (Grey)
Window size: 3840x2054 px
Font Used: Deja Vu (96 dpi)
Font Size Adjustment: 0 pt
UI Scaling: 1
Draw distance: 128 m
Bandwidth: 500 kbit/s
LOD factor: 2
Render quality: High-Ultra (6/7)
Advanced Lighting Model: Yes
Texture memory: 2048 MB (1)
Disk cache: Max size 2048.0 MB (55.9% used)
Built with MSVC version 1916
Packets Lost: 0/2,459 (0.0%)
August 15 2022 12:21:39 SLT

 

Edited August 15, 2022 by Jenna Huntsman

[Iceshaker Hotshot]

what graphic setting do you use, shadows and this? 

[Jenna Huntsman]

41 minutes ago, Iceshaker Hotshot said:

what graphic setting do you use, shadows and this? 

ALM: ENABLED
Shadows: NONE
Water Reflections: None (Opaque)
Point Lighting: Full
Draw Distance: 128m
Particle Count: 4096
Anisotropic Filtering: ENABLED
Streamed VBOs: ENABLED
Vsync: DISABLED
Antialiasing: 2x
DoF: ENABLED

 

[Henri Beauchamp]

@Jenna Huntsman

I could not measure at the exact same spot since it is protected by a security ”orb”, but on the nearby road, with the Cool VL Viewer, I get the same 80 fps at 160m DD (instead of 128), 4xAA + SMAA + sharpening shaders (instead of 2xAA + FXAA), Terrain & trees water reflections (instead of opaque water), 8192 particles (instead of 4096), everything else set to the maximum (including a generous 2x mesh LOD boost factor)...

And this is only with an old GTX 1070 Ti and a Core i7-9700K, that is theoretically a less powerful hardware than yours (in particular, the 6900XT is supposed to be of the same class as a RTX 3080, which is waaaaaay more powerful than my old GTX)...

This shows how plain lame are AMD OpenGL drivers...

The proof is here (the blurry UI is due to the fact that I only have a 1920x1200 monitor and had to push the resolution to 3840x2400 in order to match Jenna's screen size: the scaling is causing the blurr):

 

Edited August 15, 2022 by Henri Beauchamp

[Jenna Huntsman]

59 minutes ago, Henri Beauchamp said:

I could not measure at the exact same spot since it is protected by a security ”orb”,

 

I do apologize - I completely forgot I even had one! (Linden Security system hidden in the house!).

It's disabled now.

Testing now with the same version of Firestorm on Linux has roughly the same framerate (maybe averaging 2 FPS or so higher).

FWIW - I think this particular test location is being bottlenecked by something other than the GPU, as that'd explain the similar performance between radically different GPUs.

Edited August 15, 2022 by Jenna Huntsman

[Iceshaker Hotshot]

Thank you for your answers. I was hoping that AMD has caught up thanks to the new drivers. But unfortunately that doesn't seem to be the case. So I won't buy an AMD card. then I will replace my GTX 1080 AMP Extreme+ with something stronger. 

[Henri Beauchamp]

9 hours ago, Jenna Huntsman said:

I do apologize - I completely forgot I even had one! (Linden Security system hidden in the house!).

I made a quick test at your place and got the same result.

9 hours ago, Jenna Huntsman said:

FWIW - I think this particular test location is being bottlenecked by something other than the GPU, as that'd explain the similar performance between radically different GPUs.

It is true, that once the GPU is powerful enough and taken out of the equation, the viewer performances scale almost proportionally with the CPU single core performances.

However, your Ryzen 3800X is just as good as my overclocked (5.0GHz) i7-9700K, if not better: while running 300MHz slower in turbo mode than my 9700K, the 3800X got twice the amount of L2 cache (same amount of L1) and almost thrice the amount of L3 cache (32MB against 12MB), which is very important for loads such as the viewer...

Also, if your viewer got a ”Swap” fast timer, you could have a look at that figure: it reflects the amount of time per frame the CPU has to wait for the GPU to complete all the drawing calls, meaning it is a good measure of the GPU+drivers performances (the lower that time for a given scene, the best the performances). FYI, with the Cool VL Viewer, I get 5.5ms at your place for a 12ms total frame time (i.e. over 46% of the frame rendering time is taken by the GPU; ”over”, because there are other GL calls than the swap call which also incur (small) waits on the CPU side).

Edited August 16, 2022 by Henri Beauchamp

[Jenna Huntsman]

I've ran a test at my place on CoolVL - 

Cool VL Viewer v1.30.0.12, 64 bits, Aug  7 2022 22:28:44
RestrainedLove viewer v2.09.29.27
Release notes

You are at 270796.0, 249844.2, 52.4  in Crooked Posse located at
simhost-09d9fd0ddccf428d9.agni.secondlife.io (35.91.211.116:12035)
Alias: ec2-35-91-211-116.us-west-2.compute.amazonaws.com
Second Life Server 2022-07-07.573176
Release notes

CPU: AMD Ryzen 7 3800X 8-Core Processor              (3899.99 MHz)
Memory: 32681MB
OS version: Microsoft Windows 10 64 bits v10.0 (build 10586.856)
Memory manager: OS native
Graphics card vendor: ATI Technologies Inc.
Graphics card: AMD Radeon RX 6900 XT
Windows graphics driver version: 31.00.12019.5007
OpenGL version: 4.6.0 Core Profile Context 22.8.1.220810
Detected VRAM: 32677MB
J2C decoder: OpenJPEG: 1.4.0.635f
Audio driver: FMOD Studio v2.02.07
Networking backend: libcurl 7.47.0/OpenSSL 1.0.2u/zlib 1.2.11.zlib-ng
Embedded browser: Dullahan 1.12.3/CEF 91.1.21/Chromium 91.0.4472.114
Packets lost: 9/26936 (0.0%)

Built with: MSVC v1916
Compiler-generated maths: SSE2.

Compile flags used for this build:
/O2 /Oi /DNDEBUG /D_SECURE_SCL=0 /D_HAS_ITERATOR_DEBUGGING=0 /DWIN32 /D_WINDOWS /W3 /GR /EHsc /std:c++14 /EHs /fp:fast /MP /TP /W2 /c /nologo /GS- /Zc:threadSafeInit- /DLL_WINDOWS=1 /DUNICODE /D_UNICODE /DWINVER=0x0601 /D_WIN32_WINNT=0x0601 /DLL_PHMAP=1 /DLL_NETBIOS=1 /DBOOST_ALL_NO_LIB /DLL_FMOD=1 /DAPR_DECLARE_STATIC /DAPU_DECLARE_STATIC /DXML_STATIC /DCURL_STATICLIB=1

with approximately the same settings as on Firestorm - average framerate is ~65 FPS

40 minutes ago, Henri Beauchamp said:

if your viewer got a ”Swap” fast timer

I had a quick look around CoolVL, but couldn't find where that is - where did you hide it?

[Henri Beauchamp]

14 minutes ago, Jenna Huntsman said:

I had a quick look around CoolVL, but couldn't find where that is - where did you hide it?

In the fast timer view, look for ”Render” -> ”Swap”. To view it, you will likely have to fold the tree sub-branches of the Render branch (I did not modernize that rarely ever used floater, so it's still the old v1 one, without a scroll bar)

[Henri Beauchamp]

24 minutes ago, Jenna Huntsman said:

with approximately the same settings as on Firestorm - average framerate is ~65 FPS

Careful with ”approximate”... If you wish to compare the viewer performances, you would have to use the exact same settings. The Cool VL Viewer is way more generous, by default (various ”boost” factors to render more detailed stuff, such as attachment and mesh LODs, the number and quality of the textures it loads, the number of objects it loads, and the fact it got a work around for objects that stay invisible at login or fail to rez after a TP, and so rez them properly)...

The first thing that comes to my mind and could influence a lot the frame rate, is the field of view (FOV); the Cool VL Viewer got a default camera axis which is parallel to the ground (instead of pointing towards it in other viewers), meaning you catch more objects in the distance, by default...

[Jenna Huntsman]

20 minutes ago, Henri Beauchamp said:

Careful with ”approximate”... If you wish to compare the viewer performances, you would have to use the exact same settings. The Cool VL Viewer is way more generous, by default (various ”boost” factors to render more detailed stuff, such as attachment and mesh LODs, the number and quality of the textures it loads, the number of objects it loads, and the fact it got a work around for objects that stay invisible at login or fail to rez after a TP, and so rez them properly)...

The first thing that comes to my mind and could influence a lot the frame rate, is the field of view (FOV); the Cool VL Viewer got a default camera axis which is parallel to the ground (instead of pointing towards it in other viewers), meaning you catch more objects in the distance, by default...

Okay, slightly more scientific test, with my camera settings on both viewers set as follows:

My position settings:

Debug setting	Value
CameraOffsetRearView	X: -3 Y: -1.65 Z: -0.5
FocusOffsetRearView	X: 3 Y: -1.65 Z: 0.25

CameraAngle: 1.047 rad
Draw distance: 128m
Particle Count: 4096
Mesh object LOD boost: 2.0
ALM: Enabled; no shadows
Water reflections: Disabled

With these settings, the framerate between the 2 viewers is pretty much the same as far as I can tell.

39 minutes ago, Henri Beauchamp said:

In the fast timer view, look for ”Render” -> ”Swap”. To view it, you will likely have to fold the tree sub-branches of the Render branch (I did not modernize that rarely ever used floater, so it's still the old v1 one, without a scroll bar)

Measured swap time on CoolVL on my setup (in the above location) is 0.0

[Henri Beauchamp]

37 minutes ago, Jenna Huntsman said:

With these settings, the framerate between the 2 viewers is pretty much the same as far as I can tell.

Which would be the expected result, in relatively heavy render scenes; the Cool VL Viewer is better optimized at the C++ code level, which would blatantly show in frame rates in, e.g. a skybox with ALM off, with 600-1000 fps, depending on what your avatar is wearing (where Firestorm will give you 300-500 fps, at best). But we are speaking here about only 1ms per frame of difference or so, and since the viewers share almost exact same shaders, this won't make much of a difference once rendering ”everyday's” scenes.

Quote

Mesh object LOD boost: 2.0

This boost factor (which is a multiplier to the RenderVolumeLODFactor when applied to mesh objects) should be 1.0 to compare with other viewers (that do not have it).

37 minutes ago, Jenna Huntsman said:

Measured swap time on CoolVL on my setup (in the above location) is 0.0

Ah, yes... You'd have to use the Linux viewer to get this info (the Windows version does not have this FTM, IIRC)... I'll have to ”fix” that.

Edited August 16, 2022 by Henri Beauchamp

[Jenna Huntsman]

9 minutes ago, Henri Beauchamp said:

with 600-1000 fps, depending on what your avatar is wearing (where Firestorm will give you 300-500 fps

Not that these framerates are useful to anyone, given the fastest ("common") refresh rate is ~360hz, while (I believe) the most common HRR being 144 / 120 hz ; or likely for most SL users have 60hz monitors (for anyone reading: Any framerate above this just represents wasted frames, as your monitor is not fast enough to display them! Generally speaking, especially for laptop users, you should set your framerate cap to the same as your monitor's refresh rate).

13 minutes ago, Henri Beauchamp said:

You'd have to use the Linux viewer to get this info

I don't have CoolVL on my Linux install currently, but FWIW the measured swap time in Firestorm is 0.1
(Not that a Linux benchmark really helps given the OP is using Windows)

[Henri Beauchamp]

26 minutes ago, Jenna Huntsman said:

Not that these framerates are useful to anyone, given the fastest (”common”) refresh rate is ~360hz, while (I believe) the most common HRR being 144 / 120 hz ; or likely for most SL users have 60hz monitors (for anyone reading: Any framerate above this just represents wasted frames, as your monitor is not fast enough to display them! Generally speaking, especially for laptop users, you should set your framerate cap to the same as your monitor's refresh rate).

Try ”Preferences” floater -> ”Cool features” tab -> ”Misc.” -> ”Frame rate limit”. Unlike VSync or LL's old frame rate limiter, which both ”sleep” the CPU and ruin the tasks performed in the main thread, aside from just rendering (e.g. textures fetching, rezzing, etc), this frame rate limiter won't impact the performances of the viewer while keeping, whenever there is headroom, your GPU and CPU cool, and their fans quiet.

Quote

FWIW the measured swap time in Firestorm is 0.1

Yes, I am seeing this too when running the Windows build of my viewer under Wine... Perhaps the FTM is simply misplaced in the Windows part of the code and does not actually measure the swap time... I'll have to check it, time permitting (the Cool VL Viewer is primarily a Linux viewer, so such Windows minor glitches are not high on my priority list 😛 ). In the mean time, the Linux builds would give you relevant figures to measure the GPU+drivers performances.

Edited August 16, 2022 by Henri Beauchamp