Recently I've considered purchasing an AMD GPU for testing purposes and because of the recent benchmarking article I finally decided to abuse my wallet and get myself one. So, here it is along with some of my first impressions!
Update: After messing with the AMD Catalyst Installer I managed to get FGLRX 15.20.1046 (15.7) up and running. This had slight performance increases in places, for example Metro Redux and Unigine Valley. However, a lot of the time the impact was either very minimal or non-existent. The updated driver thus only affected the Valley benchmark in this article and that graph has been changed and corrections added accordingly.
You all are probably aware of the message that gets spread around about AMD GPUs on Linux. They are slow and the proprietary driver is unreliable but supposedly they have well-maintained and quite competent open source drivers. I've never actually used AMD GPUs and the only ATI I had was when I was like 7 years old and I wasn't using Linux at the time so naturally I was at the very least curious to see what the reality was like.
I picked the Asus Strix R7 370 4G Gaming mainly because of its price and secondly because it's part of the new 300 series GPUs which in my opinion makes my opinions and tests very valid in this day and age. The R7 370 is actually a rebrand of a rebrand and isn't really anything fancy or extremely new. However, this has one positive side effect which is that the card has quite mature support in the open source RadeonSI driver. And why did I pick a card that comes with 4GB of VRAM? Well, I don't actually have a technical reason that I could come up with. I simply wanted to make sure that the card, which is in theory about as fast as my GTX 760, was at least objectively better at something. I also wanted to make a joke about Liam's 3.5GB of VRAM.
The first thing I did after booting it up was testing what kind of performance I could get out of this card. I had some older Nouveau benchmark results lying around so I figured I would pit it against my 760 in a duel of benchmarks using both the open source and the closed source drivers to compare the performance. So I ran some basic benchmarks you are probably very familiar with.
Because of recent Mesa updates both of my cards have access to the tessellation shader with Nouveau being able to access features up to OpenGL 4.1. The R7 370 reported only OpenGL 3.3 with the Oibaf drivers but tessellation shader was found in the supported features making Heaven a nice benchmark. RadeonSI didn't perform too well and lost to the 760 with Nouveau (though just slightly) even when Nouveau could only reclock to the second highest power state. When I asked around on the IRC I was told that this could be because AMD's cards have fewer cores that do tessellation. The Nvidia blob was the obvious winner with FGLRX 15.20.1013 coming second.
When I turned off tessellation the results changed quite a bit. RadeonSI ran twice as fast as Nouveau and was catching up to FGLRX performance. Not a bad display for an open driver I'd say. FGLRX was once again second with the Nvidia blob running quite a bit faster than its competition.
Correction: After updating to 15.7 (15.20.1046) FGLRX started running faster than RadeonSI. Graph is updated.
Because I lack the imagination, I decided to run another Unigine benchmark. However, this is where the results got pretty interesting. RadeonSI actually beat FGLRX. I was quite surprised to see that and I had to do additional runs to verify the results. Of course the RadeonSI results were still pretty far off from the Nvidia driver.
Correction: HL2: Lost Coast runs without graphical glitches with Catalyst 15.7.
Among the benchmarks run was Half-Life 2: Lost Coast. This benchmark isn't all that GPU intensive but it can be pretty heavy with everything maxed out. Here FGLRX was almost as fast as the Nvidia driver but those results actually are probably at least a little bit off. FGLRX, while it was running quite fast, actually had graphics corruption during the benchmark. Every other driver ran the benchmark well without any noticeable glitches but FGLRX was all over the place with terrain, sky and shadows messing up quite a bit.
I also played some games to test the card out without doing comparative benchmarks. Among these was Shadow of Mordor which currently only officially supports Nvidia because of slowdowns on AMD cards. Of course this doesn't always mean that the game doesn't run nicely, for example Borderlands 2 still doesn't support AMD but my testing had it running at 50-60 FPS after dropping a setting or two. However, here the warning is there for a good reason.
As you can see, Shadow of Mordor on FGLRX is just a mess. The results are weird because the performance is constantly bad and the actual drop-off is pretty small going from Low to Very High. I'm not an expert but it kind of looks like there's a bottleneck of some sort in the driver itself. On Ultra the whole game crashed. So, AMD folks, I'd recommend you stay away from this particular game. But hey, at least I had plenty of memory to go about this time, right?
The disaster department also included other big games like Dying Light, which ran at 22 FPS or lower even when the settings were turned down and the game was running at 720p, and Metro 2033 Redux. However, the games that ran on RadeonSI mostly ran better with FGLRX.
So, that's basically what the overall situation looks like from purely performance standpoint. As you can see, AMD side does indeed perform worse in the closed driver space. Overall FGLRX is slower and more buggy than the Nvidia drivers. I'd say if you went with just the closed source driver there wouldn't really be a point to using AMD hardware. FGLRX is just a closed source driver but not as good as the competition. However, just speaking of the closed source driver wouldn't do justice to the best thing about this particular card. That thing of course being the open source driver.
Yes, I did showcase the performance and yes it's most of the time even slower than FGLRX but those are just performance numbers and they don't show some of the best things about the open drivers. First of all, the overall desktop performance seems a lot nicer with the open Mesa drivers. Shaking windows around doesn't result in as much tearing for example. And of course RadeonSI has the advantage of being partially developed by AMD and with AMD-provided documentation meaning that it get more out of the card. There aren't reclocking issues like there are with Nouveau and the driver seems very robust. Some of the games that crash the entire system on Nouveau run just fine with RadeonSI. Article taken from GamingOnLinux.com.
Update: After messing with the AMD Catalyst Installer I managed to get FGLRX 15.20.1046 (15.7) up and running. This had slight performance increases in places, for example Metro Redux and Unigine Valley. However, a lot of the time the impact was either very minimal or non-existent. The updated driver thus only affected the Valley benchmark in this article and that graph has been changed and corrections added accordingly.
You all are probably aware of the message that gets spread around about AMD GPUs on Linux. They are slow and the proprietary driver is unreliable but supposedly they have well-maintained and quite competent open source drivers. I've never actually used AMD GPUs and the only ATI I had was when I was like 7 years old and I wasn't using Linux at the time so naturally I was at the very least curious to see what the reality was like.
I picked the Asus Strix R7 370 4G Gaming mainly because of its price and secondly because it's part of the new 300 series GPUs which in my opinion makes my opinions and tests very valid in this day and age. The R7 370 is actually a rebrand of a rebrand and isn't really anything fancy or extremely new. However, this has one positive side effect which is that the card has quite mature support in the open source RadeonSI driver. And why did I pick a card that comes with 4GB of VRAM? Well, I don't actually have a technical reason that I could come up with. I simply wanted to make sure that the card, which is in theory about as fast as my GTX 760, was at least objectively better at something. I also wanted to make a joke about Liam's 3.5GB of VRAM.
The first thing I did after booting it up was testing what kind of performance I could get out of this card. I had some older Nouveau benchmark results lying around so I figured I would pit it against my 760 in a duel of benchmarks using both the open source and the closed source drivers to compare the performance. So I ran some basic benchmarks you are probably very familiar with.
Because of recent Mesa updates both of my cards have access to the tessellation shader with Nouveau being able to access features up to OpenGL 4.1. The R7 370 reported only OpenGL 3.3 with the Oibaf drivers but tessellation shader was found in the supported features making Heaven a nice benchmark. RadeonSI didn't perform too well and lost to the 760 with Nouveau (though just slightly) even when Nouveau could only reclock to the second highest power state. When I asked around on the IRC I was told that this could be because AMD's cards have fewer cores that do tessellation. The Nvidia blob was the obvious winner with FGLRX 15.20.1013 coming second.
When I turned off tessellation the results changed quite a bit. RadeonSI ran twice as fast as Nouveau and was catching up to FGLRX performance. Not a bad display for an open driver I'd say. FGLRX was once again second with the Nvidia blob running quite a bit faster than its competition.
Correction: After updating to 15.7 (15.20.1046) FGLRX started running faster than RadeonSI. Graph is updated.
Because I lack the imagination, I decided to run another Unigine benchmark. However, this is where the results got pretty interesting. RadeonSI actually beat FGLRX. I was quite surprised to see that and I had to do additional runs to verify the results. Of course the RadeonSI results were still pretty far off from the Nvidia driver.
Correction: HL2: Lost Coast runs without graphical glitches with Catalyst 15.7.
Among the benchmarks run was Half-Life 2: Lost Coast. This benchmark isn't all that GPU intensive but it can be pretty heavy with everything maxed out. Here FGLRX was almost as fast as the Nvidia driver but those results actually are probably at least a little bit off. FGLRX, while it was running quite fast, actually had graphics corruption during the benchmark. Every other driver ran the benchmark well without any noticeable glitches but FGLRX was all over the place with terrain, sky and shadows messing up quite a bit.
I also played some games to test the card out without doing comparative benchmarks. Among these was Shadow of Mordor which currently only officially supports Nvidia because of slowdowns on AMD cards. Of course this doesn't always mean that the game doesn't run nicely, for example Borderlands 2 still doesn't support AMD but my testing had it running at 50-60 FPS after dropping a setting or two. However, here the warning is there for a good reason.
As you can see, Shadow of Mordor on FGLRX is just a mess. The results are weird because the performance is constantly bad and the actual drop-off is pretty small going from Low to Very High. I'm not an expert but it kind of looks like there's a bottleneck of some sort in the driver itself. On Ultra the whole game crashed. So, AMD folks, I'd recommend you stay away from this particular game. But hey, at least I had plenty of memory to go about this time, right?
The disaster department also included other big games like Dying Light, which ran at 22 FPS or lower even when the settings were turned down and the game was running at 720p, and Metro 2033 Redux. However, the games that ran on RadeonSI mostly ran better with FGLRX.
So, that's basically what the overall situation looks like from purely performance standpoint. As you can see, AMD side does indeed perform worse in the closed driver space. Overall FGLRX is slower and more buggy than the Nvidia drivers. I'd say if you went with just the closed source driver there wouldn't really be a point to using AMD hardware. FGLRX is just a closed source driver but not as good as the competition. However, just speaking of the closed source driver wouldn't do justice to the best thing about this particular card. That thing of course being the open source driver.
Yes, I did showcase the performance and yes it's most of the time even slower than FGLRX but those are just performance numbers and they don't show some of the best things about the open drivers. First of all, the overall desktop performance seems a lot nicer with the open Mesa drivers. Shaking windows around doesn't result in as much tearing for example. And of course RadeonSI has the advantage of being partially developed by AMD and with AMD-provided documentation meaning that it get more out of the card. There aren't reclocking issues like there are with Nouveau and the driver seems very robust. Some of the games that crash the entire system on Nouveau run just fine with RadeonSI. Article taken from GamingOnLinux.com.
0 Likes
Some you may have missed, popular articles from the last month:
The comments on this article are closed.
QuoteI also wanted to make a joke about Liam's 3.5GB of VRAM.
I hate you.
QuoteOh, and by the way, has anyone seen my $75?
Oh you, couldn't resist a dig at the other guy now could you :D
0 Likes
In order to gain support for OpenGL 4.1 (and soon 4.2) with RadeonSI, you need to have Mesa built against LLVM 3.7. Oibaf only builds Mesa against LLVM 3.6, sadly. If you were using Arch Linux, it'd be a simple matter of adding the mesa-git repository as that does build Mesa against LLVM 3.7.
0 Likes
QuoteNot a bad display for an open driver I'd say.
Please not you too, Samsai.
Statements such as this, which imply an expectation of inferiority, tacitly
encourage the following fallacy:
Quote"free software [driver] => subpar [performance]"
I've read things like this before in this site (both by editors and
commenters), but I reply now because it seemed to me that you, Samsai,
understand and appreciate free software (call it open source if you want).
For those that don't immediately see the argument, I spell it out:
"Open driver" does not imply subpar (performance) or anything of the sort. If
anything strongly favours subpar results is working without proper hardware
documentation and with little or no financial support. *Not* being free/open.
0 Likes
ElectricPrism Aug 14, 2015
Quoting: GuestI know I'm wasting words, but for those saying they think AMD should fix it - no, at least not entirely. A driver team pumping man-hours into workarounds and such is trying to cure the symptom, not the cause: shitty code in the first place.
There are reasons that games suddenly get patched and work just fine on AMD cards, without driver updates. The bulk of the effort is actually in getting developers writing the games properly in the first place - that's where the most gains will be seen.
I think this comment relates to a recent benchmark on Phoronix about OpenGL Benchmarks.
I specifically wanted to hilight Xonotic performance on Ubuntu as doing terrible.
The reason this is interesting is because the "writing proper code" issues certainly plague both proprietary and open source as seen in that result.
In a open source world a developer could either fix the app or the driver, but in a closed as it's put "workarounds" become common place even breaking spec.
Last edited by ElectricPrism on 14 August 2015 at 1:27 am UTC
0 Likes
Quoting: ElectricPrismI think this comment relates to a recent benchmark on Phoronix about OpenGL Benchmarks.
I specifically wanted to hilight Xonotic performance on Ubuntu as doing terrible.
Well, I don't know what is it with Michael's setup/hardware or whatever he is doing, but he is consistently getting low Linux results in his tests.
The Shadow of Mordor benchmark is a good example where I get much better results with my older hardware than his, but this Xonotic benchmark really takes the cake.
His Xonotic results are simply ridiculous, no way Xonotic should work so bad with his hardware.
With my hardware I get these results on SteamOS:
View video on youtube.com
0 Likes
Quoting: BTRENice work, I'm glad I'm not the only AMD user anymore. You should try benchmarking Talos. I'm not 100% about the rebranding/naming scheme but you should have more or less the same card as me (7870) with more vram, right? I'd be interested to see how much of a difference that makes since I also run the RadeonSI drivers.
The GPU is exactly the same, it's just made with possibly better parts and has been optimized. They did a test with a 390x and downclocked the GPU and VRAM to 290x levels and they performed exactly the same. Not to say it's so bad really - if a 390x can perform close to a 980 for about a 100 dollars less I'd say it's a good buy. It's just that people who have the previous generation equivalents don't really have an incentive to buy a new card. Instead I grabbed a second 290x (had to downclock the a bit though).
0 Likes
FutureSuture Aug 14, 2015
Thank you so much for doing these benchmarks, Samsai. They are very much appreciated and I hope to see more in the future.
What I gathered from these benchmarks is that Nvidia still reigns supreme when it comes to playing games native to Linux, that AMD's open source driver is beating its proprietary counterpart here and there whilst being far more stable as a whole, and that AMD reigns supreme for playing games on Linux via Wine.
AMD has two advantages over Nvidia, namely its open source driver and playing games on Linux using Wine, but Nvidia still has the top advantage because playing games on Linux that are native is the most important factor to consider.
What I gathered from these benchmarks is that Nvidia still reigns supreme when it comes to playing games native to Linux, that AMD's open source driver is beating its proprietary counterpart here and there whilst being far more stable as a whole, and that AMD reigns supreme for playing games on Linux via Wine.
AMD has two advantages over Nvidia, namely its open source driver and playing games on Linux using Wine, but Nvidia still has the top advantage because playing games on Linux that are native is the most important factor to consider.
1 Likes, Who?
QuoteI also wanted to make a joke about Liam's 3.5GB of VRAM.
Oh man that quote alone was worth reading it. :) I love the way you write, Sammy. It's direct, to the point, well balanced and detailed without going into too much detail or digressions.
Also, these kind of tests are very valuable at this infant stage of Linux gaming, where a lot of uncertainty and myths are dangling around.
The only thing missing now is a Windows partition on the test machines, for comparisons of game performance.
0 Likes
Quoting: BeamboomQuoteI also wanted to make a joke about Liam's 3.5GB of VRAM.
Oh man that quote alone was worth reading it. :) I love the way you write, Sammy. It's direct, to the point, well balanced and detailed without going into too much detail or digressions.
Also, these kind of tests are very valuable at this infant stage of Linux gaming, where a lot of uncertainty and myths are dangling around.
The only thing missing now is a Windows partition on the test machines, for comparisons of game performance.
There may be news on the Windows front in future, as I have a HDD ready for it now, I just need to see if Samsai's Win7 key works for it.
0 Likes
At first thanks for benchmarks for AMD videocard.
Could you clarify please: FGLRX 15.20 is Catalyst 15.7? On Ubuntu Catalyst 15.3beta AFAIR also was marked as fglrx 15.20.
Could you also compare game's performance between Linux and Windows ports with the same video cards if WIndows OS is available for you?
Could you clarify please: FGLRX 15.20 is Catalyst 15.7? On Ubuntu Catalyst 15.3beta AFAIR also was marked as fglrx 15.20.
Could you also compare game's performance between Linux and Windows ports with the same video cards if WIndows OS is available for you?
0 Likes
Patreon
PayPal
See more from me