The Strange Behavior Of My Radeon R9 290 Is Still There
In recent days there have been a few Phoronix readers inquiring why I am not testing with my Radeon R9 290 graphics card in all our frequent comparisons and driver benchmarks. The short story is that the regression since Linux 4.7 remains and for my Radeon R9 290 and others with select Hawaii graphics cards, there still is a performance regression. Though over Christmas I hope to finally find the time to bisect it.
So for those wondering but haven't asked why the R9 290 hasn't been used, it's since there is still that pesky regression... While there was a fix for some, my HIS Radeon R9 290 and that of other select users still are having issues, likely due to differing video BIOS. AMD, meanwhile, reportedly hasn't been able to reproduce this issue with their hardware.
It's been a rather strange issue but admittedly low priority on my end when busy with testing other (and newer) graphics cards, and having a never-ending TODO list of more pressing work to pay the bills. The Phoronix Test Suite should be able to bisect it automatically, at least this gives me the time to make sure those bisecting code-paths are still in good shape, but it still takes lots of system time and unfortunately few systems with easy access to open from all of them being rackmounted.
Affected cards see very low performance, such as these latest results with Linux 4.9 and Mesa 13.1-dev still being affected:
Basically if your card is affected, the performance is unplayable. I had hoped to find the time to run the bisecting tests over Thanksgiving, but that didn't happen. My hope is now that I'll have extra time once the ad revenues drop off around Christmas to sit back, have PTS bisect it, while enjoying some Augustiner to pass the time.
When running these tests over night to confirm the regression was still in place, some workloads like OpenArena were not affected:
But what's rather strange is that after either a certain operation with GpuTest, length of time, or other internal change, the Radeon R9 290 was back to running fast for the remainder of the automated OpenGL benchmarks:
The R9 290 was performing better in the remainder of the tests. What also makes this issue strange is that the Radeon R9 290 continues working fine with the AMDGPU-PRO stack.
Anyhow, hopefully I'll find the extra time over Christmas to be able to bisect this longstanding issue.
Those wanting to dig through this data further can see this OpenBenchmarking.org result file for the metrics as of last night.
So for those wondering but haven't asked why the R9 290 hasn't been used, it's since there is still that pesky regression... While there was a fix for some, my HIS Radeon R9 290 and that of other select users still are having issues, likely due to differing video BIOS. AMD, meanwhile, reportedly hasn't been able to reproduce this issue with their hardware.
It's been a rather strange issue but admittedly low priority on my end when busy with testing other (and newer) graphics cards, and having a never-ending TODO list of more pressing work to pay the bills. The Phoronix Test Suite should be able to bisect it automatically, at least this gives me the time to make sure those bisecting code-paths are still in good shape, but it still takes lots of system time and unfortunately few systems with easy access to open from all of them being rackmounted.
Affected cards see very low performance, such as these latest results with Linux 4.9 and Mesa 13.1-dev still being affected:
Basically if your card is affected, the performance is unplayable. I had hoped to find the time to run the bisecting tests over Thanksgiving, but that didn't happen. My hope is now that I'll have extra time once the ad revenues drop off around Christmas to sit back, have PTS bisect it, while enjoying some Augustiner to pass the time.
When running these tests over night to confirm the regression was still in place, some workloads like OpenArena were not affected:
But what's rather strange is that after either a certain operation with GpuTest, length of time, or other internal change, the Radeon R9 290 was back to running fast for the remainder of the automated OpenGL benchmarks:
The R9 290 was performing better in the remainder of the tests. What also makes this issue strange is that the Radeon R9 290 continues working fine with the AMDGPU-PRO stack.
Anyhow, hopefully I'll find the extra time over Christmas to be able to bisect this longstanding issue.
Those wanting to dig through this data further can see this OpenBenchmarking.org result file for the metrics as of last night.
24 Comments