Further to Mumak's post here
http://support.amd.com/en-us/kb-articles/Pages/AMDGPU-Pro-Beta-Mining-Driver-for-Linux-Release-Notes.aspx
- Introduces Large Page Support for Blockchain Compute Workloads. This driver is provided as a beta level support driver which should be considered "as is" and will not be supported with further updates, upgrades or bug fixes. This driver is not intended for graphics or gaming workload
There is no improvement with default 64K page size installed, will enable the large page support and report back later (for rx-480).
AgentB wrote:There is no
)
No significant change.
clinfo reports a reduction in global memory size from (6.075GiB) to (5.347GiB) so i will swap it back later.
I am trying the large pages,
)
I am trying the large pages, and on linux there is some difference. At startup boinc now reports a whopping 6290 GFlops/s calculation speed, it used to be 3900 something on an RX580 4GB card. Calculation time of GPU tasks is reduced maybe 4-5%, it's hard to tell because the difficulty varies between the tasks.
Apparently the improvement on linux can be bigger because default page size is only 4 kB, I set it to 2 MB. Windows defaults to 64 kB. Don't ask me about my configuration, it is a hybrid ROCM and AMDGPU-PRO.
I'm running the amdgpu-pro
)
I'm running the amdgpu-pro 17.40 package on my RX480 Ubuntu hosts and have seen a 4-5% improvement as well, just running with the default settings. The increased page size isn't expected to help E@H, as the GR app doesn't use host memory mapped ram.
mountkidd wrote:I'm running
)
OK what version were you running before upgrading to 17.40 - (i was running 17.30)?
I was running 17.30 before.
)
I was running 17.30 before.
mountkidd wrote:I was running
)
OK that is interesting, i notice you are running kernel 4.10 and i'm back at 4.4. I may be some time...
I knew this wasn't making
)
I knew this wasn't making sense.
When you try to take short-cuts in linux, the penguins faithfully and obediently do what you ask.
If you ask them to change to the first directory matching "cd *.40" and install the driver in that directory - they will....
They won't say but we are at 17.30 this is 16.40 are you sure ? (i'd left that banana skin from a year ago)
Thanks @mountkidd, now i see the same 4-5% improvement.
First thing noticed with clinfo - the number of compute units has risen from 14 to 36. This is probably what is exciting the mining community... may try running x4 / x5 / x6 tomorrow.
AgentB wrote:may try running
)
x5 now seems about optimal, another 2% improvement on x3
clinfo reports only about 0.6 GB of free memory at x6 so no point in going further.