Playing with amdmemorytweak

[Soulkeeper]

Have any of you been experimenting with this tool ?
I've been trying different settings on my vegaFE

# default @ 945 hbm2
./amdmemtweak --i 0 --CL 17 --RAS 33 --RCDRD 16 --RCDWR 10 --RC 47 --RP 14 --RRDS 4 --RRDL 6 --RTP 5 --FAW 16 --CWL 6 --WTRS 4 --WTRL 9 --WR 16 --WRRD 1 --RDWR 17 --REF 3900 --RFC 350

# stable for me @ 988
./amdmemtweak --i 0 --CL 13 --RAS 30 --RCDRD 13 --RCDWR 5 --RC 44 --RP 13 --RRDS 3 --RRDL 4 --RTP 5 --FAW 16 --CWL 6 --WTRS 4 --WTRL 9 --WR 16 --WRRD 1 --RDWR 17 --REF 19500 --RFC 272

I was hoping looser timings might enable higher clocks to become stable, but I havn't observed this. Clocks over 1000MHz just don't seem 24/7 reliable nomatter what.

What results have you guys seen ?

original thread I found with discussion:

AMD Memory Tweak - Read/Modify Timings on the fly!

Hey all, My first post here, thought i'd make it a good one ;-) I have done a lot of debating with myself whether i would share this. Well, i decided...

forums.guru3d.com

Where to get it:

Eliovp/amdmemorytweak

Read and modify memory timings on the fly. Contribute to Eliovp/amdmemorytweak development by creating an account on GitHub.

github.com

 

[Soulkeeper]

The CL setting don't seem to have any effect. I tried 0 to 22 and benchmark results were within 1-2% of each other. The other settings seem to have an effect. I figure the fast timings net about a 3% performance gain compared to the stock timings.

 

[Micrornd]

FWIW -
I really think the HBM is at a point (stock) where it is at (or very close to) it's best, relative to what the Vega (1 and 2) cores can use.
I haven't used that tool, but on my Radeon VII anything over 1050 (1000 stock) doesn't really present any usable gains.
I can push memory to over 1300 on my Radeon VII, but the real world gains just don't appear.
The only real useable gains have been with core speed.

My definitions
Usable=in-game performance
non-usable=benchmark e-peen

 

[Soulkeeper]

FWIW -
I really think the HBM is at a point (stock) where it is at (or very close to) it's best, relative to what the Vega (1 and 2) cores can use.
I haven't used that tool, but on my Radeon VII anything over 1050 (1000 stock) doesn't really present any usable gains.
I can push memory to over 1300 on my Radeon VII, but the real world gains just don't appear.
The only real useable gains have been with core speed.

My definitions
Usable=in-game performance
non-usable=benchmark e-peen

Well, sure. It does help that your VII has about twice the mem bandwidth at stock compared to vega10. My main thing is to just toy with different hardware settings and try what's possible. Gaming is plenty fast even with my card underclocked 30% ...
I'm running opencl distributed computing stuff, so 3% is 3% for what that's worth.
New tools are always fun.

 

[Shmee]

Cool stuff. I may have to try it on the rig with my Vega 64 Nitro+