Long time already I did not collect statistical performance data.
But last release of v8 and allowance of VLAR GUPPI (GBT data) on GPU requires some refreshement of those pictures.
While data collecting let's recall what we had before, in ~2011-2012.

This picture shows all collected data.
Immediately some misconception appears. That CPU process VLAR at about same speed as non-VLAR. No, data definitely show (and rising curve at lowAR confirms) that even opt CPU app processes VLAR at least twice (actually more) longer than VHAR.
So, VLAR is hard. Even for CPU.

I have asked legitimate questions...
Why is offended someone here and responds aggressively?

Cause while asking you made illegitimate comparison with other projects. If I provide some options for "perfectionists" it doesn't mean other soft can't be improved similarly - just nobody there cared about that. Try to go there and ask to implement tuning options....

The people who know me, know me - I'm a perfectionist.
I would like to squeeze out the max performance.

So do that, you have "the buttons" to play with.

For the following GPUs (1 or 2 WUs/GPU?), which GPU apps (in Lunatics Installer or d/l URLs of stock apps?) and which cmdline settings I should use?

You tell us? How is it possible to know what is better for your setup if you don't carry benches and just asking others for best tuning line?

All that looks strange to me. You have tunable app, you have benchmark package, you have readmes... And then all questions are "what better tuning line".
What exactly your problems with KWSN bench?

EDIT:

-tune N Mx My Mz
-oclfft_tune_gr N
-oclfft_tune_lr N
-oclfft_tune_wg N
-oclfft_tune_ls N
-oclfft_tune_bn N
-oclfft_tune_cw N

Try to find table I posted with experimental findings what is worked for my own C-60.
Also, maybe someone made smth similar but I not aware. With fast GPU you could make your own table quite fast. As rule of thumb ( should I teach "perfectionist" how to make tuning?? ) one should take best recommended ones (RTFM!) and vary them from those values.
If you "full perfectionist" - then do full parameter space scan. WG for AMD limited to 256. Most of options are fool-proof. Other will show improper input as invalid result (and benchmark will catch that).

It's a pity, sad and annoying that I can't let run SETI@home like other projects (Plug&Play or "just out of the box and fine") for to get max. performance.

What hell make you think you can do that for other projects Or do you think that if there is no options then max possible performance already achieved?
What a perverted way of thinking started to spread recently... Ok, forget about options. Just run stock. It's the best you can do. Just as with other projects.

I'm a semi advanced user.
I know a bit how to make bench test runs.

AstroPulse worked well, MultiBeam not (slightly deeper explanation).

Of course, I read the readme.txt files.
But, there I can't read this which I asked in the above linked message.

E.g. (after my experiences)...

AP bench test runs (on NV GT730):

APbench211_minimal.7z (which have the WU #ap_genwis.dat incl. This kind of file isn't in the MultiBeam tool, because of this always two times superfluous runs)
ap_Zblank_2LC67.wu of ap_Zblank_2LC67and6LC67.7z

In short (something like this, for an entry level, low end VGA card)...
cmdline tests:
-unroll 2 up to 10 (+1)
-ffa_block 256 -ffa_block_fetch 128 up to -ffa_block 2048 -ffa_block_fetch 1024 (+128, rough) (+/-64, fine)
-tune 1 8 4 1 up to -tune 1 64 4 1 (1st N *2)
-tune 2 8 1 1 up to -tune 2 256 1 1 (1st N *2)
-oclFFT_plan 16 8 32 up to -oclFFT_plan 256 32 256 (all possible N's specially written)
-sbs 32 up to -sbs 512 (N *2)
-skip_ffa_precompute (without N's)


Like this above for AP, I would like to have this also for MB.
Until now I have no ideas which N's are possible and if the cmdline settings are alone (e.g. -unroll N) or connected together (e.g. -ffa_block N and -ffa_block_fetch N).

The currently available bench test MB WUs are too small/fast for very fast VGA cards.

For MB is possible (I have no idea which N's are possible and if the cmdline settings are alone or connected together):

-period_iterations_num N
-spike_fft_thresh N
-sbs N
-use_sleep_ex N
-no_caching
-tune N Mx My Mz
-oclfft_tune_gr N
-oclfft_tune_lr N
-oclfft_tune_wg N
-oclfft_tune_ls N
-oclfft_tune_bn N
-oclfft_tune_cw N


Of course -use_sleep is needed for AP and MB NV OpenCL apps.

Also the download possibilities of all available apps (stock are newer than Installer ones).

It's a pity, sad and annoying that I can't let run SETI@home like other projects (Plug&Play or "just out of the box and fine") for to get max. performance.


Until I finshed this AP and MB bench test runs (for to find the best/fastest cmdline settings) I don't let run SETI@home in my BOINCs.


Thanks.

Because of my Message in the SETI-Beta Forum which I wrote for 15 days...

I made something wrong because of the .wisdom file creation?

It's possible to make a MB/AP Bench Test Instruction?

Thanks.