Want legit 7970 testing/benchmarking and tuning for cgminer and Diablominer?

[coretechs]

I finally had the time to install a pair of Gigabyte 7970s last night.

I tried a few of the settings in the other thread and also found the best I could get out of cgminer was around 475Mh/s per GPU.

I downloaded diablo miner, oc'd the cores to 1100mhz on stock voltage and dropped the memory down to 600mhz.  Started up and bam, 1300Mh/s.   It was very late and I had no time to do any tinkering, but 650Mh/s per gpu out of the box on stock voltage with a nice oc is pretty awesome.

[DiabloD3]

Curious about how much room for optimization there is, I made some calculations to see how close Diablominer is to the maximum theoretical performance of the 7970 (see here) and unless I'm missing something in my assumptions I found it to already be within ~2% of the maximum theoretical performance.

That'd be depressing if all I can get is another 2%. I'm trying to get the damned Kernel Analyzer to work in either Win in VM or in Wine, both is a no go, has to have a real running copy of the drivers, and the native Linux version locks up soon as DM calls an a CL function.

Damnit AMD

[DeathAndTaxes]

Curious about how much room for optimization there is, I made some calculations to see how close Diablominer is to the maximum theoretical performance of the 7970 (see here) and unless I'm missing something in my assumptions I found it to already be within ~2% of the maximum theoretical performance.

Your link had a link to https://bitcointalk.org/index?topic=7964.msg550288#msg550288 which server dies on.  Can you find a link to original 3375 ops calculation.

[rjk]

Curious about how much room for optimization there is, I made some calculations to see how close Diablominer is to the maximum theoretical performance of the 7970 (see here) and unless I'm missing something in my assumptions I found it to already be within ~2% of the maximum theoretical performance.

Your link had a link to https://bitcointalk.org/index?topic=7964.msg550288#msg550288 which server dies on.  Can you find a link to original 3375 ops calculation.

Add .php after index.

[1onevvolf]

...

Your link had a link to https://bitcointalk.org/index?topic=7964.msg550288#msg550288 which server dies on.  Can you find a link to original 3375 ops calculation.

Add .php after index.

Thanks, I fixed the link.

[1onevvolf]

That'd be depressing if all I can get is another 2%.

Yeah, it kinda takes the fun out of optimizing if all you can get is 2%.

Any idea if there's a GCN feature that we can exploit for more performance like fixed function hardware or new instructions/amd-specific extensions? I initially thought that using size 16 vectors would help, thinking that the GCN SIMD cores were akin to x64 SIMD registers, but it seems that the card exploits its SIMD cores by running a scalar instruction on 16 threads at once.

I'm trying to get the damned Kernel Analyzer to work in either Win in VM or in Wine, both is a no go, has to have a real running copy of the drivers, and the native Linux version locks up soon as DM calls an a CL function.

Don't even bother until the release a new version because Kernel Analyzer won't even correctly list the kernel stats for tahiti. Running the profiler via command line does give some useful output however, and I have .il, .cl, .isa  and a comma separated value file of a profile run with -v 1 from my 7970 if you're interested.

[DiabloD3]

That'd be depressing if all I can get is another 2%.

Yeah, it kinda takes the fun out of optimizing if all you can get is 2%.

Any idea if there's a GCN feature that we can exploit for more performance like fixed function hardware or new instructions/amd-specific extensions? I initially thought that using size 16 vectors would help, thinking that the GCN SIMD cores were akin to x64 SIMD registers, but it seems that the card exploits its SIMD cores by running a scalar instruction on 16 threads at once.

I'm trying to get the damned Kernel Analyzer to work in either Win in VM or in Wine, both is a no go, has to have a real running copy of the drivers, and the native Linux version locks up soon as DM calls an a CL function.

Don't even bother until the release a new version because Kernel Analyzer won't even correctly list the kernel stats for tahiti. Running the profiler via command line does give some useful output however, and I have .il, .cl, .isa  and a comma separated value file of a profile run with -v 1 from my 7970 if you're interested.

I'm more interested in it for generic 2.6 optimization. I'm hoping I can get 2.6 running as fast as 2.5 on my 5850 and _IF_ I can get it running as fast as 2.1, I can abandon 2.1 support (but I don't see how 2.6 could get that fast)

[mc_lovin]

I find it funny how so many people criticised giving you a 7970 to do diablominer tuning to and now you have delivered a miner that like EVERYONE is using for their own gain.

[DiabloD3]

I find it funny how so many people criticised giving you a 7970 to do diablominer tuning to and now you have delivered a miner that like EVERYONE is using for their own gain.

Yeah, and now cgminer also has the kernel among the ones it supports. I think I delivered on the promise as best as possible, although I'm still looking for new ways of beating the compiler.

To be fair, AMD's GCN compiler (not the same as the 5xxx/68xx/69xx compiler(s), seems to be ground up brand new) is doing a pretty good job, I haven't been able to find any useful optimizations after I did that two week brutal death march of tuning for 2.6.

I think all in all, the community benefited from their investment.