Re: [JCE] Ultrafast CN-Heavy/Tube/HVX miner, low power, Vega56 1750+ h/s

Quote from: JCE-Miner on November 09, 2018, 07:36:17 PM

I'm claiming to be the fastest on old GPU and I don't even provide a decent autoconfig for old cards, what a shame...
I'll update some autoconf for those cards like Cayman, Tahiti, Bonaire or Pitcairn.

Perf on HD6990:
https://github.com/jceminer/cn_gpu_miner/issues/26

This guy shared his config that goes to 270 h/s per GPU (HD6990 is a dual GPU) with no memory strip hack (which is lame, i will advise him to patch his BIOS).

Quote

Settings used : 32 worksize and 864 intensity.

I guess that's on one thread per GPU, the dual threading is for GCN cards.
It translates:

Code:

{ "mode" : "GPU", "worksize" : 32, "alpha" : 64, "beta" : 8, "index" : ...., "multi_hash":864 }

Thanks all for report and sharing your optimal config. Don't forget to tell what algo you're mining Wink

so i can put those config in the Github doc.

Quote

Tuning is just a bit of a hobby of mine

I usualy get complains about the parameters being too numerous (there are four). Not to few.
Greeks above beta are deprecated, worksize is very like in XMRstak, multi_hash is like XMRstak intensity (but unlike SRB intensity) and alpha/beta are internal data packing parameters, no general rule, sometimes doing x2 or /2 on the defaults gives a little boost. Depends on the card and algo you mine. My autoconfig always defaults to 64 and 8.

Ok so I will start from the beginning.
Recently I have reinstalled my 4xHD6990 rig from Win7 to Win10 since JCE miner is the fastest CNv8 miner for those old cards.
Unfortunately for those old cards setting everything up under Win10 is not a easy task. I have failed in 12.5%

First of all use catalyst 14.4 driver which is the latest which can detect full HD6xxx card memory. It does not matter that it is driver for Win8.1. It works great for a single card but for HD6990 it is a little bit tricky. The best way to install the driver is to connect single cards and install the driver not using the installer but choosing the INF file from the driver package. Once this is done plug in the second card. Driver will be installed automatyczny. And so on with all the cards ( like it was with Vega nad Blockchain Compute Driver). With HD6990 you cannot add single cards since there are two cores per card. That lead to problem later on.
After the driver setup we have to disable "EnableUlps". That will gain some stability or, like in my case, start hashing.
And the most important. After the reboot always disable and enable every card in device manager before the miner start. Script will do the trick.

Code:

{ "mode" : "GPU", "worksize" : 32, "alpha" : 64, "beta" : 8, "index" : ...., "multi_hash":864 }

I came to the same settings after some time.
I can squeeze 282h/s per core @930mhz in CNv8. Which is 12,5% better than xmr-stak and xmr-stak does not work properly yet.

Unfortunately the stability of the driver is now the main issue. I can have 7 out of 8 cards working at average 282h/s but that single one... it just does not hash at all. Maybe because is had a monitor connected during driver installation.
Nevertheless the rig is producing 2365h/s vs 2270h/s on Win7. So even without one GPU working I have 4% better results all in all.

This is what I know. If there is some kind of black magik which I can use to start my last gpu, be my gest.

Edit:
Testing Heavy on 4x Vega 64 0.33b6 and it was hashing about 2070h/s but after about 40min a hashrate jump can be noticed. From time to time (every 15-20s) hashrate jumps to 2100+ for about 3-4s and than levels to 2070h/s. After about a minut one card or more drops hashrate to 1985h/s. No jumps any more. Rest of the cards hashing ok.