Author

Topic: Novel/Experimental CryptoNight GPU Optimizations; Open Source; AMD+Nvidia (Read 259 times)

newbie
Activity: 10
Merit: 0
Hey there,

just going to point out I’m still working on this, updated with CN/2 and multi algo switching included via xmrig base: https://github.com/949f45ac/xmrig-HIP

There’s interesting CN/2 performance in some cases, and CN/1 I have also been able to make faster. Of course, RX Vega generally works well on Linux now, so this miner has become a bit more useful.

Current exemplary numbers:

Card descriptionCore, mem clkMiner configCN/2 hashrateCN/1 hashrate (TxB, if different)
470 4GB1126, 17508x124 * 2870840 (4x496)
460 2GB1020, 175064x14439442
Vega 641630@905, 110032x64 + 32x6020002180
Vega 561485@900, 92032x56 * 218301910 (32x112)

If you’re interested in trying new stuff, give this one a spin~ I’ll be glad to help out with any issues, and hear about results.

Cheers
newbie
Activity: 10
Merit: 0

Work continues here:
https://github.com/949f45ac/xmrig-HIP
(See first reply itt as well, for current status.)



Original OP:

Hey everyone.

I’ve developed a new miner on the basis of xmr-stak-nvidia. Here is the link: https://github.com/949f45ac/xmr-stak-hip
It is all open source and can run both AMD and Nvidia cards, usually a bit faster than the other open source miners. (I did not test against closed source.)
Building requires a quite specialized setup with ROCm. If you run Nvidia cards, there is a portable binary in the "releases" section. I could upload one for AMD as well, but you’ll still need to setup ROCm to run it, so you may as well build the miner yourself.

As for the hashrates, here is a small table with some test results. ("Previous" = open source CUDA/OpenCL monero7 implementations that come with xmr-stak/xmrig.)
As you can see, the HIP miner does especially well on power-optimized settings.

CardCard settingPrevious T x BPrevious H/sx-s-hip T x Bx-s-hip H/s
1050Ti 4GBpl=52.5W64x3029632x48313
1050Ti 4GBpl=75W64x3030332x48319
460 2GB modsclk=38x4483858x112416
460 2GB modsclk=48x4484058x112419
Vega 56 referencesclk=0 mem+10%8x1928 * 21075-11208x224 * 21170 stable
Vega 56 referencesclk=5 mem+10%8x1928 * 21240-13258x224 * 21304 stable

Notes:
  • "T x B" indicates "threads x blocks" configuration, or "worksize x intensity" in the case of OpenCL.
  • RX 460 had modded BIOS, but memory not overclocked. (Polaris memory is hard to OC on Linux.)
  • OpenCL miner cannot use higher intensity than 448 on the 2 GB RX 460 for some reason.
  • Vega 56 fluctuates quite a bit on OpenCL.
  • Vega 56 numbers for the HIP miner have two processes started. Scheduling two GPU workloads in the same process does not work for some reason. But hashrate for a single 8x448 process is only about 15 H/s less in both cases.
  • All Vega numbers are with Kernel driver patch applied. (Slightly increases both OpenCL and HIP miner speed.)

Dev donation address in the miner is unchanged. If you insist on giving me coins personally, addresses are in the README.
You may also just give a star to the repo or send me some thanks, if you use this miner and like it.

Cheers
Jump to: