search

fireice-uk / xmr-stak

Free Monero RandomX Miner and unified CryptoNight miner
GNU General Public License v3.0
4.05k stars 1.79k forks source link

Graft - really slow hashrates on 2.5.2 #2035

Open rhlug opened 5 years ago

rhlug commented 5 years ago

OLD

# ./xmr-stak -v
Version: xmr-stak 2.5.0 9fe30b2

HASHRATE REPORT - AMD
| ID |    10s |    60s |    15m | ID |    10s |    60s |    15m |
|  0 |  881.7 |  829.7 |   (na) |  1 |  881.7 |  921.9 |   (na) |
Totals (AMD):  1763.5 1751.6    0.0 H/s
-----------------------------------------------------------------
Totals (ALL):   1763.5 1751.6    0.0 H/s
Highest:  1795.3 H/s
-----------------------------------------------------------------

NEW

# ./xmr-stak -v
Version: xmr-stak 2.5.2 752fd1e

HASHRATE REPORT - AMD
| ID |    10s |    60s |    15m | ID |    10s |    60s |    15m |
|  0 |  520.4 |  509.9 |   (na) |  1 |  520.4 |  539.8 |   (na) |
Totals (AMD):  1040.9 1049.7    0.0 H/s
-----------------------------------------------------------------
Totals (ALL):   1040.9 1049.7    0.0 H/s
Highest:  1078.9 H/s
-----------------------------------------------------------------

Compiled against same rocm 1.9.x

# ldd /opt/xmr-stak-2.5.[02]/bin/libxmrstak_opencl_backend.so | grep OpenCL
    libOpenCL.so.1 => /opt/rocm/opencl/lib/x86_64/libOpenCL.so.1 (0x00007f8f7a895000)
    libOpenCL.so.1 => /opt/rocm/opencl/lib/x86_64/libOpenCL.so.1 (0x00007f6cd7e9f000)

Is the Graft CNv2 that much slower to prove or is there something else going on here?

Spudz76 commented 5 years ago

CN2v2 requires retuning, from scratch, from stock. Delete configs, regen, don't rely on anything you used to know (favorite harmonic settings) it's all different.

Look for good settings from the Monero users they forked a couple weeks ago. Defaults definitely aren't great. You will probably never hit the same speed, but some GPU types can, while some CPU types actually go faster at the CN2v2 vs CN1

psychocrypt commented 5 years ago

Thats strange. Do you used the same config for bith versions? Both tests run with rocm 1.9.X?

If I remember correct there are no changes in the amd part which can create a drastical decrease in the hashrate.

psychocrypt commented 5 years ago

Ahh I missed that you are mining craft. Yes it is now using cryptonight_v8 amd this algo is much slower. Never the less one user here get nearly 2k out of his vega with rocm (version from the rocm dev branch) Try to play with the work_size (8 or 16 should be good) and play around with unroll.

rhlug commented 5 years ago

Cool, I will mess around a bit. Just thought maybe it was something simple. I will return with my findings.

gurupras commented 5 years ago

I'm also impacted by this issue. I tried all combinations of work_size (8/16) x unroll (1-8). I ended up with the best settings @ work_size: 16 and unroll: 1. Even at these settings, I was only getting ~1.5k with a Vega56. Is this normal? I used to get ~1.9k+ on the same card