neoscrypt miner performance analysis

bluebox

— minor update/edits 1/8/18, 1/21/18 —

I’ve been mining UFO exclusively for the past 4 months, but here’s a miner test that applies to all neoscript coins, at least on CentOS 7 with NVidia GPUs. Four versions of ccminer that I know about, along with the new nsgminer 0.9.4.

This is the pc I use for all of my mining. I had everything but the GPU’s leftover from earlier builds for my kid’s pc’s or my gaming box. I chose the specific EVGA 1060’s because they were cheeeep (~$210/ea), have some OC capability, and can be modded to remove the fan cover easily and used with risers for maximum cooling. Yes, removing the cover additionally reduces temps in spite of already being separated on risers — these models have fans mounted directly to the heatsink, and are separate from the cover. Not all GPU’s can have their covers removed — some EVGA’s are fixed to the board and can’t be separated from the heatsink (like my old 970’s)…

4 EVGA GTX1060 3GB part # 03G-P4-6160-KR
ASRock H97 Anniversary, Pentium G3220 3GHz, 4GB RAM, small SSD boot
Seasonic Prime Titanium 750W

All miners were freshly compiled on CentOS 7.4.1708 kernel 3.10.0-693.11.1.el7.x86_64 with gcc 4.8.5, individually with CUDA 7.5.18, 8.0.61_375.26, and 9.0.103_384.59. I didn’t feed them any special flags; they were all compiled using the given ./build.sh or autogen-make process. I could not test 9.1 as there isn’t a compatible driver for 1060’s yet. Also left behind v6.5 as it is not compatible with Pascal cards.

nvidia-xconfig was used to add the gpu’s to the xorg.conf, then coolbits was added so the fan and gpu clock can be changed. Option “Coolbits” “12” is valid for these 1060’s.

NVidia Settings control panel was then used to change PowerMizer Settings to ~~Prefer Maximum Performance Mode~~ Auto***, Graphics Clock was increased to +200, and fans were set to 90%. This maximizes performance on these cards (P2 mode; P0 cannot be set) while keeping temps below 70°C to avoid clock throttling.

Each miner was run for between 5 and 10 minutes until reported hash rates leveled off. NSGMiner was run longer, as it does not report rates the same way; I had to jot down the combined rates shown and average them, but they’re still only a ballpark estimate.

So, to the numbers (kH/sec):

ccminer-djm34
cuda 7.5: 3080
cuda 8.0: 3088
cuda 9.0: **
cuda 9.1: 3130 1/21/18 †

ccminer-KlausT
cuda 7.5: *
cuda 8.0: 2708
cuda 9.0: 2890

ccminer-ghostlander
cuda 7.5: 2818
cuda 8.0: 2841
cuda 9.0: 2862

ccminer-tpruvot
cuda 7.5: 2525
cuda 8.0: 2475
cuda 9.0: **

nsgminer v0.9.4
cuda 7.5: ~2100
cuda 8.0: ~2100
cuda 9.0: **

* compute 61 not supported - this is one of the newest ccminer builds, so it has newer flags
** compute 20 not supported - usually the scrypt algo chokes the compile, I wasn’t about to try stripping it out of the build
*** 1/8/18 I found that for whatever reason I was “losing” a gpu (or two) every 6-12 hours when I went back to full-time mining. Perhaps it was the updated nvidia driver version, which I neglected to mention — 384.98. Setting it back to Auto (default setting) eliminated the problem; system has been running stable 24/7 since then. This little bugger just ran for months on end last year when I put it together in the spring. My Windows 7/10 dual GTX970 box never, and I mean never, could do this. The only troubles I’ve had are if the pool/stratum hiccups or the house power goes out.
†Finally got the time to install compatible 1060 drivers with cuda9.1, and modified the makefiles to remove compute_20 dependency so the whole enchilada would compile. My first run was on 9.1 and it pulled yet another 10-15K per card over 8.0. The best gets better. :)

djm34 is still the king, even though it hasn’t been updated in 2 years. KlausT and ghostlander’s are a close second on CUDA 9, with tpruvot lagging and having old code incompatible with CUDA 9. NSGMiner was far behind; even though I report the combined hash rate for 4 cards, the individual card hash rates showed lower numbers in the 400-500kH/s range with occasional peaks to 700kH/sec.
(Edit 1/21: I need to play with intensity more with nsgminer; I’m not convinced the default maximizes things fully.)

What all of this shows is that your mileage may vary, since I didn’t try different compilers (intel, portland group, etc.), different gcc versions, different cards (too expensive!), or different kernels (ubuntu, fedora, etc.). This is just a snapshot, keeping many things the same. Hope this helps.

jimmy24651

I have a question…
rookie question? yes…
When I click on the link for the DJM miner above, I am taken to the GitHub page for it, obviously, but what do I click on to actually download the miner? I am currently running CCMiner 8.14, and I am getting great results, but a little more definitely wouldn’t hurt!

wrapper

@bluebox said in neoscrypt miner performance analysis:

All miners were freshly compiled on CentOS

Thanks for the update.

bluebox

@jimmy24651

In gihub, any binaries the dev makes are listed in the Releases link at the top. His latest (2 years old this month) includes all the binaries in one zip file rather than individual .exe’s (I’m assuming you’re running windoze…).

If you want to compile, you can either use git to clone the repo (copy locally) or download the zip or tar.gz’d source in the Releases page.