I don't know if it has always been this way and I just never noticed...
I woke up this morning and noticed that across all of my rigs there was a noticeable drop in hashrate of about 1+ Mh/s per rig.
All AMD cards -- 280X's & 370's.
Anyone else notice this?
1 ·
Comments
I should also add -- my observations are not based on pool reports, but actual calculated hashrate output by ethminer.
I just find it odd that such a noticeable drop occurred with one DAG change. In fact, now that I think about it more, is it really possible that it dropped this much every DAG change for the past 6 months?
How many DAGs have there been since day 1? Every 30k blocks, right?
If so, that's ~43 DAG files.
~1 Mh/s drop across a 6 card rig last night is about ~0.1666 Mh/s per card.
0.1666 * 43 = 7.1666 Mh/s loss per card since day 1? That seems high.
280X's started out hashing at ~25+ Mh/s and are now down to 20-21 -- not 18-19 which is where that math would put them.
So yeah, something seems to be slightly up.
Bueller?
Use same size for chunk and max to simulate current ethminer DAG implementation. You also need all the setx / export stuff ethminer requires these days. Batch files may not be fully up to date, I made this months ago.
Shall we apply "CAGR" to this? At least this is much more predictable than mining difficulty or ETH pricing.
But I did not run benchmark command. Just noticed when I checked window to see if everything is running well.
He may be using Qtminer, which pretty much every freakin' time you stop it and restart it, it regenerates "a" DAG--not sure if it's the genesis DAG or the next one up, though.
Hope this sheds some light ^_^
First time I have ever seen it after a week of mining.
It's the output of my DAGGER simulator. For soem reason it won't run on my 970 any longer, but that one kind keeps a steady 20MH/s until 2GB and then it takes a steep nose dive. Kind of like the 780 (old pic taken from thread):
Interesting graphs. When you post 7950, do you mean 2gb or 3gb cards?
If the latter, then why would 3gb cards be so negatively affected starting at 1.5 gb DAG.
If the former, then I suggest to make a clarification.
For normal GPU operations that TLB is large enough to cache most requested addresses, but the problem with ethash is that it does pseudo-random lookups accross a very large memory range. As the size of the DAG increases, the TLB runs full quicker, causing more memory lookups to have to "replayed". This is called TLB trashing.
As you can see both AMD and Nvidia cards suffer from it, but in a different way. Between Compute and GCN versions there's also a difference. I haven't seen too many results of people with GCN 1.1 and 1.2, would be interesting to see.
I am still a bit unclear about which graph to follow for 7950-I assume upper and not lower. However, in this case simulation is not exactly correct because so far i did not observe any difference in MSI HD7950 card (at steady 19.3Mh/s) during the last 1.5 mo while on the graph it should be decreasing steadily on upper picture. If it is the lower pict, then it is different, of course.
At https://forum.ethereum.org/discussion/3947/dagger-simulator
one guy (chinhdn) had done 7950 stimulation and he has nothing going wrong at 256mb chunks until DAG size 2560 mb.
Does the single chunk alloc command in Linux:
export GPU_MAX_ALLOC_PERCENT=100
export GPU_SINGLE_ALLOC_PERCENT=100
that people use now basically negates the problem or exacerbates it?
Thanks for the prior explanation; could you comment further?
I would appreciate it very much.
The chunking strategy I had been working on was bugged. I fixed it the other day and it has a negative impact on hashrate. So it's quite useless even more so now people have found the right environment vars to set.
I can't imagine it's just the dag, but I cant regain the lost hashing power.