SILENTARMY v5: Zcash miner, 115 sol/s on R9 Nano, 70 sol/s on GTX 1070

[zawawa]

The murderous 8 hour Black Friday shopping is finally over... It's time for coding!

Great! I switched to windows so I guess I can provide tests if necessary.

That would be wonderful. As I became more familiar with SA's code, I can now see a rather glaring problem in the current implementation.
This is a huge bottleneck both for AMD and NVIDIA, guys. I don't think it's impossible to catch up with Claymore once it's fixed.
My wife is working today, so hopefully I can get some results in today.

[1714614336]

1714614336

[1714614336]

1714614336

[1714614336]

1714614336

[1714614336]

1714614336

[qwep1]

The murderous 8 hour Black Friday shopping is finally over... It's time for coding!

Great! I switched to windows so I guess I can provide tests if necessary.

That would be wonderful. As I became more familiar with SA's code, I can now see a rather glaring problem in the current implementation.
This is a huge bottleneck both for AMD and NVIDIA, guys. I don't think it's impossible to catch up with Claymore once it's fixed.
My wife is working today, so hopefully I can get some results in today.

+1

[laik2]

The murderous 8 hour Black Friday shopping is finally over... It's time for coding!

Great! I switched to windows so I guess I can provide tests if necessary.

That would be wonderful. As I became more familiar with SA's code, I can now see a rather glaring problem in the current implementation.
This is a huge bottleneck both for AMD and NVIDIA, guys. I don't think it's impossible to catch up with Claymore once it's fixed.
My wife is working today, so hopefully I can get some results in today.

Depending on Timezone "today" isn't constant value ( 9pm here)

[zawawa]

It's 11 a.m. here in California, so you never know
The actual work doesn't look that easy, though. We will see.

[ioglnx]

Good zawawa.
I finally managed to bring your code into my TFS for autobuilds.So it syncs with every new commit and builds.
I splitted for AMD and NVIDIA meanwhile due to different OpenCL libs.

Meanwhile you figure out the bottleneck..guess the vector calculations :-D I will do my workout and pump. Good luck

[zawawa]

Good stuff, good stuff.

Meanwhile you figure out the bottleneck..guess the vector calculations :-D

I would think so too, but it was actually algorithmic. What a surprise after days of head-scratching...

[laik2]

Good stuff, good stuff.

Meanwhile you figure out the bottleneck..guess the vector calculations :-D

I would think so too, but it was actually algorithmic. What a surprise after days of head-scratching...

Anxious to see what your genius has came up with

[nerdralph]

This is indeed a major rewrite. Now I'm convinced I can do this, but it's very time-consuming. We will see.

I think I can modify ht_store to get performance on par with Optiminer.  This would be straight OpenCL.  If I have enough time I should have something ready to test tomorrow.

The OpenCL compiler isn't behaving as I would like, so no luck with the relatively simple optimization.
https://bitcointalk.org/index.php?topic=1679855.msg17000600#msg17000600

[nerdralph]

Good stuff, good stuff.

Meanwhile you figure out the bottleneck..guess the vector calculations :-D

I would think so too, but it was actually algorithmic. What a surprise after days of head-scratching...

In my most recent discussions with Marc, he says the atomic_add in ht_store is still a problem.  Although the counter table is small enough to fit in L2 cache, he says he's seeing a 60% miss rate.  The L2 cache write-back must be lazier than we want, so if the cache line for the slot can be flushed at the end of ht_store the hit rate should be improved.

[Biodom]

Thanks, guys...zec miners at the linux fort are anxiously awaiting for the cavalry to arrive just in time to save the fort from the v8 "hordes" 

[giagge]

It's 11 a.m. here in California, so you never know
The actual work doesn't look that easy, though. We will see.

Great zawawa , boost my Nvidia gtx 1070 pls .

[zawawa]

Good stuff, good stuff.

Meanwhile you figure out the bottleneck..guess the vector calculations :-D

I would think so too, but it was actually algorithmic. What a surprise after days of head-scratching...

In my most recent discussions with Marc, he says the atomic_add in ht_store is still a problem.  Although the counter table is small enough to fit in L2 cache, he says he's seeing a 60% miss rate.  The L2 cache write-back must be lazier than we want, so if the cache line for the slot can be flushed at the end of ht_store the hit rate should be improved.

Yeah, that function is a total b*tch... I also noticed those counters are slowing things down considerably. I don't know if we can have that kind of precise control over the L2 cache, though. I suspect the root cause is at a higher level. Luckily, I still have 8 hours my time before the end of the day  We shall see.

[xeridea]

Optiminer builds no longer available due to people circumventing the absurd devfee for subpar miner, lolz.

[laik2]

https://forum.z.cash/t/about-linux-dev-fees/9600/37

Someone read my post obviously

[xeridea]

https://forum.z.cash/t/about-linux-dev-fees/9600/37

Someone read my post obviously

Nah I just happened to check on progress, and saw notice on the Github page.  I would have no interest in removing a devfee, I just find it interesting Optiminer attempted an absurd devfee, on a miner that is not the fastest, which incentivised people to remove it, to get speeds similar to the fastest.

[laik2]

https://forum.z.cash/t/about-linux-dev-fees/9600/37

Someone read my post obviously

Nah I just happened to check on progress, and saw notice on the Github page.  I would have no interest in removing a devfee, I just find it interesting Optiminer attempted an absurd devfee, on a miner that is not the fastest, which incentivised people to remove it, to get speeds similar to the fastest.

Well...I would have used it even with 20% fee IF the owner guaranteed stability. There was no such guarantee, so even 1% was too much. Stupidity is hardcoded into human nature
I actually coudn't believe that someone is trying to sell stupid proxy and wasted numerous hours to secure something so stupid that could be bypassed by simple iptables ...oh what a day...

[zawawa]

Alright, coding is done. I need to tweak parameters quite a bit to get optimal performance, though.
I will keep you guys updated.

[Amph]

Optiminer builds no longer available due to people circumventing the absurd devfee for subpar miner, lolz.

i guess this is not possible with clymore right? so it was something that optiminer did wrong with his fee in his build

Alright, coding is done. I need to tweak parameters quite a bit to get optimal performance, though.
I will keep you guys updated.

200 sol per 1070 coming?

[reb0rn21]

sadly crappy 280x can do 235sol/s at 1200/1600

[xeridea]

Optiminer builds no longer available due to people circumventing the absurd devfee for subpar miner, lolz.

i guess this is not possible with clymore right? so it was something that optiminer did wrong with his fee in his build

Alright, coding is done. I need to tweak parameters quite a bit to get optimal performance, though.
I will keep you guys updated.

200 sol per 1070 coming?

Not sure, though he does have countermeasures, and it slows down ~5% (same as -nofee option) if you try to avoid it.  But 2.5% for a top notch miner is a lot more reasonable than 10-15% on a subpar miner.