Yes you are right, the case is designed in a slightly bad way for that bottom card! but believe it or not its running at around 58 degrees so it's not too hot at the moment
Fans are all at the Max and I got a Noctua Fan in front of the Gpu's i don't know if you spotted it :-) it's running at 3000 RPM - Plus and to be honest the system is not too load cant here it from the basement
-Windows 10 -crimson 15.12 driver -Asus H170 pro gaming -4*MSI R9 390X -8 Go OC msi after burner
Up to 144 H/s :-)
I see you don't have the option to undervolt them ? Msi afterburner is not appearing the voltage control in your picture,, it's strange as you have msi cards ... So you get around 33.5mhz / card Why someone previous said 40mhz / card with gigabyte 390x ?
Are you promoting Gigabyte or what ? Your hashrate can't be real ! check again and confirm please
Dwarfpool, and probably other pools as well, don't show the correct hashrate all the time. They sometimes report 50%-100% higher hashrates than normal. So, a single snapshot from Dwarfpool isn't really accurate, and @highwalker may simply be misunderstanding the readings.
The hashrate counter of ethminer isn't very sophisticated. When a kernel execution finishes between the "farm-recheck" time (regardless of when that kernel started running), the amount of nonces hashed in that kernel run (== --cl-local-work * --cl-global-work) are added to a total and after the farm-recheck period finishes, it gets divided by farm-recheck and you have your "hash rate". Then the total gets reset. Because you have a relatively large work size and a short recheck, most of the time you have two kernel invocations finish between your recheck period, but sometimes just 1. Looking at the screenshot, you have 23 samples with 18 averaging about 39MH and 5 at about 20.5. That's 18 * 39 + 5 * 20.5 = 804,5 / 23 = 35MH/s. You can verify that by either lowering your global work or (temporarily) increase farm-recheck to like 3000 or so.
do --cl-local-work --cl-global-work affect hashrate? Im getting 14.5 on my gtx680 and 20.5 on 280x, without the additional settings or clocking, is it worth it to try to modify them?
@newkidONdablock in most cases the defaults are fine. Ramping them up will show higher hashrates but also a less stable one. Local work always needs to be a multiple of 8, in my fork it should be a power of 2. Global work is like "intensity" seen in other miners where 2 ^ intensity ~ global-work. If you increase it above the level where the GPU is saturated, the only effect will be a longer kernel runtime.
You can sort of compare it to the entrance of a big venue or a festival or something, where each person going in has to go through one of these rotating barrier gates where only person (1 GPU thread, 1 nonce) can pass through at a time. Local work is the amount of gates you place side by side. Global work is the queue in front of each gate. When everybody is inside, your kernel finishes. Of course there are limitations to amount of gates you can place side by side. And there are ideal sizes as well, as people going though the gates appear to be holding hands while they cross the entrance. That's 64 people on AMD (a wavefront) and 32 on CUDA (a warp). So it makes sense to use multiples of 32/64 for local work. Inside the ethash kernel, groups of 8 people (threads) are actively exchanging information during the kernel run, hence the multiples of 8 restriction. I could go on about the real limiting factor here which is register usage, but I don't have a festival gate analogy at hand
The hashrate counter of ethminer isn't very sophisticated. When a kernel execution finishes between the "farm-recheck" time (regardless of when that kernel started running), the amount of nonces hashed in that kernel run (== --cl-local-work * --cl-global-work) are added to a total and after the farm-recheck period finishes, it gets divided by farm-recheck and you have your "hash rate". Then the total gets reset. Because you have a relatively large work size and a short recheck, most of the time you have two kernel invocations finish between your recheck period, but sometimes just 1. Looking at the screenshot, you have 23 samples with 18 averaging about 39MH and 5 at about 20.5. That's 18 * 39 + 5 * 20.5 = 804,5 / 23 = 35MH/s. You can verify that by either lowering your global work or (temporarily) increase farm-recheck to like 3000 or so.
Well like i said before. This was a test with local work @ 16k and global @ 256. When i lower that to 8192 and 128 it will be stable and around 39. As i said before. There will be no fluctuations whatsoever.
I took the screenshot at a wrong moment. Again, i will redo it. because most here need to see it with their own eyes to believe it. I would too..
Thank you for your detailed explanation none the less. Appreciate it!
@newkidONdablock in most cases the defaults are fine. Ramping them up will show higher hashrates but also a less stable one. Local work always needs to be a multiple of 8, in my fork it should be a power of 2. Global work is like "intensity" seen in other miners where 2 ^ intensity ~ global-work. If you increase it above the level where the GPU is saturated, the only effect will be a longer kernel runtime.
You can sort of compare it to the entrance of a big venue or a festival or something, where each person going in has to go through one of these rotating barrier gates where only person (1 GPU thread, 1 nonce) can pass through at a time. Local work is the amount of gates you place side by side. Global work is the queue in front of each gate. When everybody is inside, your kernel finishes. Of course there are limitations to amount of gates you can place side by side. And there are ideal sizes as well, as people going though the gates appear to be holding hands while they cross the entrance. That's 64 people on AMD (a wavefront) and 32 on CUDA (a warp). So it makes sense to use multiples of 32/64 for local work. Inside the ethash kernel, groups of 8 people (threads) are actively exchanging information during the kernel run, hence the multiples of 8 restriction. I could go on about the real limiting factor here which is register usage, but I don't have a festival gate analogy at hand
I get a whole extra MH/s with VERY stable results on 280x using --cl-local-work 64 --cl-global-work 12288 --farm-recheck 200 versus default mining.
When I use --cl-global-work 16384, my cards crash like crazy. Even with default clock and relatively high voltage 1000/1500 @ 1.093v for 280x. 8192 gives invalid shares but can run the cards forever.
I would love to try --cl-local-work 64 --cl-global-work 12288
What good does it do me if i would lie about it? Yeah.. none..
41MHS is your peak hash rate. In the other screen it dropped down to 19MHS. Your average hash rate would be in the 30's Instead of ethminer look at the average effective hash rate from the pool.
how to tweak the best for xfx r9 380 4g? and hd7950? cl-local-work? --cl-global-work? and tweaking use afterburner? less energy? more ram or higher gpu?
how to tweak the best for xfx r9 380 4g? and hd7950? cl-local-work? --cl-global-work? and tweaking use afterburner? less energy? more ram or higher gpu?
Try combinations of cl work first. Then keep ethminer running and play with afterburner. Both GPU nad Ram will affect the hash rate. To use less energy you have to underclock and undervolt.
Comments
min/mean/max: 43108900/43283117/43486912 H/s
Win 7 x64 / MSI AMD Radeon R9 270X Gaming 2G
min/mean/max: 17354176/17816444/18169260 H/s
-Windows 10
-crimson 15.12 driver
-Asus H170 pro gaming
-4*MSI R9 390X
-8 Go
OC msi after burner
Up to 144 H/s :-)
is it overclocked and how much (270x)?
Msi afterburner is not appearing the voltage control in your picture,, it's strange as you have msi cards ...
So you get around 33.5mhz / card
Why someone previous said 40mhz / card with gigabyte 390x ?
Your hashrate can't be real !
check again and confirm please
My 390 NItro does 40Mh as well, but core clock to 1260.
I have a screenshot of that.
http://s29.postimg.org/iixxgyo5j/Schermafbeelding_2016_03_21_om_23_13_31.png
Voila
If you know that 1150 gets you to 32-33 easily, than 1260 will get you to 40.
Common sense....
But because it's so hard to believe here i'll do a rerun and i'll take another screenshot.
It's a Sapphire R9 390 Nitro.
The hashrate counter of ethminer isn't very sophisticated. When a kernel execution finishes between the "farm-recheck" time (regardless of when that kernel started running), the amount of nonces hashed in that kernel run (== --cl-local-work * --cl-global-work) are added to a total and after the farm-recheck period finishes, it gets divided by farm-recheck and you have your "hash rate". Then the total gets reset. Because you have a relatively large work size and a short recheck, most of the time you have two kernel invocations finish between your recheck period, but sometimes just 1. Looking at the screenshot, you have 23 samples with 18 averaging about 39MH and 5 at about 20.5. That's 18 * 39 + 5 * 20.5 = 804,5 / 23 = 35MH/s. You can verify that by either lowering your global work or (temporarily) increase farm-recheck to like 3000 or so.
You can sort of compare it to the entrance of a big venue or a festival or something, where each person going in has to go through one of these rotating barrier gates where only person (1 GPU thread, 1 nonce) can pass through at a time. Local work is the amount of gates you place side by side. Global work is the queue in front of each gate. When everybody is inside, your kernel finishes. Of course there are limitations to amount of gates you can place side by side. And there are ideal sizes as well, as people going though the gates appear to be holding hands while they cross the entrance. That's 64 people on AMD (a wavefront) and 32 on CUDA (a warp). So it makes sense to use multiples of 32/64 for local work. Inside the ethash kernel, groups of 8 people (threads) are actively exchanging information during the kernel run, hence the multiples of 8 restriction. I could go on about the real limiting factor here which is register usage, but I don't have a festival gate analogy at hand
I took the screenshot at a wrong moment. Again, i will redo it. because most here need to see it with their own eyes to believe it. I would too..
Thank you for your detailed explanation none the less. Appreciate it!
Core clock 1280, memclock 1500, +100mV Core. As you can see..
Picture. https://drive.google.com/file/d/0B_cBBoZ_PTwOa0JXQ0hlZ3BMZjM1bkxQeDFKWV9KakFIMW5z/view?pref=2&pli=1
What good does it do me if i would lie about it? Yeah.. none..
I would love to try --cl-local-work 64 --cl-global-work 12288
41MHS is your peak hash rate. In the other screen it dropped down to 19MHS.
Your average hash rate would be in the 30's Instead of ethminer look at the average effective hash rate from the pool.
Both GPU nad Ram will affect the hash rate. To use less energy you have to underclock and undervolt.