You might be missing fact that it has 68**/58** series has simd wich perform 5 operation,thats how we get 1120/1600,of the actual count of processors is 224 for 6870 and 320 for 5870, so i cant find any connection to gpu_thread_concurrency...
According to OpenCL File this is size of x,the amount of main threads which would be the amount of threads we run i think...
Topic: LTC mining with Reaper on a 7970 - page 6. (Read 33070 times)
Am I right in thinking a 6870 has 1120 shaders,4 threads so 4480?
Yeah, the 6850 has 960 shaders according to AMDs site.
So 960 * 5 = 4800
I get exactly 250kh/s at aggression 17.
With it at 6144 I get 259kh/s with aggression 17 and 250kh/s with aggression 16.
All 68xx series cards are still vliw5 to my knowledge, which I think means they have a queue of 5 threads per shader. So a 6870 would be 5 * 1120 = 5600
Am I right in thinking a 6870 has 1120 shaders,4 threads so 4480?
Well, that's just what I've been told. And I only have 6950's to test on.
However, that said I believe 6850 is VLIW5, so might want to try 5 x 960 = 4800
However, that said I believe 6850 is VLIW5, so might want to try 5 x 960 = 4800
Funny, I get higher kH/s on my 6850 with 6144 then I do its optimal number (3840).
I think I can answer the gpu_thread_concurrency number. This should be set to [shaders per gpu] multiplied by the [shader "threads" supported].
Examples:
A 6950 has 1408 shaders, and each shader can queue 4 threads, so the optimal number for this card is 5632.
A 6990 has 1536 shaders, and each shader can queue 4 threads, so the optimal number for this card is 6144.
I think the 5xxx cards were 5 threads queued per shader, so there it would be 5 times the shader count.
I do not know how many "threads" each shader on a 7 series can queue.
Examples:
A 6950 has 1408 shaders, and each shader can queue 4 threads, so the optimal number for this card is 5632.
A 6990 has 1536 shaders, and each shader can queue 4 threads, so the optimal number for this card is 6144.
I think the 5xxx cards were 5 threads queued per shader, so there it would be 5 times the shader count.
I do not know how many "threads" each shader on a 7 series can queue.
Anyone else tried this? What are your litecoin.conf settings? Especially if you have a 7970. I'm under Win7 using the newest 64-bit reaper from the wiki. http://wiki.solidcoin.info/wiki/Reaper
I'm getting about 236 kh/s at stock clocks and voltage, and Afterburner shows the GPU at 96% usage with the following settings in litecoin.conf:
worksize 256
aggression 12
threads_per_gpu 4
sharethreads 18
lookup_gap 2
gpu_thread_concurrency 6144
If I increase the aggression above 12 I get lots of stales/invalids and the message "Weird response from server"
I'm wondering what the last three settings do (sharethreads, lookup_gap, gpu_thread_concurrency). I noticed that gpu_thread_concurrency is 3x my 2048 stream processors and tried other multiples of 2048 but that didn't help. I also tried different values for lookup_gap but none were any better.
If I reduce threads_per_gpu below 3, performance drops into the high 220's and gpu usage in afterburner drops to low 90's
I'd also like to know how this is performing on 5850's and 5970's, since I have a bunch more of those.
I'm getting about 236 kh/s at stock clocks and voltage, and Afterburner shows the GPU at 96% usage with the following settings in litecoin.conf:
worksize 256
aggression 12
threads_per_gpu 4
sharethreads 18
lookup_gap 2
gpu_thread_concurrency 6144
If I increase the aggression above 12 I get lots of stales/invalids and the message "Weird response from server"
I'm wondering what the last three settings do (sharethreads, lookup_gap, gpu_thread_concurrency). I noticed that gpu_thread_concurrency is 3x my 2048 stream processors and tried other multiples of 2048 but that didn't help. I also tried different values for lookup_gap but none were any better.
If I reduce threads_per_gpu below 3, performance drops into the high 220's and gpu usage in afterburner drops to low 90's
I'd also like to know how this is performing on 5850's and 5970's, since I have a bunch more of those.
Jump to: