this is the image phateus produced for vectors and memory clocks
but it lacks 200mhz which seems to be best for my card
Topic: Worksize and GPU memory clock (Read 1939 times)
this is the image phateus produced for vectors and memory clocks
but it lacks 200mhz which seems to be best for my card
For this discussion I'm using HD5750 under Windows 7 with SDK2.1
I'm running fpgaminer's fork of poclbm with phatk2.1 with vectors which seems to be the fastest without 100% CPU usage bug
I've been told that downclocking the GPU clock could actually make the GPU go faster. At first this did not happen at all. Here's why: I've experimentally determined that worksize of 128 was best at stock.
However, I've found a better configuration than stock clock and worksize 128. It's 200Mhz clock (down from 1150) and worksize 256. All lower work sizes don't seem to work as well. Higher clocks also seem a tiny bit slower. Lower clocks are slower as well.
It seems for each worksize a different memory clock is optimal. On worksize 64 I think it's something like 350mhz. I wonder why that is.
So the performance goes like this:
Most random configs get 140mhash/s (some random memory clock and some random work size)
Stock clock and worksize 128 gets about 144
350 mhz and worksize 64 gets about 144.3
200 mhz and worksize 256 gets about 144.5
this basically makes no sense, and I don't want to experiment more because clock speed 400 froze my card so I had to reboot
I'm running fpgaminer's fork of poclbm with phatk2.1 with vectors which seems to be the fastest without 100% CPU usage bug
I've been told that downclocking the GPU clock could actually make the GPU go faster. At first this did not happen at all. Here's why: I've experimentally determined that worksize of 128 was best at stock.
However, I've found a better configuration than stock clock and worksize 128. It's 200Mhz clock (down from 1150) and worksize 256. All lower work sizes don't seem to work as well. Higher clocks also seem a tiny bit slower. Lower clocks are slower as well.
It seems for each worksize a different memory clock is optimal. On worksize 64 I think it's something like 350mhz. I wonder why that is.
So the performance goes like this:
Most random configs get 140mhash/s (some random memory clock and some random work size)
Stock clock and worksize 128 gets about 144
350 mhz and worksize 64 gets about 144.3
200 mhz and worksize 256 gets about 144.5
this basically makes no sense, and I don't want to experiment more because clock speed 400 froze my card so I had to reboot
Jump to: