I have one 5770 card (default 850/1200 clocks) on which I was trying to mine litecoin in my home computer and then moved it to a HP server in datacenter, however there I can only reach about 60% of performance, but don't know why.
At home, I had Windows 7 64bit, 4 GB RAM, cgminer 3.1.0, latest AMD driver and SDK and card in x16 PCIe 2.0 slot. With TC 3968 (more TC was not working) and intensity 18, I was getting about 200 kHash/s which is good for this card.
On server, I have 64bit debian linux 6.0 (squeeze), but only 3 GB RAM and the PCIe slot is only version 1.0 (x8 using a riser). Cgminer is also 3.1.0, driver 12.10, SDK 2.8 (I followed this guide
http://www.scribd.com/doc/132505278/Debian-GPU-Litecoin-cgminer-install-guide ).
I was trying various TC and other options, intensities 12-20, but can't get more than 117 kHash/s. What was different is that I was able to compile bin with more TC (even more than 5000), but it had no effect on performance. Even OC didn't have any effect. Bin files for Windows had "l4" on end, those on linux have "l8". From documentation it is "size of long". Does it matter in something? Can for example linux cgminer use bin file prepared on windows version or is it incompatible?
What could be causing this?
1) Maybe is the PCIe 1.0 the bottleneck? How does scrypt mining work? Does cgminer transfer huge amounts of data to GPU memory? Or mainly the GPU itself does majority of GPU memory accesses?
2) Or only 3 GB of system RAM? But cgminer is using only 300 MB of RAM, which should be OK for just 200 KHash/s... I have read that scrypt mining needs a lot of system RAM. How does it use it? For some pre-computations or does it transfer the data to GPU memory?
3) Or is it just bad combination of driver/SDK versions?
BTW bitcoin performance is OK on the linux server (about 195 MHash/s), only scrypt is affected.