Topic: GPU to hash metrics (Read 3637 times)

The GPU chips in a 5970 are the same as those in a 5870 and will use the same power at a given clock speed and voltage. In addition, a 5970 has an extra PCI bridge chip in it which uses some power, though this is probably not significant.

The most significant thing is heat dissipation - the heatsink and fan in the 5970 has to get rid of the heat from 2 GPU chips and is constrained to be the same size as that in a 5870 which only has to get rid of the heat generated by 1.

According to Wikipedia, http://en.wikipedia.org/wiki/Comparison_of_AMD_graphics_processing_units

The standard clock for 5970s is 725MHz, at which speed the card uses 294W of power.
Standard clock for a 5870 is 850MHz, at which speed each card uses 188W of power, 376W for 2 cards.

In terms of work done, a 5970 offers 725*3200/294=7891 shader.Mhz/Watt and 2 5870s offer 2*1600*850/376=7234 shader.Mhz/Watt. The 5970 is more power efficient at standard clock speeds because it is run at a lower voltage - the standard voltage of the GPUs in a 5970 is 1.05V and for 5870s is 1.15V. To overclock the GPUs in a 5970 to the same level as would be achieved with a 5870 would require the voltage to be set to the the same level - not impossible, but the power use would then be the same.

I overclock my 5870 to 980MHz at standard voltage, at which frequency it uses about 215W of power. If you do this on a 5970 it would use about 430W - not impossible, but cooling it would become a major issue.

I believe this is false, the 5970 can clock just as high as 2 crossfired 5870s. It can also do this using less power.

For a given architecture, work done will be closely proportional to clock speed multiplied by number of stream processors.

Bus width seems to matter very little or maybe not at all for hashing. I performed an experiment by plugging my 5870 via an adaptor into a PCIe x1 mainboard slot. I could not see any difference in the hashing rate, though performance for graphics was noticeably degraded.

Memory speed seems to make no significant difference either. When mining, one might set the memory speed as low as it goes to reduce power use.

My HD 5870 manages about 362000 khash/s at a clock speed of 980MHz, so is giving 226.25 khash/s per stream processor. From this I would estimate the clock speed of your 5970 at 980*175/226.25 = 758MHz.

This is close to the standard speed of 750MHz, so you are probably not overclocking your GPU.

If you want it to go faster turn the clock speed up a bit, but be aware that higher clock speeds mean more power used (may stress your power supply) which means more heat generated (may stress you GPU) which means the cooling fan goes faster (may stress your ears)

If you plan to buy additional GPUs for mining, I would recommend 5870s rather than 5970s as these have a higher standard clock and more headroom for overclocking (assuming you have the motherboard slots for twice as many GPUs)

Are the number of Stream Processors the most effective specification to directly relate a specific GPU to hash/s metrics?

For instance:

My ATI "Redwood" has 400 stream processors and can push about 36,000 khash/s.  (@90 khash/s per stream)

An ATI HD 5970 can put out 560,000 khash/s with 3,200 stream processors (@175 khash/s per stream)

My NVidia Quadra NVS 135M has 8 stream processors (hahah!) and does 950 khash/s  (@~120 khash/s per stream)


Other metrics don't seem to matter that much, i.e. GPU memory and memory speed, GPU clock speed, bus width, etc.