Bitcoin Forum

Other => CPU/GPU Bitcoin mining hardware => Topic started by: NeonFlash on September 29, 2011, 09:42:03 AM



Title: Underclocking the Memory Frequency?
Post by: NeonFlash on September 29, 2011, 09:42:03 AM
I read in many places that the Memory Clock Frequency is downclocked to 300 Mhz. What is the reason behind this? Does it help in reducing some kind of overhead?

My assumption is that Memory Clock frequency doesn't play much of a role for a task like Bitcoin Mining. What matters more is the GPU core clock frequency and stream processors.

Also, will it help to reduce the operating temperature of the GPU by decreasing the Memory Clock Frequency?


Title: Re: Underclocking the Memory Frequency?
Post by: dark_silverstar on September 29, 2011, 09:46:40 AM
not only reducing temperature, also reduce wattage as well  ;D


Title: Re: Underclocking the Memory Frequency?
Post by: NeonFlash on September 29, 2011, 09:56:57 AM
Do you mean that decreasing the Memory Clock Frequency will help in reducing the voltage as well besides the operating temperature? Any stats that go along with it as an example?


Title: Re: Underclocking the Memory Frequency?
Post by: DeathAndTaxes on September 29, 2011, 10:18:43 PM
It varies from card to card.  Best thing to do is test yourself.

I have a 3x5970 water cooled rigs.  With memory at full speed temps are 30C over ambient and it pulls 1080W at the wall.  With core overclocked to 900MHz (from 735MHz) and memory underclocked to 200MHz temps are 24C over ambient and it pulls 960 at the wall.

Best thing you can is experiment yourself.

Use a wall meter like Kill-a-watt, and set fan to static speed (say 70%) to record temps. 

Set it to one setting, run miner for 3 hours (for good sample size).  record temp, kWh used, and temp.  Adjust memory clock and repeat.


Title: Re: Underclocking the Memory Frequency?
Post by: ssateneth on October 01, 2011, 01:43:53 AM
In addition to using less power (which means lower electricity bill = more net profit), there are certain memory timing thresholds at certain frequencies, as well as a few peaks of performance relating to worksize. I did an exhaustive amount of benches at various memory speeds. It's a little out of date and I really should run it on current kernels.

https://docs.google.com/spreadsheet/ccc?key=0AjXdY6gpvmJ4dEo4OXhwdTlyeS1Vc1hDWV94akJHZFE&hl=en_US#gid=0