Bitcoin Forum

Why do so many people underclock their video memory?

lower gpu temps

On my 6850 lowers temp 64 to 63c but gives 3mhs faster also.

A more technical answer:

There is no need to have fast memory, or much memory, when hashing since nothing is being cached in the RAM on the card. When you play games, the textures needs to be cached somewhere for faster rendering since keeping it in RAM would be quite slow due to the distance from the GPU. Having faster memory on the card is optimal in gaming since the DDR memory on graphic cards are faster and closer to the GPU.

so this is why i wonder if its really worth it as its not even being used so musn't be using squat power

Since the RAM on graphics cards are Dynamic Random Access Memory(DRAM), they need their values refreshed on an interval since the capacitors storing memory leak the voltage. This is done regardless if a address in memory is being pointed to or not, or "being used" in simpler terms. The memory clock controls this refresh interval, thus lower memory clock equals lower intervals to refresh voltages, which uses less power.

thanks

Doesn't seem to make much difference in hash rate, from what I can tell..

i tried under clocking mine in afterburner but it has no effect ... over clocking the core or memory works though. and  i cant do it in ccc due to restrictions.

any tips ?

No. The clock speed has nothing whatsoever to do with the refresh interval.

I underclock memory

use rivatuner to underclock the memory.

Out of curiocity, how much does this save in power consumption and heat?

Honestly i don't think it helps Temps or Power.

Having found this thread, I have underclocked the memory on my RADEON 5850HD and the temp has dropped from 81°C to 72°C and the power consumption has dropped by over 100watts according to my energy monitor, or somewhere between £0.01 and £0.02 per hour.

 Video card RAM is not used for textures.  Those are stored on the system memory and hard drive.  The RAM on a video card is only used to store the rendered frames before they are sent to the monitor.

That was true a decade ago, it's definitely true now. Here's two ways to see that this can't possibly be true:

1) 32-bit color, 1900x1200 = less than 9MB per frame. You really think they put 1GB on video cards so they can hold 117 frames?!

2) How could that possibly work? The GPU is running at blistering speed. How could it possibly apply textures if it had to keep copying between the GPU and system memory? Low end cards these days exceed 40 billion texels per second. You really think they could sustain that rate to system memory (that's already busy with system stuff) that tops out at less than 8GB/s?

The whole point to having all that fast memory is so that the GPU can render textures without bothering the CPU or main memory.

 Do I really think they put 1GB on graphics cards to hold 117 frames?  YES.  How else are you going to get smooth framerates?  In Rift I get 40 - 60 FPS.  In PlanetSide I get 100+. In EQ2 I get 30 - 50.  If the rendered frames had to compete with textures then your computer would grind to a halt.  I run LOTRO with Hi-Rez textures.  That folder ALONE is over 11GB.  That very fact makes it impossible to store textures on the VRAM.  Since it would have to pick and choose what textures it needs, it would have to call them from the hard drive, to system RAM and finally the VRAM.  That's idiotic.  Especially when it makes far more sense to only send the textures you need for a given frame. 

You should research video cards before trying to make such arguments, or at least know when to look things up when someone tells you you're wrong.

"The RAM on a video card is only used to store the rendered frames before they are sent to the monitor." - This statement is just wrong, mainly because of that "only." VRAM stores lots of things besides acting solely as a frame buffer. This Memory section of the Video Card entry in wikipedia (http://en.wikipedia.org/wiki/Video_card#Video_memory) explains very briefly what VRAM is used for.

You're statements of 11GB of textures and the way you think they're loaded makes it seem like you have no idea how caching systems work. Maybe you should read about that before trying to act like an expert.

If I've learned one thing from these forums, it'd be JoelKatz is usually right about hardware.

Read the wikipedia entry again.  It has the word 'may' in it.  Just because it's capable of different things doesn't mean it is used for those things. 

Any time you enter a new cell in a game the textures needed are loaded into VRAM for that cell. This is why your VRAM usage goes up with the use of high-res texture packs. It is also why your VRAM usage goes up with the use of Anti-Aliasing and Anisotropic Filtering which is applied to those textures present in the VRAM.

I really don't want to bring this down to the personal insult level, but seriously, RESEARCH before posting nonsense. You're more stubborn than talking to a brick wall. There's nothing bad about being wrong about something, buts it's bad when you continue to try to spread misinformation even though many are weighing in.

I doubt you even googled VRAM usage, since I found many results that I looked at before posting that explained what VRAM is used for these days.

For the love of god! Think, man! If they rendered more than a few frames ahead, the lag when you tried to move would be intolerable.  No modern game renders more than 3 frames ahead.

No, because thankfully they have GPUs with fast local memory that are designed from the ground up to do *exactly* that.

It also makes it impossible to fit the textures in system memory. On modern systems, the video card memory and the system memory are of the same order of magnitude. Games load those textures they need for a particular region onto the video card. They pick the texture resolution (and in some cases, use compression) based on the available RAM on the video card.

Why? That's actually remarkably efficient. The hard drive controller can write directly into system RAM. And the GPU can read directly from system RAM. This keeps the CPU out of the loop.

You have to be crazy. There is no way you want to be doing anything per-texture, per-frame except on hardware 100% designed for exactly that purpose, which would be the GPU and its fast, local memory.