Never really paid much attention to memory, I tried a bit higher 4514Mhz and it was a no go with 1087mv (core but seems to help memory as well)



Think I should pencil mod the memory voltage?

Depends. What kind of performance increases in games are you getting with the memory overclock?

Basically none in this case, the performance increase at stock clocks /w that memory overclock is like 5%.

Now I remember why overclocking Fermi was core first, then memory.

i never got to play with a GTX 470. If overclocking the memory 34% over stock has such a small 5% performance increase, overclocking it further shouldn't be a priority unless you just like to volt mod for fun

We'll it's funky because even though it was gaming stable for about an hour as soon as I turned the core up to 800 which at 1087mv is well over what I'd need at lower memory speeds it crashed the driver.

So either ram vid is getting reduced by the higher clock speed, or the added core speed stressed the memory further thus resulting in a driver crash.


Either way I'm ready for GK110

It's usually that way with overclocking. You get a max core overclock with stock memory; then you set the core back to stock and find the max memory overclock. However, when you attempt to put the two 'maxes' together, it may not be stable - depending on how well the card is build for overclocking.

Many times i have had to adjust my core and/or memory clocks downward from what was otherwise maximum stable overclock when it was overclocked separately.

Fermi's memory controllers were FUBAR and GF11x did little to fix that, Kepler addressed it quite nicely. Memory overclocking on Fermi seems to be a pointless exercise in futility to me, especially if you are having to overvolt to do it.

i almost forgot .. it's like Fermi was ages ago


Yes, the surprise for me was in overclocking the memory of the GTX 680 when i first got it. Voltage may be locked down, but at least the memory not only overclocks well, it makes a solid performance improvement.

I just don't think the card is fast enough to take advantage of a huge memory bandwidth, even going from 1850 to 2100 at 900/918 core only netted an addtional performance increase of 8%.

1850 is 148 GB/s 2100 is 168 GB/s that's a 14% bandwidth increase with a 2% core/shader increase netting a mere 8% performance increase and probably the worst scaling I've seen in clock speeds.

It depends on what games you're using. If you're using the latest and the most demanding DX11 games that are already rather limited by the GPU's shader performance, then the bandwidth won't help that much.

Generally, the bandwidth helps more for older games. When Fermi first came out, the memory overclock yielded a higher overall performance gain than it does now... for the newer games!

Your 35% memory overclock should've yielded at least 8% overall gain by itself when considering all modern games - not the select few ones that are commonly used among the forum enthusiasts as popular benchmark "scales of measurement".