Memory Bus

[Honeydiddle]

I see many people saying don't buy the 4GB 770 because is has a 256 bit bus and generally you should have 128 bits per 1GB of VRAM, hence why the 7970 and 780 are 3GB and have a 386 bit bus. However the titan has 6gb of VRAM and has a 386 bit bus?

 

I know there there must be some sort of logical explanation?

[TheCroaker]

I dont know much so just.. to say it really coz I thought of it, is it possibly because Titan is supposed to be a workbench card not a gaming card?

[M-ursu]

Lets just say that if you buy 4GB 760, you will be fine whit it, because it have enough big mem bus.

So just do not worry about buss, let manifactures worry about it.

[Glenwing]

I have not heard that rule before...

 

The memory bus is split between the memory modules, so naturally there will always be a common factor that divides evenly into both the memory amount and the memory bus... that is the only corallation between bus width and memory amounts from an engineering standpoint.  For example on the GTX 770, there are 8x or 16x 256MiB modules (total 2GiB or 4GiB), and on a reported 256-bit memory bus, this means every 256MiB DRAM is on a 32-bit bus, so when combined you have 2GiB of DRAM (8x 256MiB) on a 256-bit bus (8x 32-bit).  When you add the additional 8x modules on the back of the card, for a total of 4GiB of DRAM, they are doubled up so you have 2 modules on each 32-bit bus.  The Titan has 24x memory modules doubled up on 32-bit busses, so 12x 32-bit busses makes an overall 384-bit bus.  And doubled up 256MiB modules gives you 24x256MiB or 6GiB of DRAM.  But having two memory modules doubled up on each 32-bit bus doesn't mean the extra RAM is unusable or something.  It works perfectly fine.  It is not as good as if each module had its own dedicated 32-bit bus but I don't see a reason to avoid cards with doubled up memory modules on each bus.  I have never heard of this causing noteworthy performance problems.

[Dravic]

I have not heard that rule before...

 

The memory bus is split between the memory modules, so naturally there will always be a common factor that divides evenly into both the memory amount and the memory bus... that is the only corallation between bus width and memory amounts from an engineering standpoint.  For example on the GTX 770, there are 8x or 16x 256MiB modules (total 2GiB or 4GiB), and on a reported 256-bit memory bus, this means every 256MiB DRAM is on a 32-bit bus, so when combined you have 2GiB of DRAM (8x 256MiB) on a 256-bit bus (8x 32-bit).  When you add the additional 8x modules on the back of the card, for a total of 4GiB of DRAM, they are doubled up so you have 2 modules on each 32-bit bus.  The Titan has 24x memory modules doubled up on 32-bit busses, so 12x 32-bit busses makes an overall 384-bit bus.  And doubled up 256MiB modules gives you 24x256MiB or 6GiB of DRAM.  But having two memory modules doubled up on each 32-bit bus doesn't mean the extra RAM is unusable or something.  It works perfectly fine.  It is not as good as if each module had its own dedicated 32-bit bus but I don't see a reason to avoid cards with doubled up memory modules on each bus.  I have never heard of this causing noteworthy performance problems.

 

That said, there's even more to it, because if there was any loss of speed it would be negligible compared to the performance boost when card doesnt have to tax the processor and HDD/SSD/RAM for getting new textures onboard for calculations thanks to the vram being bigger. Or the other way around, the loss of speed when the card IS taxing the system to get new textures onboard.

 

Generally it should be noted, TODAY - you dont need 4GB unless you use 2 cards or more in SLI/Crossfire.

Today.

Now for "TOMORROW" we don't know, but is only logical that we FINAAALYYY will see high res textures in games (consoles have enough RAM, 4-5GB).

[Kuzma]

I have not heard that rule before...

 

The memory bus is split between the memory modules, so naturally there will always be a common factor that divides evenly into both the memory amount and the memory bus... that is the only corallation between bus width and memory amounts from an engineering standpoint.  For example on the GTX 770, there are 8x or 16x 256MiB modules (total 2GiB or 4GiB), and on a reported 256-bit memory bus, this means every 256MiB DRAM is on a 32-bit bus, so when combined you have 2GiB of DRAM (8x 256MiB) on a 256-bit bus (8x 32-bit).  When you add the additional 8x modules on the back of the card, for a total of 4GiB of DRAM, they are doubled up so you have 2 modules on each 32-bit bus.  The Titan has 24x memory modules doubled up on 32-bit busses, so 12x 32-bit busses makes an overall 384-bit bus.  And doubled up 256MiB modules gives you 24x256MiB or 6GiB of DRAM.  But having two memory modules doubled up on each 32-bit bus doesn't mean the extra RAM is unusable or something.  It works perfectly fine.  It is not as good as if each module had its own dedicated 32-bit bus but I don't see a reason to avoid cards with doubled up memory modules on each bus.  I have never heard of this causing noteworthy performance problems.

Look at the 680 and 660 TI 4GB benchmarks; the cards weren't using more than 2.4-2.6GB VRAM no matter how hard they pushed them until they overclocked the memory clock. The memory bus wasn't big enough until they ran the cards in 3x SLI. GK110 seems to stick it's middle finger up at this though :/ 12GB on 384bit bus.

[Glenwing]

Look at the 680 and 660 TI 4GB benchmarks; the cards weren't using more than 2.4-2.6GB VRAM no matter how hard they pushed them until they overclocked the memory clock. The memory bus wasn't big enough until they ran the cards in 3x SLI. GK110 seems to stick it's middle finger up at this though :/ 12GB on 384bit bus.

 

Oh yes, I forgot about 600 series... memory bus was a bit small there.  680 core was never meant to be positioned as the flagship card lol.  Although the 770 addresses this a bit with the higher clocked memory out of the box.

[Kuzma]

Oh yes, I forgot about 600 series... memory bus was a bit small there.  680 core was never meant to be positioned as the flagship card lol.  Although the 770 addresses this a bit with the higher clocked memory out of the box.

True but not by enough :p the 2.6 was with a effective memory clock of 6500mhz on the 680 so meh the full 4GB still probably wouldn't be filled unless you watercool.

[Honeydiddle]

So if i got a 4gb 770 and overclocked it with water, it could be better and semi more future proof than a 780?

[AMD_Nemesis]

It's already more future proof at 1440 it beats the 780 sometimes not always and a few people have told me this VRAM thing doesn't exist they have used like 3.2 gb of it and didn't overclock I'm going to tell you this a 7704gb or even 760 4gb is not a waste at all.

[Honeydiddle]

Interesting, i may still go the 780 due to newer features though

[Kuzma]

So if i got a 4gb 770 and overclocked it with water, it could be better and semi more future proof than a 780?

If you pushed your memory clock to about 7500; which should be possible on a water loop then yes it would be. Thanks for reminding me about gpu boost 2.0 :p

 

It's already more future proof at 1440 it beats the 780 sometimes not always and a few people have told me this VRAM thing doesn't exist they have used like 3.2 gb of it and didn't overclock I'm going to tell you this a 7704gb or even 760 4gb is not a waste at all.

Look at 680 benchmarks then tell me otherwise; either those few people got really good overclocks on their memory bus' or are lying.

 

Interesting, i may still go the 780 due to newer features though

Taking into account the overclock on a 4GB 770 it makes a lot more sense to go for it.

[Honeydiddle]

Tough choice haha

[AMD_Nemesis]

Tough choice haha

eh not really its good you get over 2 gigs I got my 770 today overclocked it a bit and in bioshock infinite and saw 2200mb used. So even 2gb is outdated now that's why I got 4 gigs also ran a texture test and got up to 3.7 gb

[ExDreamer]

It depends on the clock speed of the memory itself.

Technically you could have a 64-bit memory bus, but if clocked at 16GHz, it would have more bandwidth than a 512-bit bus clocked at 1GHz.