Memory Bus size and how it effects your VRAM usage

[Kuzma]

Why is compute a exception? Doesn't the vram ( water) STILL have to got through the bit tunnel?

with compute the analogy changes since all the water doesn't need to be accessed at the same time.

[maazster]

Finally, some one who gets it, I have been saying this for ages but no, no one would listen. Oh yeah as usual great topic, Kuzma.

[TTesla]

with compute the analogy changes since all the water doesn't need to be accessed at the same time.

Why does compute NOT need all of the water at once? And why do game DO need ALL of the water at once?

OH and what does OP stand for? As in you are the op? Since yiu started the topic? Sorry for the randomnes!

[Kuzma]

Why does compute NOT need all of the water at once? And why do game DO need ALL of the water at once?

OH and what does OP stand for? As in you are the op? Since yiu started the topic? Sorry for the randomnes!

Games are fast paced and random so every texture has to be available to be displayed at any given point in time and are very much variable whereas compute is generally in a sequence (e.g. rendering a video with OpenCL) you know what the next instruction is and as such there's no need to have every texture available at the same time

[Kuzma]

Finally, some one who gets it, I have been saying this for ages but no, no one would listen. Oh yeah as usual great topic, Kuzma.

posts like this that put a smile on my face :3 thank you

[Gunjob]

posts like this that put a smile on my face :3 thank you

Very informative and clear, definitely wasn't something I knew and I consider myself to be fairly knowledgeable.  

[Askew]

Still though none of this explains the GK110 anomaly, if it is true what is written about it's ability, then there is a reason for the occurrence.

 

Someone must be aware of the reason, that's the most interesting part of this for me.

[Kuzma]

Still though none of this explains the GK110 anomaly, if it is true what is written about it's ability, then there is a reason for the occurrence.

 

Someone must be aware of the reason, that's the most interesting part of this for me.

I only discovered the anomoly a few hours before posting this thread (it took me about a day to write this out ._. + research and cross-ference and benchmark reading etc. ) so I haven't been able to find a single thing out but I am very much interested myself too, I only posted that because I hope that someone else can find the answer to thaat which I could not

 .

[Kuzma]

Why does compute NOT need all of the water at once? And why do game DO need ALL of the water at once?

OH and what does OP stand for? As in you are the op? Since yiu started the topic? Sorry for the randomnes!

OP = Original Poster

[Askew]

I only discovered the anomoly a few hours before posting this thread (it took me about a day to write this out ._. + research and cross-ference and benchmark reading etc. ) so I haven't been able to find a single thing out but I am very much interested myself too, I only posted that because I hope that someone else can find the answer to thaat which I could not

 .

 

Ok well I'm following this thread so if it comes to light and is posted here I'll appear to talk about it

[maazster]

posts like this that put a smile on my face :3 thank you

You are welcome.

[MdX MaxX]

Can you elaborate on what you mean by accessible?

 

I'm not understanding this notion that the amount of accessible memory is limited by the size of the memory bus.  The bus size represents the amount of data that can be read in a clock cycle (qualitatively speaking).  How does that affect the address space in any way?  Regardless of bus size, as long as the instruction size is sufficiently large, say 64 bits, all memory locations should be accessible if you have less than close to 2^64 bytes of memory (which we won't see on a graphics card in our lifetimes).

 

Are you saying, rather, that only a certain amount of memory is accessible per some unit of time?  That makes much more sense.  As game environments are constantly changing and rendered in "real time," it would make sense that only a certain amount of memory would be usable per unit time before the GPU would have to move on and render the next thing.  In that case, what unit of time are you referring to?  How does this factor into your calculations?  I'm not quite sure what your Titan calculations are getting at either.

 

I don't mean to sound like I'm attacking you or anything; I'm just trying to figure out why a 6GB 780-Ti would be "useless."

[jjblack]

This question is one that drives me nuts as well.

Sometimes it seems like that extra memory may act like a cache or perhaps like a ramdisk for textures. Often many textures in one rendered frame are reused to create that image. I would hope that developers and nvidia would have figured out a way to not reload every texture in a game that is not being used to render that specific image. That is the whole point of a buffer. Maybe some of that6gb is used as a buffer for transitioning frames to improve average framerate. The card may load 6gb of textures but in any given frame with improved coding only needs access to 3 to generate the image at x resolution and settings. Next frame might use 3.5.

Maybe the reason we dont see huge performance gains in cards that have more memory is we are hitting some other limitations with gddr5 unrelated to bus bandwidth or perhaps we arent as far along in core architectural design to piece image renderings together efficiently enoughto make use of all that vram. Yeah the ram is "being used" but it doesnt really translate to significant performance gains. I would hope how these images are constructed and accessed is as important to red and green as giving us more cores, bandwidth and ram etc.

[Ashaira]

would anyone care to explain how much of a gain would 2 4GB cards with 256 width over a single 4GB card with 512 width. for a more concrete example: 2 270X 4GB vs an R9 290X.

 

from the 2 benchmarks i have seen for crossfire 270X it seems they are a bit better than a 290X but since in crossfire one card basically has to replicate the memory of the other isn't some of the width wasted on that so overall you wouldn't have 512 total from this crossfire.

 

I am at a total loss when it comes to memory buss and crossfire.

[R4STABAN]

4 GB version of 2 GB cards are not a complete 'waste'. The ability to cache extra textures can make the difference between playing a game and not playing it.

 

However, using the GTX 680 as an example (and ignoring skyrim..), the card does not have the gpu grunt to run settings that will require more than 2GB of vram anyway (at 3 x 1080p) until you have 3 cores in SLI

 

OP is right about the important of the memory bus though. Sticking with the 680 example, it's 256bit bus means it read from 4 chips at once. When you get a 4GB variant you double the amount of chips. The card can utilise all of the vram available to it, but the extra vram results in higher latency. This is why double vram cards have lower memory (and therefore overall) performance and slightly higher frame times.

[Ebinlaw]

Then can you somehow calculate the maximum amount of ram(water) the card can access in games at certain memory clock? Like 384-bit x 6000mhz = 2.304 Gb/sec? So max amount of ram that this theoretical card could use is 2.3GB?

[cliff_182]

thnx man 

very helpful

[R4STABAN]

Then can you somehow calculate the maximum amount of ram(water) the card can access in games at certain memory clock? Like 384-bit x 6000mhz = 2.304 Gb/sec? So max amount of ram that this theoretical card could use is 2.3GB?

 

It can access all the memory on the board.

 

384 bit bus means reading 64 bits from 6 chips for each clock.

 

In your theoretical example the maths is way off. A 384 bit bus with 6000mhz memory creates a memory bandwidth of 288 GB/s:

 

384 x 6000MHz = 2304000(million) bits

 

2304000(million) / 8 = 288000(million) bytes = 288GB

 

(note this is gigabytes in the storage sense (decimal) and not the traditional gigabyte which is usually referred to as a gibibyte(binary))

[Ebinlaw]

Then why you shouldn't buy a gtx 770 4gb, if it can still access all the memory on the board?

Is 6 memory chips the maximum amount 384 -bit can handle? That's why 770 4gb is useless? In order to effectively use all 4gb you need 512-bit bus?

Am I getting it right?

[R4STABAN]

Then why you shouldn't buy a gtx 770 4gb, if it can still access all the memory on the board?

 

 

With the exception of inefficient skyrim mods, the game settings / resolution required to fill more than 2GB of memory on a 680/770 will bring the GPU core to it's knees unless you have 3 of them in SLI. 2 in SLI is borderline @3x1080p, you have to turn windows aero off etc.

Is 6 memory chips the maximum amount 384 -bit can handle?

 

64bits is read from each chip, 64x6 = 384. 384 bits read each 'transaction', clockspeed = number of 'transactions' per second

 

 

That's why 770 4gb is useless?

 

No, see first answer. Also, 680/770 has a 256 bit bus

 

In order to effectively use all 4gb you need 512-bit bus?

Am I getting it right?

 

No. It would slightly help frame times etc if all other variables were the same, i.e memory chip denisty, speed, clockspeed, etc. But it's not a necessity.

 

The 680/770 has 8 ram chips on it btw.

[DEvilgodspidER]

Reviving an old thread but, I'm thinking of buying a GTX 770 4GB from Gigabyte, what do my memory clocks need to be in order for me to fully utilize 3GB and 4GB VRAM efficiently?

The memory bus is 256Bits.

[7850OC]

Reviving an old thread but, I'm thinking of buying a GTX 770 4GB from Gigabyte, what do my memory clocks need to be in order for me to fully utilize 3GB and 4GB VRAM efficiently?

The memory bus is 256Bits.

the news are that the new line of 9xx is coming in short time (they say)  <_<

 

you may want wait for that, before buy the old line friend :rolleyes:

 

(or buy a R9 285) 

[D2ultima]

Reviving an old thread but, I'm thinking of buying a GTX 770 4GB from Gigabyte, what do my memory clocks need to be in order for me to fully utilize 3GB and 4GB VRAM efficiently?

The memory bus is 256Bits.

Read the vRAM guide in my signature. The short answer is "it will already utilize 4GB vRAM without issue".

 

the news are that the new line of 9xx is coming in short time (they say)  <_<

 

you may want wait for that, before buy the old line friend :rolleyes:

 

(or buy a R9 285) 

And as well, you may want to wait for a GTX 970 or something. Should be a similar price point and should stomp the 770 into the ground. It's to be announced in a couple days (I think) or at worst on the 18th of September; so unless you need a card RIGHT now, then you can look for them then.

[DEvilgodspidER]

Read the vRAM guide in my signature. The short answer is "it will already utilize 4GB vRAM without issue".

 

And as well, you may want to wait for a GTX 970 or something. Should be a similar price point and should stomp the 770 into the ground. It's to be announced in a couple days (I think) or at worst on the 18th of September; so unless you need a card RIGHT now, then you can look for them then.

 

 

the news are that the new line of 9xx is coming in short time (they say)  <_<

 

you may want wait for that, before buy the old line friend :rolleyes:

 

(or buy a R9 285) 

Thank you

I have decided to wait for the GTX 970, which already came out, so right now I'm waiting for some sort of Christmas deal or for AMD's GPU's to come out in order for the GTX 970's price to go down.

I know that the price is perfect, but I just don't think spending 371€ (plus shipping) where I live isn't the same MSRP relation here as it is in the US (GTX 770 4GB - GTX 970).

What are the prices there? Before the GTX 970 came out how much was the GTX 770 4GB?