Explain to me how memory width (128/192/256 bit, etc) is related to memory amount.

[MaxAwesome]

Hello guys!

I'm asking this out of curiosity, because I've been reading GTX 660 TI reviews, and in some forums people are saying that it's ridiculous to have 192bit memory bus paired with 2GB of VRAM, because it can't use it all.

Why is this the case? How does the bus width affect amount of memory on a GPU?

 

[mlee49]

Memory bus can be thought of literally as lanes of traffic. More lanes dedicated for traffic, the greater the flow. The graphics processor is connected to the RAM on the card via a memory bus.

As far as affects on gpu memory bus, I've seen mixed results. A wider bus doesn't simply mean the card is faster or will perform better. From my experience, I purchased an eVGA GTX 275 FTW with a wider memory bus and my gpu memory overclocked better than stock.

 

[MaxAwesome]

Memory bus can be thought of literally as lanes of traffic. More lanes dedicated for traffic, the greater the flow. The graphics processor is connected to the RAM on the card via a memory bus.

As far as affects on gpu memory bus, I've seen mixed results. A wider bus doesn't simply mean the card is faster or will perform better. From my experience, I purchased an eVGA GTX 275 FTW with a wider memory bus and my gpu memory overclocked better than stock.

This much I know.

You actually didn't mention anything about memory size and how it's related to bus width

I'll give the example above again: how does 192bit bus interfere with the 2GB of VRAM available on the GTX 660 Ti? Can all the VRAM be adressed or not? Is that the problem?

 

[Zubasa]

The way it works is quite simply the bus width pretty much controls the number of memory chips that can be used on the card.
A GPU with 256-bit bus has 8 memory chips minimum, since each memory chip have a 32-bit wide bus.
Some cards with double the memory amount will have 16 memory chips.
The 8 extra chips usually on the back of the card shares the bus with the 8 chips soldered to the opposite (front) side of the card.

The 660TI is a 192-bit bus which normally will have 6 memory chips resulting in 1.5GB of memory.
But in this case nVidia added 2 memory chips to the back of the 660TI 2GB to achieve this.
The 2 extra chips shares the bus with 2 memory chips on the front side of the card.

The first card that I know to employ this technique was the 550Ti.
All the vram can be addressed yes, but it is quite an unorthodox way to have such a set up.

 

[Frag_Maniac]

Yep, as said it's bandwidth, or 256bit being like 4 lanes of traffic (data), 192 like 3 lanes, 128, like 2, etc.

That said, many have speculated the Ti 660 with 192bit interface will be crippled, but others have pointed out that it appears Nvidia has found a way to make better use of less bandwidth.

So it's not just specs, a lot of it has to do with improvements in architecture and how well they can make use of a given GPU's specs.

 

[TRWOV]

The only drawback I see from the 192bit bus is that the last 512MB of those 2GB are going to deliver less bandwith as it uses a shared 64bit link, so if your GPU memory usage goes >1.5GB you could get an fps hit... at least that's how I understand it, please correct me if I'm wrong.

 

[hat]

The way it works is quite simply the bus width pretty much controls the number of memory chips that can be used on the card.
A GPU with 256-bit bus has 8 memory chips minimum, since each memory chip have a 32-bit wide bus.
Some cards with double the memory amount will have 16 memory chips.
The 8 extra chips usually on the back of the card shares the bus with the 8 chips soldered to the opposite (front) side of the card.

The 660TI is a 192-bit bus which normally will have 6 memory chips resulting in 1.5GB of memory.
But in this case nVidia added 2 memory chips to the back of the 660TI 2GB to achieve this.
The 2 extra chips shares the bus with 2 memory chips on the front side of the card.
http://www.techpowerup.com/reviews/Palit/GeForce_GTX_660_Ti_Jet_Stream/images/front.jpghttp://www.techpowerup.com/reviews/Palit/GeForce_GTX_660_Ti_Jet_Stream/images/back.jpg

The first card that I know to employ this technique was the 550Ti.
All the vram can be addressed yes, but it is quite an unorthodox way to have such a set up.

They're still doing that goofy shit? I haven't seen that in a long time.

 

[Hilux SSRG]

Taking from all these comments, it seems to me that 2gb on a 660Ti is more marketing than usefulness. Now I wonder about those 3gb 660Ti?

 

[mlee49]

This much I know.

You actually didn't mention anything about memory size and how it's related to bus width

I'll give the example above again: how does 192bit bus interfere with the 2GB of VRAM available on the GTX 660 Ti? Can all the VRAM be adressed or not? Is that the problem?

I would only tell you what I know.

So, back to traffic: Larger bus handles more memory, but higher memory(ie 2GB) thats alot of bits(cars) on the bus(road/lanes).

So just widening a highway and adding a lane does not mean traffic jams won't happen. Opening lanes for more cars is fine, but adding more cars(ie vram) will just fill up the lanes just as fast.

In theory, all the VRAM can be addressed. In reality, there is a buffer that will limit the amount of VRAM addressing before performing some other action(like offloading it to the cpu).

If Nvidia was a city planner and hoped to add this ultra fast highway with 4 lanes of traffic, but said "This highway will handle the amount of traffic 2 other highways" then they might have just created a bigger traffic jam than before.

Nvidia is smart though and certainly will use smart techniques to ensure this larger vram is warranted. Think Physx

 

[qubit]

The way it works is quite simply the bus width pretty much controls the number of memory chips that can be used on the card.
A GPU with 256-bit bus has 8 memory chips minimum, since each memory chip have a 32-bit wide bus.
Some cards with double the memory amount will have 16 memory chips.
The 8 extra chips usually on the back of the card shares the bus with the 8 chips soldered to the opposite (front) side of the card.

The 660TI is a 192-bit bus which normally will have 6 memory chips resulting in 1.5GB of memory.
But in this case nVidia added 2 memory chips to the back of the 660TI 2GB to achieve this.
The 2 extra chips shares the bus with 2 memory chips on the front side of the card.
http://www.techpowerup.com/reviews/Palit/GeForce_GTX_660_Ti_Jet_Stream/images/front.jpghttp://www.techpowerup.com/reviews/Palit/GeForce_GTX_660_Ti_Jet_Stream/images/back.jpg

The first card that I know to employ this technique was the 550Ti.
All the vram can be addressed yes, but it is quite an unorthodox way to have such a set up.

That's some weird sht and not something I'd like to see on my card. It should always be kept to a power of 2 for optimal design.

Good explanation.

 

[MSI650ti]

Hello guys!

I'm asking this out of curiosity, because I've been reading GTX 660 TI reviews, and in some forums people are saying that it's ridiculous to have 192bit memory bus paired with 2GB of VRAM, because it can't use it all.

Why is this the case? How does the bus width affect amount of memory on a GPU?

Yes a lower bit card can be crippled if the memory is set at a low speed like a GT 610 or a ATI 5450.
And yes the theory of the 128bit 192 bit 256 bit being lanes on a highway is correct, But there is one thing not mentioned about the smaller bit memory buses "or more narrow lanes" Is this, The wider highway (256 bus) could have a slower speed limit than the 2 lane (128 bit bus or 192 bit bus)
So this means that some newer cards even though they have a smaller memory bus does not mean they are slower than a card with a wider 256 bit bus. There are actually some 128 bit memory bus cards that are faster than 256 bit memory bus cards. This is because memory clock speed can be different and so on. The actual important thing to look at is the memory bandwidth of the card, How many (GB/sec) it is. The basic 660 has 144.2 GB a second in memory bandwidth, Which is very fast, And more than enough to handle even 3-4GB of video memory. I hope this helps.

 

[Vario]

I've learned to always buy cards with larger bus width, they seem to work smoother overall, just my personal experience.

Also there were a few 500 series cards that had enormous quantities of ddr3 with small 128 bus. What a joke..

example: http://www.tweaktown.com/news/20598...ew_overwhelms_gtx_550_ti_w_4gb_ram/index.html

 

[Aquinus]

I've learned to always buy cards with larger bus width, they seem to work smoother overall, just my personal experience.

Also there were a few 500 series cards that had enormous quantities of ddr3 with small 128 bus. What a joke..

example: http://www.tweaktown.com/news/20598...ew_overwhelms_gtx_550_ti_w_4gb_ram/index.html

That's perfectly fine if you don't really plan to play video games on it very often. DDR3 is cheap enough to produce anyways so I suspect that it was the best cheap option nVidia had for the price range. You want GDDR5? Pay for it. It's that simple.

 

[Pill Monster]

That's perfectly fine if you don't really plan to play video games on it very often. DDR3 is cheap enough to produce anyways so I suspect that it was the best cheap option nVidia had for the price range. You want GDDR5? Pay for it. It's that simple.

You missed his point. The card has 4GB of DDR3......completely stupid amount it could never use it.

 

[Aquinus]

You missed his point. The card has 4GB of DDR3......completely stupid amount it could never use it.

Like I said, they went with the cheapest ICs they could get their hands on. DDR3 is fairly cheap in comparison. Of course it's rediculous but from a cost perspective, it very well could have costed less to produce the video card than putting 2GB or even 1GB of GDDR5 on it instead. Not to mention you need a controller on the GPU to handle it.

It's all about cost to produce the product.

Also, if someone is doing CUDA, OpenCL, or DirectCompute work, you may not need a ton of GPU horse power to get a lot done. This card clearly doesn't target gamers, that's my point.

 

[Pill Monster]

This card clearly doesn't target gamers, that's my point.

Yeah that's true.....even so they prob could've made the card cheaper by using less memory.....but I digress. Hell with DDR3 discrete card u may as well get an A10 APU anyway. lol

 

[Vario]

Like I said, they went with the cheapest ICs they could get their hands on. DDR3 is fairly cheap in comparison. Of course it's rediculous but from a cost perspective, it very well could have costed less to produce the video card than putting 2GB or even 1GB of GDDR5 on it instead. Not to mention you need a controller on the GPU to handle it.

It's all about cost to produce the product.

Also, if someone is doing CUDA, OpenCL, or DirectCompute work, you may not need a ton of GPU horse power to get a lot done. This card clearly doesn't target gamers, that's my point.

Its not that. Its arrangement is completely suboptimal for performance. It does not have enough bandwith to even use that ram. its a total gimmick. http://techreport.com/news/21536/point-of-view-insults-intelligence-with-geforce-550-ti-4gb

Ladies and gentlemen, I present to you the world's most pointless graphics card. Nvidia partner Point of View has come up with a GeForce 550 Ti with four gigabytes of graphics memory, which makes about as much sense as putting a spoiler on a Smart car. Epic fail.

The problem with this new card, of course, is that the lowly GeForce 550 Ti doesn't need anywhere close to that much RAM. When we tested the 550 Ti back in March, we discovered that the budget GPU is really only fast enough to play games at the relatively low resolution of 1680x1050, which works out to less than two megapixels. No wonder most of the 550 Tis on the market have just 1GB of memory.

While Point of View loads the card up with high-density memory chips, they're not very fast ones. The GeForce 550 Ti is meant to be paired with GDDR5 memory with an effective data rate of 3.6 GT/s. Point of View hits the 4GB mark with much slower DDR3 memory capable of pushing bits at less than one third that speed.

Graphics card makers have long tried to dress up low-end models with obscene amounts of memory, but this is the most egregious example yet. Some consumers are easily swayed by bigger numbers, I guess. Thankfully, it looks like the card isn't available in North America right now. According to Computer Base, Euros can order the thing for €113.

 

[Pill Monster]

LOL I just looked at the specs on POV homepage, as if DDR3 wasn't bad enough by itself it's only 1066mhz. What were they thinking.....0_o

actually speed prob doesn't matter they could make it 2133mhz it would still take a week for data to get to the GPU just like being on a Greyhound bus. (get it?) lol