How Does GPU Spec Change Performance?

[piggykid1]

So I know each card has a different chipset and that's what makes each card better or worse than the other.

 

But how does each specification of a card change the performance of a card. All I really know about it is that VRAM helps run greater resolutions and is better for mining, folding, and boinc etc.

 

So how does each of the following effect performance (if I missed one tell me please)

 

1. VRAM/ Memory

 

2. Core/ Boost Clock

 

3. Memory Clock

 

4. Texture Fill Rate

 

5. Memory Bit

 

7. Memory Bandwidth

 

Also I am a little confused about the 16x and 8x PCI-E thing.

 

Thanks!

[Shd0w2]

vram can influence performance positively with the more you have at higher resolutions. you wont need 4gb of ram for 1080p thought generally. but with 4gb of veam at 1440p, it will have a framerate increase over a card with 2gb of vram.

core clock is how fast the processing units are running in the gpu, it will increase with a higher clock,depending in gpu, there will be a performance increase,s sometimes slight, or huge.

memory clock is how fast the vram is running, just like normal ram, a hgiher speed will also usually slightly increase performance

texture fill rate, i dont really know how to explain it

memory bit, neither

memory bandwith, meh

[ericlee30]

vram can influence performance positively with the more you have at higher resolutions. you wont need 4gb of ram for 1080p thought generally. but with 4gb of veam at 1440p, it will have a framerate increase over a card with 2gb of vram.

core clock is how fast the processing units are running in the gpu, it will increase with a higher clock,depending in gpu, there will be a performance increase,s sometimes slight, or huge.

memory clock is how fast the vram is running, just like normal ram, a hgiher speed will also usually slightly increase performance

texture fill rate,

Very well put, couldn't of said it any better.

[Guest]

So I know each card has a different chipset and that's what makes each card better or worse than the other.

 

But how does each specification of a card change the performance of a card. All I really know about it is that VRAM helps run greater resolutions and is better for mining, folding, and boinc etc.

 

So how does each of the following effect performance (if I missed one tell me please)

 

1. VRAM/ Memory

 

2. Core/ Boost Clock

 

3. Memory Clock

 

4. Texture Fill Rate

 

5. Memory Bit

 

7. Memory Bandwidth

 

Also I am a little confused about the 16x and 8x PCI-E thing.

 

Thanks!

 

1. VRAM is complicated. More shaders/stuff to process by the GPU in-game means more memory taken up. That doesn't mean that the GPU is fast enough to make use of all of it. For example, a GTX 770 does not need 4GB and cannot use 4GB by itself to its advantage, however, because in SLI the VRAM is cloned, meaning 2 GPUs run off the same amount of memory, 4GB is better because the setup is fast enough to need it. Nothing over 2GB is required for 1080p.

 

2. Core/Boost clock: this is highly GPU dependent. If the GPU is extremely large and has an extremely large amount of stream processors/texture units/ROPs, then obviously its large size produces more heat than a small GPU, so it won't necessarily achieve higher clockspeeds. Generally, higher clock = better performance. Before 2012, this was even more complicated because there were different clockspeeds for the GPU core and the shader part of the GPU. Now they are unified. Boost is like a CPU, when it's under heavy load, it can boost up to an even higher frequency, but most cards can't stay there for long because it's like the CPU: Boost clock causes more power consumption, temperature and noise.

 

3. Memory clock: again, highly dependent on the specific model. Higher clock = higher bandwidth, but generally memory bus width and memory type is more important.

 

4. Texture Fill Rate: OK, I'm no expert, but this is pretty much as straightforward as it sounds. The more powerful your GPU is, the more textures it can process in the same amount of time. It has become murky as of late; for example, the GTX 750 Ti is pretty meager on paper, but it performs far better than it should theoretically. Remember that the rates you're talking about here are usually some number x number of processors x clockspeed. keep in mind that the processor we're talking about here are the ROPs and TMUs, NOT the shaders! (can't remember which is which for texel and fill)

 

5. Memory bit width: WHen you look at videocard reviews and they show you a shot of the entire PCB, you'll notice that on high end cards, there are more memory modules than on low end cards. Because of this, they have more memory bit width. For example, if a card like the 270X has a 256-bit width, then it has 8 memory chips, each with 32-bit width. 32 x 8  = 256. This is arguably more important than memory clock. This was why the GTX 660 and GTX 660 Ti's 192-bit bus raised some eyebrows; it was highly asymmetrical and people doubted whether the GPU could efficiently use all of the VRAM. This was probably also why they ended up being slower than the 256-bit HD 7870 and 256-bit GTX 760 respectively.

 

6. Memory bandwidth: some calculation involving the memory bus width and the memory clock. However, keep in mind that this is the bandwidth that exists between the GPU and the graphics memory in your card, and not between the card and the system. That is reliant on the PCIe slot speed/width. Not really a problem for high end cards today in games.

 

 

About the x8 and x16, it is rather complicated as well. CPUs and chipsets come with a set amount of "lanes". Each lane is represented as x1 (x16 is 16 lanes, etc.). x16 is faster than x8, but each generation of PCIe is also faster (3.0 is twice as fast as 2.0). Thus, PCIe 3.0 x8 is equal to PCIe 2.0 x16.

 

 

Mining/Folding don't require any more than a 2.0 x4 speed at maximum. VRAM may help, but that doesn't justify getting a 4GB GTX 770 over a 2GB model; it's probably more dependent on the GPU itself.

 

 

Hope I helped, i'm in kind of a rush

[Createinator]

1. VRAM/ MEMORY allows the card to have larger amount of DATA for quick access available in VRAM making it work faster and games/ output to be smoother(stutter/ tearing free)2
2. Core/ Boost Clock is the rated frequency the card runs at stock and the rated frequency card can "boost"(run faster) to when the demand arises, it is very similar to CPU that can do 3.5 GHz at stock and boost to 3.7GHz on demand.
3. Memory Clock Speed indicates the speed at which GPU can access data stored in VRAM
4. TMU and ROPs give an idea about the speed and precision of 3D output and Color reproduction.
5. Memory Bit or BUS is the available bandwidth for GPU to access VRAM at any given point of time. Therefore any card with large amount of VRAM but small BUS will not be able to fully utilize its VRAM.
6. CPU allows a certain number of PCIe lanes, each lane allows for limited bandwidth. An X16 PCIe slot has 16 lanes whereas an X8 slot has half as many, making it relatively slower.

[Cubit481]

memory bit, neither

memory bandwith, meh

higher memory bit decreases the time required to access the vram...so ya it matters...

[Shd0w2]

1. VRAM is complicated. More shaders/stuff to process by the

Hope I helped, i'm in kind of a rush

great job, i really rushed mine, but nebertheless. amazing glad to see dedication on here.

[Shd0w2]

higher memory bit decreases the time required to access the vram...so ya it matters...

i meant i didnt really know how to explain it.

[Cubit481]

i meant i didnt really know how to explain it.

okay... ..great job in explaining rest of the things though...

[Shd0w2]

okay... ..great job in explaining rest of the things though...

Thanks, excuse my errors in punctuation. Im using a crap pad mini.

[Cubit481]

Thanks, excuse my errors in punctuation. Im using a crap pad mini.

ohh i can understand...used it myself...

[Shd0w2]

ohh i can understand...used it myself...

Yeahhhh, it wouldnt recognize cpu,gpu,vram, etc , so i disabled autofix pr whatever

[piggykid1]

1. VRAM is complicated. More shaders/stuff to process by the GPU in-game means more memory taken up. That doesn't mean that the GPU is fast enough to make use of all of it. For example, a GTX 770 does not need 4GB and cannot use 4GB by itself to its advantage, however, because in SLI the VRAM is cloned, meaning 2 GPUs run off the same amount of memory, 4GB is better because the setup is fast enough to need it. Nothing over 2GB is required for 1080p.

2. Core/Boost clock: this is highly GPU dependent. If the GPU is extremely large and has an extremely large amount of stream processors/texture units/ROPs, then obviously its large size produces more heat than a small GPU, so it won't necessarily achieve higher clockspeeds. Generally, higher clock = better performance. Before 2012, this was even more complicated because there were different clockspeeds for the GPU core and the shader part of the GPU. Now they are unified. Boost is like a CPU, when it's under heavy load, it can boost up to an even higher frequency, but most cards can't stay there for long because it's like the CPU: Boost clock causes more power consumption, temperature and noise.

3. Memory clock: again, highly dependent on the specific model. Higher clock = higher bandwidth, but generally memory bus width and memory type is more important.

4. Texture Fill Rate: OK, I'm no expert, but this is pretty much as straightforward as it sounds. The more powerful your GPU is, the more textures it can process in the same amount of time. It has become murky as of late; for example, the GTX 750 Ti is pretty meager on paper, but it performs far better than it should theoretically. Remember that the rates you're talking about here are usually some number x number of processors x clockspeed. keep in mind that the processor we're talking about here are the ROPs and TMUs, NOT the shaders! (can't remember which is which for texel and fill)

5. Memory bit width: WHen you look at videocard reviews and they show you a shot of the entire PCB, you'll notice that on high end cards, there are more memory modules than on low end cards. Because of this, they have more memory bit width. For example, if a card like the 270X has a 256-bit width, then it has 8 memory chips, each with 32-bit width. 32 x 8 = 256. This is arguably more important than memory clock. This was why the GTX 660 and GTX 660 Ti's 192-bit bus raised some eyebrows; it was highly asymmetrical and people doubted whether the GPU could efficiently use all of the VRAM. This was probably also why they ended up being slower than the 256-bit HD 7870 and 256-bit GTX 760 respectively.

6. Memory bandwidth: some calculation involving the memory bus width and the memory clock. However, keep in mind that this is the bandwidth that exists between the GPU and the graphics memory in your card, and not between the card and the system. That is reliant on the PCIe slot speed/width. Not really a problem for high end cards today in games.

About the x8 and x16, it is rather complicated as well. CPUs and chipsets come with a set amount of "lanes". Each lane is represented as x1 (x16 is 16 lanes, etc.). x16 is faster than x8, but each generation of PCIe is also faster (3.0 is twice as fast as 2.0). Thus, PCIe 3.0 x8 is equal to PCIe 2.0 x16.

Mining/Folding don't require any more than a 2.0 x4 speed at maximum. VRAM may help, but that doesn't justify getting a 4GB GTX 770 over a 2GB model; it's probably more dependent on the GPU itself.

Hope I helped, i'm in kind of a rush

A rush? That post was awesome made lots of sense thanks for the help

[Cubit481]

Yeahhhh, it wouldnt recognize cpu,gpu,vram, etc , so i disabled autofix pr whatever

lol... ...i see u r getting m8...great phone, planning on getting it too...

[Clarkey]

1. VRAM is complicated. More shaders/stuff to process by the GPU in-game means more memory taken up. That doesn't mean that the GPU is fast enough to make use of all of it. For example, a GTX 770 does not need 4GB and cannot use 4GB by itself to its advantage, however, because in SLI the VRAM is cloned, meaning 2 GPUs run off the same amount of memory, 4GB is better because the setup is fast enough to need it. Nothing over 2GB is required for 1080p.

 

Please stop saying this bull. I regularly go over 2 gbs of vram on my 780ti.

[Pomfinator]

Please stop saying this bull. I regularly go over 2 gbs of vram on my 780ti.

Games bloat VRAM usage, if you have more than 2GB of VRAM they will gladly suck it up, but if you don't, they run perfectly fine.

[xAcid9]

Games bloat VRAM usage, if you have more than 2GB of VRAM they will gladly suck it up, but if you don't, they run perfectly fine.

 

More texture popups for u mate.

[Pomfinator]

More texture popups for u mate.

Or you can be watch_dogs and use all of my 3GB of RAM on my 780Ti and still have horrible pop-in.  That game is a mess.

[Guest]

Please stop saying this bull. I regularly go over 2 gbs of vram on my 780ti.

Please stop saying this bull. The GTX 770 does not have 3GB of VRAM like the 780 Ti.

No one told you to get a 780 ti for 1080p gaming in the first place.

[Recon-UK]

Not everything is as straight forward as one may think.

 

Take Fermi, the GTX 480 a 700mhz GPU, then take the GTX 660Ti, if you were to clock the GTX 480 at 660Ti speeds, it would hands down rip the 660Ti a new arsehole, but due to such a hot running GPU, it was much harder to run at a higher clock.

 

 

A GTX 480 only has 480 Cuda Cores, all of which are tremendously faster then Kepler cores, but it's speed and compute power make too much heat and power consumption, condemning it to a lower core clock.

 

 

Here is how a 480 performs when edging close to stock 660Ti core clocks...

 

My own venture into some DIY was the first generation of Nvidia Dx11 GPU's known as Fermi 400 series, i had 2 480's.

 

Heat is a known issue with stock Fermi GPU's, however Nvidia did not do a good job of researching a decent cooler for their 400 series GPU's up until the GTX 460 arrived on the scene.

 

My 480's in SLi reached 102c, i later gave my 2nd 480 to a good friend who made good use of it for 1050p shenanigans, 102c is crazy, but it was my fault, spacing for the cards was literally not there.

 

So, me being the overclocker and tinkerer i am, i wanted MOAR from my single card to make up the difference as much as i can..

 

Here is images of what i did...

 

Cooler off the GPU.

 

 

 

 

Image of the single card with a replaced cooler: GELID Icy Vision REV 2.0

 

 

This was good for stock voltage bios and 850mhz on the core, but i still wanted more.

 

I wanted thermal performance that would come close to a basic water loop.

 

I did a lot of experimenting with different fans of all sizes and types...

 

Xigmatek fans 120mm:

 

 

 

Phobya G silent 120mm:

 

 

 

None of the above was going to cut it, i needed some quality fans which are capable of forcing air in a small space, not overlapping the entire heatsink.

 

 

So i went for Noctua 92mm fans, these were amazing, inaudible and deadly at cooling!

 

 

I have no images for this, but i noticed i was not gaining any OC headroom, this is where i decided to order a Thermalright VRM-G2, which would replace the entire array of tiny heatsinks that covered the VRM's.

 

 

I then proceeded to modify the cards bios with a ridiculous amount of voltage overhead, a serious bios not really meant for air, but gave me the raw card to play with.

 

Tweaked phases and more voltage on tap than the local power grid would earn me the fastest air cooled GTX 480 on overclock.net (OCN)

 

The cooling capacity was pretty insane and impressive in one go... see for yourself:

 

 

 

 

Performance numbers:

 

 

3D Mark 11:

http://www.3dmark.com/3dm11/5654604

 

vs stock 7950:

http://www.pcchip.hr/wp-content/uploads/2012/02/Sapphire-Radeon-HD-7950-3DMark11.jpg

 

 

Unigine Valley run:

 

 

 

Video at a lower clock:

 

 

These next shots are downsampled images at 1440p and higher:

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

The 660Ti and it's measly bus width would cave in sooner than the Fermi GPU with the wider bus, however the 660Ti has an advantage of higher clocked memory, and literally no compute performance, so it's TDP and heat can be much lower as well as focused on pure gaming (Kepler was made this way)

 

Kepler is a redesign of Fermi but with the compute performance tacked off, only until Titan and 780Ti will users get fully fledged Kepler, Kepler also runs on a smaller process helping that all important efficiency.

 

Here is a result of the 660Ti (up to date from the Valley thread here on LTT)

 

 

 

 

Sorry if this is way too much info to take in, but it is the way i understand it, a GPU from yesteryear is not exactly slower, it just is not as efficiently designed, Kepler is nothing big over Fermi for example, it is just a more efficient way of computing.

[Recon-UK]

Games bloat VRAM usage, if you have more than 2GB of VRAM they will gladly suck it up, but if you don't, they run perfectly fine.

[Guest]

Snip

This is really interesting. Fermi stream processors are far superior to those of Kepler, but often they're so hobbled by low clock and high voltage.

So there really isn't a way to make the 480 truly competitive again because the voltage required would instakill the GPU, oui?

[Recon-UK]

This is really interesting. Fermi stream processors are far superior to those of Kepler, but often they're so hobbled by low clock and high voltage.

So there really isn't a way to make the 480 truly competitive again because the voltage required would instakill the GPU, oui?

I killed that 480 through voltage so yes.