Joined
·
12,389 Posts
Basic GPU Choosing
I thought I would create a thread that would have a huge amount of compiled information and be able to help people understand what they really need in a graphics card. Hopefully this will help cut down the huge amount of "Do I need 1.5GB or 2GB of VRAM!?!?!?" threads. My goal is that by the end of reading all this, most, people will have a fair idea of what a graphics card really does and what they need for their personal uses.
What this is not intended to be in any way shape or form: AN AMD VS NVIDIA THREAD.
Words marked with * have notes explaining them at the end.
PART ONE - Info
First off, what makes up a graphics card. What exactly separates a GTX580 from a Radeon 6970? Listed below are basic descriptions of what each individual component of a card does. These were complied via Google for the sake of having it all in one place.
Core Clock
The actual speed at which the graphics processor on a video card operates. Core clock is measured in megahertz (MHz). The core clock speed can sometimes be changed on newer cards where users want to gain a performance boost. This is called overclocking and it can usually be done using third-party utilities or the drivers provided by the video card manufacturer.
LINK
Memory Clock
The memory clock, along with the size of the memory bus, tells us the amount of memory bandwidth a graphics card has. The more memory bandwidth a card has, the better it can handle higher resolutions and high levels of AA and AF.
Memory comes in several different varieties, most of which are some form of DDR (double data rate). DDR memory can read from/write to memory twice every clock cycle. So if your DDR memory clock is 500MHz, the effective clock speed is 1000MHz. You'll notice, for cards that use DDR memory, the card pages show both the actual clock speed and the effective speed of the memory.
Like the core clock, the memory clock of most cards can be manually increased through the driver.
LINK
Shader Clock
The core clock runs some functions on the multiprocessor level, like the instruction decoder, and the shader clock runs the individual processors. The shader clock is the fastest of the two, and this sets the speed of arithmetic operations by the processor.
LINK
Pixel Pipelines
The part of a video card that transfers pixel information. The amount of pixel pipelines a graphics card has can have a great impact on the speed of the image rendering. The more pixel pipelines, the faster the video card can process pixels. A card with eight pipelines can process twice as many pixels as a card of the same speed with four pipelines.
LINK
Stream Processors
A stream processor is one of many (sometimes hundreds) parallel processing units that is part of the architecture of modern day graphics processors. The importance of it being parallel is stressed because a GPUs ultimate function is to continuously render pixel (millions of them at a time) imagery/content based on requests from applications and games. With an army of parallel units performing an enormous amount of similar incoming calculation requests (whether FP or INT arithmetic), as long as there are more requests than there are such processing units to handle them, the more stream processors, the better. General purpose CPUs are designed completely differently, as it can perform only so many operations in parallel - most CPUs (probably all) only have one (1) arithmetic logic unit (ALU) that handles all of the arithmetic such as Add, Subtract, Multiply, Divide, Bit Shift, and more, for FP and INT, whereas a typical GPU consists of hundreds of ALU-like units (what we call stream processors) designed for a more limited set of FP/INT calculations.
LINK
VRAM
Short for video RAM, and pronounced vee-ram. VRAM is special-purposememory used by video adapters. Unlike conventional RAM, VRAM can be accessed by two different devices simultaneously. This enables the RAMDACto access the VRAM for screen updates at the same time that the videoprocessor provides new data. VRAM yields better graphics performance but is more expensive than normal RAM.
LINK
PART 2 - What you need
Ok cool story bro that's a lot of reading, now why should I choose certain cards over others?
There are a few basic questions you need to ask yourself before purchasing a graphics card.
1. What are you going to be using your rig for?
2. What's your resolution?
3. What's your budget?
4. What motherboard have you chosen?
5. Lastly and least importantly, are aesthetics important to you?
Starting with the first question, rigs can be used for many many things. Gaming. Video editing. General web browsing. Mass processing (Folding for example).
The second question is mostly concerning how much VRAM you are going to want, but has to do with general power as well. Obviously you're going to need a lot more powerful card(s) to game at 5760x1200 than to game at 1024x768.
The third question is a big determinant for well everything. If you can afford a nice motherboard and PSU, you can buy more, bigger graphic cards. You must also balance it with a nice CPU to avoid bottlenecking*. Basically the more you spend on a graphics card, the more you're going to have to spend on everything else.
The fourth question goes mainly with question three. The spacing of a motherboards PCI-E slots is very important if you plan on SLi or CF. If you buy single slot cards this obviously isn't a big factor, but if you are buying a dual slot card (which is most common) then you are going to have to plan accordingly if you want to fit more cards in. Then there are triple slot monsters, you really have to plan well to fit them babies.
Lastly the fifth question is only if there are multiple cards that match what you need and you are stuck choosing between manufacturers. Choose what you like and what matches the rest of your system because eventually you will probably end up wanting to match everything up, even if you don't think you will now.
All of these questions go hand in hand with each other so I'm going to try to create a basic flow chart to describe the choosing process.
I'm sure many opinions will be shared about thisflow chart but this is just general options!
Following information about workstation cards is thanks to chrisguitar. You can thank him at the link
Entry level (comfortable performance):
Quadro 400 and 600
Mid range (excellent performance):
Quadro 2000 and 2000D
High end (For professionals):
Quadro 4000 and 5000
Ultra high end (For people who are crazy ;D)
Quadro 6000 and Plex 7000
Then you have the Teslas's which are insane! They are used for Reservoir simulation, Molecular Dynamics. Numerical Analytics, Computational Visualization (ray tracing)
To break it down though, while using a Quadro 2000 on BFBC2 I was on all medium settings with 4x AA and AF and no AO running at; (with dxtory)
Min: 45 FPS (under a lot of stress with explosions and smoke)
Avg: 80 FPS (no smoke but moving)
Max: 130 FPS (not moving at all)
I used my Xeon cpu as well so image with an i5 or i7, it would be great!
In summary Quadro's are for professionals, gaming is playable but not a max settings @ high fps @ high resolution.
If you're just going to be using your computer for general computing; office, internet, email, very light gaming, etc. I wouldn't recommend a card above a GTS 450. These are pretty solid cards, they can OC to 950Mhz without any problems and they play games decently. Solid cards for the cheap price.
For [email protected], it all depends on how much PPD you want. nVidia cards fold better than AMD/ATi cards, so more PPD you want, the higher end nVidia card you choose, it's that simple.
PART 3 - VRAM
How much VRAM do I really need!?
First, to get one fact straight. Just because you can SLi/CF multiple cards, does not mean it adds to the VRAM. The card with the lowest amount of VRAM, is the amount is usable. This applies with dual gpu cards as well.
4x GTX 580s with 1.5GB of VRAM =/= 6GB of RAM! Only 1.5GB.
1x GTX 590 "3GB" really is two 1.5GB GTX580s so you do not have 3GB usable, only 1.5GB! <-- this is very important to understand.
If you SLi a 1GB 460 and a 768MB 460, only 768MB will be usable!
Same applies with CF and the 6990.
So how much do you need...more the merrier right?
Not entirely, results have shown that overkill VRAM can sometimes hinder performance slightly. For example better results were found when a 1.5GB 580 was used at 1080p compared to when a 3GB 580 was used at 1080p.
As a general rule of thumb you can use this list below, but do keep in mind some games use a lot more VRAM then others so this is just a general idea.
Surround/Eyefinity --> 3GB+
1600p/1440p-->3GB
1200p/1080p -->2GB-1.25GB
<1080p -->1GB
PART 4 - Budget
Of course a huge factor in choosing cards is, how much money do you have?
With computers everything gets exponentially expensive. You buy a bigger CPU, you'll probably get a bigger GPU, and then a better motherboard, and then a bigger case, etc etc.
You can't focus to much money on one component because then you'll experience bottle necking, which do you not want. Try to keep it even. If you buy a 990X or 2600K, buy some 580s or 6970s. If you buy some mid ranged processor, get a mid ranged GPU, etc.
Watercooling is expensive so you have to take that into account when purchasing as well.
PART 5 - Cooling & Spacing
By now, you should have hopefully gotten your huge options of graphic cards a lot smaller and more limited. Moving on to another factor you must take into consideration, is cooling.
Now you must decide how you're going to cool your card(s).
Right off the bat if you're doing a build with 3 or 4 cards I would recommend watercooling.
With one or two cards you can do either or.
Now you have a choice to make, you can either buy reference** cards or non-reference cards. Non-reference cards generally come overclocked right off the bat and have a better cooling system. However, waterblocks are for the most part made for reference cards.
If you plan on buying an aftermarket air cooler like the Accelero series you should be shopping for reference cards as well.
Now is when you also need to make some motherboard decisions. What you need to pay attention to is spacing of the PCI-E slots. If you plan on air cooling then it's best to leave slots/spacing in between each card. So this requires some planning on your part to make sure you have adequate spacing and room to fit your cards on your mobo.
For watercooling the main idea that you need to know is that for radiator size you're going to want at least one 120mm rad per GPU (240mm rad for dual gpu cards).
*Bottlenecking; occurs when the GPU is not fully being utilized because your CPU is not fast enough to be able to keep up. You can determine this by looking at GPU load in Afterburner while playing a game without v-sync on.
**Reference & non-reference. Reference cards are based off AMD/nVidia's original design. Non-reference cards are redone by the manufacturers to perform better or to fit their heatsink/fan setup.
Additional Info:
Graphics Card Ranking - Made by Anth0789
PSU Calc made by Phaedrus2129 & FiX
Small PPD List (Anyone find a better one please let me know)
Everything on here was written myself except the beginning information which is linked to the source and the bit added by chrisguitar. I would love to hear all comments, questions, suggestions and feedback! Unless it's bad feedback then I'd appreciate a PM instead.
Thanks all and I hope this helps.
I thought I would create a thread that would have a huge amount of compiled information and be able to help people understand what they really need in a graphics card. Hopefully this will help cut down the huge amount of "Do I need 1.5GB or 2GB of VRAM!?!?!?" threads. My goal is that by the end of reading all this, most, people will have a fair idea of what a graphics card really does and what they need for their personal uses.
What this is not intended to be in any way shape or form: AN AMD VS NVIDIA THREAD.
Words marked with * have notes explaining them at the end.
PART ONE - Info
First off, what makes up a graphics card. What exactly separates a GTX580 from a Radeon 6970? Listed below are basic descriptions of what each individual component of a card does. These were complied via Google for the sake of having it all in one place.
Core Clock
The actual speed at which the graphics processor on a video card operates. Core clock is measured in megahertz (MHz). The core clock speed can sometimes be changed on newer cards where users want to gain a performance boost. This is called overclocking and it can usually be done using third-party utilities or the drivers provided by the video card manufacturer.
LINK
Memory Clock
The memory clock, along with the size of the memory bus, tells us the amount of memory bandwidth a graphics card has. The more memory bandwidth a card has, the better it can handle higher resolutions and high levels of AA and AF.
Memory comes in several different varieties, most of which are some form of DDR (double data rate). DDR memory can read from/write to memory twice every clock cycle. So if your DDR memory clock is 500MHz, the effective clock speed is 1000MHz. You'll notice, for cards that use DDR memory, the card pages show both the actual clock speed and the effective speed of the memory.
Like the core clock, the memory clock of most cards can be manually increased through the driver.
LINK
Shader Clock
The core clock runs some functions on the multiprocessor level, like the instruction decoder, and the shader clock runs the individual processors. The shader clock is the fastest of the two, and this sets the speed of arithmetic operations by the processor.
LINK
Pixel Pipelines
The part of a video card that transfers pixel information. The amount of pixel pipelines a graphics card has can have a great impact on the speed of the image rendering. The more pixel pipelines, the faster the video card can process pixels. A card with eight pipelines can process twice as many pixels as a card of the same speed with four pipelines.
LINK
Stream Processors
A stream processor is one of many (sometimes hundreds) parallel processing units that is part of the architecture of modern day graphics processors. The importance of it being parallel is stressed because a GPUs ultimate function is to continuously render pixel (millions of them at a time) imagery/content based on requests from applications and games. With an army of parallel units performing an enormous amount of similar incoming calculation requests (whether FP or INT arithmetic), as long as there are more requests than there are such processing units to handle them, the more stream processors, the better. General purpose CPUs are designed completely differently, as it can perform only so many operations in parallel - most CPUs (probably all) only have one (1) arithmetic logic unit (ALU) that handles all of the arithmetic such as Add, Subtract, Multiply, Divide, Bit Shift, and more, for FP and INT, whereas a typical GPU consists of hundreds of ALU-like units (what we call stream processors) designed for a more limited set of FP/INT calculations.
So what's the difference between Nvidia's and ATI's GPU architecture and why do seemingly comparable ATI cards have more stream processors than Nvidia ones? The answer lies in their different approach in implementation. Nvidia's GPUs are flooded with fewer stream processors (CUDA technology), where each one is identical in look, feel, design, and function (FP and INT arithmetic) to its neighbor. To be more exact, for every 8 identical stream processors, there is one special functional unit that keeps things in check. So if you look at a Geforce GTX 280 with 240 stream processors, it's really using only about 88% (1 of every 9 sp's are there police the other eight) of its advertised FP/INT arithmetic processing power.
Nonetheless, Nvidia's GPU architecture is easier for application and games developers to program for due to its simplicity (every stream processing unit performs the same function)- as long as the units are fed numbers to crunch from the apps, you are getting fast raw results every clock cycle. Nvidia's architecture has been deemed analogous to American moto engines- simple raw power, gas guzzling.
ATI's architecture is a bit different- not every stream processor (Brook+ technology) is identical to its neighbor. For every block of 6 stream processing units, 4 are identical, the 5th carries different FP/INT arithmetic functions, and the 6th keeps things in check. So essentially, each block of 5 ATI stream processors (ignoring the special unit) is comparable to 1 Nvidia stream processor. The math isn't that simple, but its a good generalization to make that helps demystify why a high-end ATI Radeon HD 4870 card with a rocking 800 stream processors is relatively weaker than an Nvidia GTX 280 with only 240. Because of ATI's GPU architecture, apps and game developers have a tougher time programming to take full advantage of every stream processor on board the fact that specific FP/INT arithmetic functions can only be "worked on" by one out of every five units (per block). In order to take full advantage of ATI's architecture, an app or game must be optimally coded- something like baiting the hook to suit the fish..
If a program is not optimized for the architecture, the work, to ultimately have as many blocks of stream processing units working every cycle of the clock, relies on the GPU scheduler - the Ultra Threaded Dispatch Processor. All in all, current Nvidia graphics cards lead ATI implementations in most (if not all) game benchmarks, but from a cost/performance standpoint, ATI is definitely the better bang for the buck. Choose wisely.
Nonetheless, Nvidia's GPU architecture is easier for application and games developers to program for due to its simplicity (every stream processing unit performs the same function)- as long as the units are fed numbers to crunch from the apps, you are getting fast raw results every clock cycle. Nvidia's architecture has been deemed analogous to American moto engines- simple raw power, gas guzzling.
ATI's architecture is a bit different- not every stream processor (Brook+ technology) is identical to its neighbor. For every block of 6 stream processing units, 4 are identical, the 5th carries different FP/INT arithmetic functions, and the 6th keeps things in check. So essentially, each block of 5 ATI stream processors (ignoring the special unit) is comparable to 1 Nvidia stream processor. The math isn't that simple, but its a good generalization to make that helps demystify why a high-end ATI Radeon HD 4870 card with a rocking 800 stream processors is relatively weaker than an Nvidia GTX 280 with only 240. Because of ATI's GPU architecture, apps and game developers have a tougher time programming to take full advantage of every stream processor on board the fact that specific FP/INT arithmetic functions can only be "worked on" by one out of every five units (per block). In order to take full advantage of ATI's architecture, an app or game must be optimally coded- something like baiting the hook to suit the fish..
If a program is not optimized for the architecture, the work, to ultimately have as many blocks of stream processing units working every cycle of the clock, relies on the GPU scheduler - the Ultra Threaded Dispatch Processor. All in all, current Nvidia graphics cards lead ATI implementations in most (if not all) game benchmarks, but from a cost/performance standpoint, ATI is definitely the better bang for the buck. Choose wisely.
VRAM
Short for video RAM, and pronounced vee-ram. VRAM is special-purposememory used by video adapters. Unlike conventional RAM, VRAM can be accessed by two different devices simultaneously. This enables the RAMDACto access the VRAM for screen updates at the same time that the videoprocessor provides new data. VRAM yields better graphics performance but is more expensive than normal RAM.
LINK
PART 2 - What you need
Ok cool story bro that's a lot of reading, now why should I choose certain cards over others?
There are a few basic questions you need to ask yourself before purchasing a graphics card.
1. What are you going to be using your rig for?
2. What's your resolution?
3. What's your budget?
4. What motherboard have you chosen?
5. Lastly and least importantly, are aesthetics important to you?
Starting with the first question, rigs can be used for many many things. Gaming. Video editing. General web browsing. Mass processing (Folding for example).
The second question is mostly concerning how much VRAM you are going to want, but has to do with general power as well. Obviously you're going to need a lot more powerful card(s) to game at 5760x1200 than to game at 1024x768.
The third question is a big determinant for well everything. If you can afford a nice motherboard and PSU, you can buy more, bigger graphic cards. You must also balance it with a nice CPU to avoid bottlenecking*. Basically the more you spend on a graphics card, the more you're going to have to spend on everything else.
The fourth question goes mainly with question three. The spacing of a motherboards PCI-E slots is very important if you plan on SLi or CF. If you buy single slot cards this obviously isn't a big factor, but if you are buying a dual slot card (which is most common) then you are going to have to plan accordingly if you want to fit more cards in. Then there are triple slot monsters, you really have to plan well to fit them babies.
Lastly the fifth question is only if there are multiple cards that match what you need and you are stuck choosing between manufacturers. Choose what you like and what matches the rest of your system because eventually you will probably end up wanting to match everything up, even if you don't think you will now.
All of these questions go hand in hand with each other so I'm going to try to create a basic flow chart to describe the choosing process.

I'm sure many opinions will be shared about thisflow chart but this is just general options!
Following information about workstation cards is thanks to chrisguitar. You can thank him at the link

Entry level (comfortable performance):
Quadro 400 and 600
Mid range (excellent performance):
Quadro 2000 and 2000D
High end (For professionals):
Quadro 4000 and 5000
Ultra high end (For people who are crazy ;D)
Quadro 6000 and Plex 7000
Then you have the Teslas's which are insane! They are used for Reservoir simulation, Molecular Dynamics. Numerical Analytics, Computational Visualization (ray tracing)
To break it down though, while using a Quadro 2000 on BFBC2 I was on all medium settings with 4x AA and AF and no AO running at; (with dxtory)
Min: 45 FPS (under a lot of stress with explosions and smoke)
Avg: 80 FPS (no smoke but moving)
Max: 130 FPS (not moving at all)
I used my Xeon cpu as well so image with an i5 or i7, it would be great!
In summary Quadro's are for professionals, gaming is playable but not a max settings @ high fps @ high resolution.
If you're just going to be using your computer for general computing; office, internet, email, very light gaming, etc. I wouldn't recommend a card above a GTS 450. These are pretty solid cards, they can OC to 950Mhz without any problems and they play games decently. Solid cards for the cheap price.
For [email protected], it all depends on how much PPD you want. nVidia cards fold better than AMD/ATi cards, so more PPD you want, the higher end nVidia card you choose, it's that simple.
PART 3 - VRAM
How much VRAM do I really need!?
First, to get one fact straight. Just because you can SLi/CF multiple cards, does not mean it adds to the VRAM. The card with the lowest amount of VRAM, is the amount is usable. This applies with dual gpu cards as well.
4x GTX 580s with 1.5GB of VRAM =/= 6GB of RAM! Only 1.5GB.
1x GTX 590 "3GB" really is two 1.5GB GTX580s so you do not have 3GB usable, only 1.5GB! <-- this is very important to understand.
If you SLi a 1GB 460 and a 768MB 460, only 768MB will be usable!
Same applies with CF and the 6990.
So how much do you need...more the merrier right?
Not entirely, results have shown that overkill VRAM can sometimes hinder performance slightly. For example better results were found when a 1.5GB 580 was used at 1080p compared to when a 3GB 580 was used at 1080p.
As a general rule of thumb you can use this list below, but do keep in mind some games use a lot more VRAM then others so this is just a general idea.
Surround/Eyefinity --> 3GB+
1600p/1440p-->3GB
1200p/1080p -->2GB-1.25GB
<1080p -->1GB
PART 4 - Budget
Of course a huge factor in choosing cards is, how much money do you have?
With computers everything gets exponentially expensive. You buy a bigger CPU, you'll probably get a bigger GPU, and then a better motherboard, and then a bigger case, etc etc.
You can't focus to much money on one component because then you'll experience bottle necking, which do you not want. Try to keep it even. If you buy a 990X or 2600K, buy some 580s or 6970s. If you buy some mid ranged processor, get a mid ranged GPU, etc.
Watercooling is expensive so you have to take that into account when purchasing as well.
PART 5 - Cooling & Spacing
By now, you should have hopefully gotten your huge options of graphic cards a lot smaller and more limited. Moving on to another factor you must take into consideration, is cooling.
Now you must decide how you're going to cool your card(s).
Right off the bat if you're doing a build with 3 or 4 cards I would recommend watercooling.
With one or two cards you can do either or.
Now you have a choice to make, you can either buy reference** cards or non-reference cards. Non-reference cards generally come overclocked right off the bat and have a better cooling system. However, waterblocks are for the most part made for reference cards.
If you plan on buying an aftermarket air cooler like the Accelero series you should be shopping for reference cards as well.
Now is when you also need to make some motherboard decisions. What you need to pay attention to is spacing of the PCI-E slots. If you plan on air cooling then it's best to leave slots/spacing in between each card. So this requires some planning on your part to make sure you have adequate spacing and room to fit your cards on your mobo.
For watercooling the main idea that you need to know is that for radiator size you're going to want at least one 120mm rad per GPU (240mm rad for dual gpu cards).
*Bottlenecking; occurs when the GPU is not fully being utilized because your CPU is not fast enough to be able to keep up. You can determine this by looking at GPU load in Afterburner while playing a game without v-sync on.
**Reference & non-reference. Reference cards are based off AMD/nVidia's original design. Non-reference cards are redone by the manufacturers to perform better or to fit their heatsink/fan setup.
Additional Info:
Graphics Card Ranking - Made by Anth0789
PSU Calc made by Phaedrus2129 & FiX
Small PPD List (Anyone find a better one please let me know)
Everything on here was written myself except the beginning information which is linked to the source and the bit added by chrisguitar. I would love to hear all comments, questions, suggestions and feedback! Unless it's bad feedback then I'd appreciate a PM instead.
Thanks all and I hope this helps.
