[Haun]

I have been looking at computers when today i realized i didnt actually know what was important.

So can some one educate me on what makes a GPU good?

Thanks

[mctw]

It does what it says on the tin. Graphics Processing Unit.
In simple terms, its basically a mini computer, that is purely dedicated to processing raw data from the CPU, and turning it into images. This allows the workload for the CPU to be lessened.

[Haun]

i was thinking important specs... like # of cores, frequency, Vram, stuff like that.

[Razer]

Cores in GPU's are into hundreds. A GPU is just like a processor but designed for graphics. A graphics card would need to have good VRAM and such.

A good graphics card has: High speed VRAM (GDDR3-5) a GPU around the 700MHz-1GHz margin and a good data bandwidth (256/512bit). It also has to have a cooler capable of keeping all the components within a defined range. Newer graphics cards have smaller die size (55nm) which help with power consumption, heat and noise.

[xactly]

There are many things to take into account when considering what is important in a GPU and what we should look out for, here are a few:

ROPs: Also known as Raster Operators apply final effect such as Anti Aliasing, “The more the better� you may think then but you would be wrong. Have too many and you’ll overwhelm the memory interface with data, use too few and they won’t be able to keep up with the data from the cores until the current data is outputted temporarily causing a stall.

Stream Processors: Referred to as Floating Point Unit as they can only handle certain instructions. ATI use Superscalar stream processors whilst nVidia make use of more capable and accurate Scalar stream processors.

VRAM: Also known as Video Random Access Memory is used to store the frame-buffer of 3D applications. Graphics Double Data Rate or GDDR is the name for the memory used by the GPU. GDDR5 is the current, fastest available VRAM which allow for ludicrous figures in terms of Memory bandwidth.

Frequency: Usually measured in MHz or GHz. There are 3 frequencies to take into account: shader clock, core clock and memory clock. In this scenario it is true to say “the higher the better “.

Core clock: This is the main frequency to consider, nVidia tend to have the core clock of their GPUs at around 600-700MHz whilst ATI opt for higher clocks of 700-800MHz. The 4890 to date is the fastest GPU with a core clock of 850MHz whilst Sapphire have managed to reach 1000MHz with the release of their “Atomic 4890�.

Shader clock: Very important in shader heavy games such as Crysis and Oblivion. ATI’ shader clocks are known as “Stream Processors� and therefore operate at the same frequency of the Core as they are “linked�. nVidia’ shader clocks run at much higher rates and are usually more than double the speed of the core clock e.g. a standard 8800GT has a 600MHz Core and a 1500MHz Shader Clock, the shader clock of nVidia’ GPU’ are “unlinked� to the core clock.

Memory Clock: This determines the speed of which the memory runs at to communicate to the GPU and is measured in either MHz or GHz.

GDDR2/3/4: 800MHz (actual) = 1600MHz (effective) as the data is output twice during every clock cycle.

GDDR5: 800MHz (actual) = 3600MHz (effective) as GDDR5 operates with Command Clocks and Write Clocks. These output 4x as much data as that of GDDR3.

Memory Interface: The form of interface between the GPU and the VRAM. The wider the interface the more the data that can be put through, effectively allowing for a faster performing graphics card, however, there are very few scenarios where an interface will run out of bandwidth when compared to the compute power of GPU e.g. The 4870 has a 256-bit interface which very rarely run out of bandwidth because of the compute power (or lack of it). The GTX280 has a 512-bit interface to keep up with the vast amount of data throughput; if it were a 256-bit interface it would cause a bottleneck in terms of performance.

RAMDAC: Random Access Memory Digital to Analog Converter. These convert the image rendered by the graphics card from a digital signal to an analog signal so that the image can be displayed on a monitor. These are found on all graphics cards and usually run at 400MHz.

Die Size: In this case it is the smaller the better as less power will be consumed therefore less heat will be emitted. 45nm is an ideal die size as it will allow for greater overclocking than 65nm or even 90nm as much less heat is produced. However, the die size does not have a direct effect on performance and should only be viewed when conscious about heat or power.


When considering a graphics card you should not look at specific areas of the graphics card but as an overall package and how well it performs in the real-word. A good GPU is the right ratio of all of the mentioned criteria which come together at a reasonable price, The GTX275 and HD4870 are examples of this.