Speed
There's no such thing as a single speed number which you can use to compare
how fast different video cards are at games. Video card performance can vary
wildly from one game engine to another. One video card can be faster than
another at one game and then the positions are reversed with another game.
Game performance can vary among different versions of display drivers. GPU
makers tend to improve weak spots in their drivers over time so some cards
get very different scores depending on whether they're tested just after
release or six months later. Some cards are better at multi-texturing while
others are better at pixel shaders. Some are better with full screen
anti-aliasing while others are better with none.
Basically, there's just no way to assign
a single performance number to a video card because its performance varies
so much depending on the situation. But since people want a single speed
number anyway, this column provides one. Just remember that it's an
estimate. The speed number gives the approximate speed of the video card at
3D
games. Higher numbers mean higher performance. If the speed number of one
card is 1.5 times the number of another card then it means that it is about
1.5 times as fast.
This speed number is an average across a wide range of different game
benchmarks gathered from various website reviews. You shouldn't try to
compare the speed numbers of cards from different technology generations
(for example, one card is DirectX 8 and the other
is DirectX 9). AGP and PCI-Express numbers are only roughly comparable
because the tests are run on different hardware and more of the AGP tests use
older games. The benchmarks selected
give emphasis to playable frame rates so the tests for high-end cards are
run at very high screen resolution and image quality. Benchmarks for low-end
cards tend to be lower resolution and image quality. As a result, the speed
ratios between fast cards and slow cards aren't very accurate.
These speed numbers have been derived from a large number of gaming
benchmarks. A few of these benchmarks
are synthetic benchmarks like
3DMark but almost all are taken from
real games. Some game engines which favor one kind of
GPU
or another are included (Doom 3 favors some NVIDIA cards, Half Life 2 favors
some ATI cards) but
they are just a small part of the gaming benchmarks which include a broad
range of different games. These benchmarks are run with high-end
CPUs so the CPU won't be the performance bottleneck. Adding
more benchmark samples can make the speed numbers move around quite a bit.
Having a video card go up or down by 5% is very common and 10% is not unusual
so there's no point in worrying about differences of a few percent between
two cards. The speed number isn't that accurate. Just use the speed number as
an estimate.
Video cards which contain a "?" in the speed column do not have any benchmark
data. Their number is an informed guess based on the numbers of similar cards.
If there are too few cards in the same speed range with benchmark data then
the card is given a "?" with no speed number. High-end
cards tend to get lots of attention from reviewers so they have lots of
available benchmark data. Low-end cards or slowed down models (for example,
use 64 bit video RAM instead of the standard 128
bit) often have little or no benchmark data available. Some cards are too
obscure to be worth the time to compile the data (which takes a lot of
time).
DirectX
This table column gives the version of DirectX that this video card was
designed to support. This tells you the basic features of the video card.
This has nothing to do with the latest version of DirectX
display driver
you can get for the video card. You can get DirectX 9 drivers for DirectX 7
and 8 video cards but they don't have hardware support for the new features
which were added in DirectX 9. You can read more about the 3D capabilities of
various DirectX versions on this page.
- 7.0 -
Supports vertex-lit, multi-textured, bump-mapped models and introduced
hardware T&L.
- 8.0 -
First version of DirectX which supported pixel and vertex shaders. Some
useful shader functionality was still missing so this generation of
shaders were still not general purpose shaders.
- 8.1 -
Added support for PS 1.3 and
PS 1.4 which were designed to extend DirectX 8
pixel shader capabilities for certain video card families.
- 9.0 -
DirectX 9 cards delivered higher color precision and solved a few other
problems with pixel shaders. DirectX 9 is considered to be the first
flexible and general purpose implementation of shaders.
- 10.0 -
Reduced GPU dependence on the CPU to get many jobs done. As a result,
scenes with far more visible objects will be possible. ATI and NVIDIA
will both use unified shader architectures which make sure that the
shader hardware is fully utilized for any kind of 3D scene. That's as
opposed to previous versions of DirectX where part of the shader hardware
was often idle for some kinds of scenes.
- 10.1 -
Added support for SM 4.1. This is a minor update to
DirectX 10.0 which primarily adds
more flexibility to image quality control.
Max watts
This table column provides the approximate peak power consumption of the video
card. This is the wattage which has to be delivered by the computer's power
supply to the video card. This is not the wattage load delivered by the AC
power cord to the computer. The AC wattage is a higher number because power
supplies are not 100% efficient. If the video card has more than one possible
configuration then values between "[" and "]" tell which configuration is
listed. The first value is the GPU clock, the
next is the RAM size, and the last is the
RAM clock. If that video card has only one value
listed in its table column (the only GPU clock is 600 MHz), then the value
is left out. For example, "[,128,600]" means that the wattage is for a video
card using the only GPU clock value, a RAM size of 128 MB, and a memory clock
of 600 MHz.
The power consumed by two apparently identical video cards can vary somewhat.
The websites which measured these values only measured one card. You
also need to take into account that different versions of the same video
card sometimes use different kinds of video RAM (DDR, GDDR2, GDDR3) so the
RAM power consumption varies somewhat from card to card. Just remember that
these wattage numbers are an approximation.
Unfortunately, there is not that much accurate information available for the
power consumption of video cards. A few websites measure the total power
consumption of video cards by using AC clamp meters. These
measure the current going through an AC power cord to within a few percent.
This is a very easy way of measuring power. But it's not very good at telling
you how much wattage the video card will draw from your power supply. First
of all, you have to remember that the computer's power supply is not 100%
efficient. 75% efficiency is a normal value but it varies depending on the
power supply and the load. For a 75% efficient power supply, the wattage
delivered on the outputs of the power supply to the hardware in the computer
is only 75% of the power reported by the clamp meter. Then you need to take
into account how they measure idle and peak loads. The most common practice
is to measure the clamp meter wattage when the computer is idling. That
usually is the wattage drawn while the computer is sitting at the Windows
desktop running no programs. Then they measure the peak wattage of the video
card while running a game with high image quality settings to push the video
card as hard as possible. But the CPU load is also substantially higher when
running a video game than when idling at the desktop and other parts of the
hardware consume more power as well. They end up measuring
both the increase in power consumption of the video card (which is probably
loaded 100%) as well as the increase in power consumption of the CPU (which
may not be loaded at 100% but is still far more loaded than when idling at
the desktop). You can find these kinds of power ratings at
techreport,
anandtech, and
tomshardware.
You can also read about them (if you're good with
Babel Fish) on pages like
this,
this,
this,
this,
this,
this,
and this.
What you'd really like to know is the peak wattage used only by the video
card. That's what the power supply actually has to deliver to the video card.
Fortunately, some enterprizing Russians at
xbit labs have gone to
the trouble of measuring the actual wattage used by many video cards. It's
hardly surprising that they're the only ones who do it considering that it
involves some soldering. You can read their main video card power articles
here,
here,
here, and
here.
These are the most accurate ratings available because they're measuring the
peak current delivered by the power supply to the video card. If a video card
has an xbitlabs wattage measurement then that value is given much higher
weight in
the final outcome than other measurements. Video cards which have xbitlabs
measurements are shown in the wattage column using bold
green characters.
The wattages shown which are not bold green
characters do not have xbitlabs measurements and are
approximations. Most of them are filled in by using
the AC wattage measurements and making an adjustment for power supply
efficiency and extra CPU load. So basically, those numbers are educated
guesses. Even direct measurements on video cards vary a bit because silicon
chips are not identical and their power consumption can vary. So basically,
just use these power wattage numbers as a rough estimate. They're not perfect
but they can give you a good idea of about how much a given video card will
load your power supply.
Outputs
This table column provides the video outputs included on the standard
configuration for this video card. "2 X" means that there are two copies of
that kind of output. Values separated by "&" means that the video card has
all of the outputs listed.
- VGA - Includes a standard VGA output.
- DVI-I - Includes a single link DVI-I output.
- DL-DVI-I - Includes a dual link DVI-I output.
- DVI-D - Includes a single link DVI-D output.
- DL-DVI-D - Includes a dual link DVI-D output.
- XFIRE - Includes a single link ATI
CrossFire output. A CrossFire cable can convert this into a single link
DVI-I output.
- DL-XFIRE - Includes a dual link ATI
CrossFire output. A CrossFire cable can convert this into a dual link
DVI-I output.
- AIW-VGA - Includes an ATI All-In-Wonder output
which can be converted into a VGA output using an adapter cable.
- AIW-VGA2 - Includes an ATI All-In-Wonder output
which can be converted into two VGA outputs using an adapter cable.
- HDMI - Includes an HDMI output.
Aux power
This table column lists any auxiliary power connectors required to power the
card. This is used by video cards which need more power than can be drawn
through the
motherboard
expansion slot.
Noise
This table column lists the amount of noise made by the card's cooling
system when running at peak load for the standard cooling configuration. Some
video cards have custom cooling solutions which make different amounts of
noise. Many video cards adjust the speed of the
cooling fan based on the load and are relatively quiet when running anything
other than heavy
3D
drawing loads. This column lists the sound of the card at its loudest which
happens most commonly when running games. The amount of noise which is
considered annoying varies from person to person so this value represents the
most common opinion expressed by reviewers.
- Okay - This card is quiet enough that it won't bother
most people.
- Loud - This card is loud enough to be annoying.
- Ouch - This card is so loud that it should be bundled with
earplugs.
Notes
| Note |
Explanation |
| 2S |
The standard implemention of this video card is wide enough that you
can't put an expansion card in the slot next to the video card. This is
usually because the video card has a very thick cooling system.
|
| 3S |
The standard implemention of this video card is so wide that it takes
up three slots. This is because the video card has a very thick cooling
system.
|
| D1 |
This video card is actually two video cards operating as a single unit.
It's functionally the same as running two NVIDIA cards in
SLI
mode but
it is physically just one large video card.
|
| DUAL |
This is a speed test of two cards running together in
SLI or
CrossFire
mode. The gaming benchmarks selected for pairs of cards are very
high-resolution and high image quality tests. People run pairs of
cards together to improve performance under those conditions. Under
less challenging conditions, pairs of cards usually don't improve
performance very much. Pairs of cards also don't usually improve
performance in games which haven't been configured to support dual
cards. The game
benchmarks included in the Speed column only
include games which properly support dual cards. Games which don't
support dual setups are not included.
The specifications for each video card can be
found from the "card" link. If this kind of card has a range of possible
specifications (for example, it has a range of possible
GPU clocks) then the value of the tested
cards is listed in the appropriate column instead of a "card" link.
|
| FX1 |
The NVIDIA GeForce 5 generation of video cards have a complicated
shader architecture. NVIDIA has not released specifics about its internal
design. There's more than one way to characterize the pixel shader
hardware. Articles on the subject often have titles like
"Deciphering the pipeline mess" or "GeForce FX shader mysteries". As a
result, the number displayed in the
GPU hardware column reflects the
performance of the
GPU
relative to other video cards in its class. That value may differ from
values you find in other video card tables. Comparing the number of
shaders between different architectures is often risky, but in the
GeForce 5 series of cards, it's especially so.
|
| FX2 |
The GeForce FX 5200 and 5500 cards have abysmally bad pixel shader
performance. They are listed here as having one pixel shader but zero
would
probably have been closer to their real performance. Technically, they
qualify as DirectX 9 cards. But for games, their pixel shader
performance is so slow that you really have to treat them as DirectX 8
cards. Reportedly, they are fast enough to get serviceable performance
from the Aero user interface in Windows Vista.
|
| HDCP |
This video card usually supports HDCP. Some manufacturers cut costs by
leaving out the crypto-ROM needed for HDCP support so you should check
the specific version of a video card to make sure HDCP is supported.
Video card models which don't have this note usually don't support
HDCP.
|
| HM1 |
This video card implements ATI HyperMemory. It has a total useable
amount of RAM which is the sum of both the video RAM actually on the
video card plus RAM it borrows from the motherboard. The total useable
RAM may be limited by the amount of RAM on the motherboard. The
following table gives you the total useable RAM for each combination of
actual video RAM on the video card and amount of RAM on the
motherboard. The video card manufacturers tend to put out very
confusing descriptions of how much total useable RAM is available for a
given video card (presumably because it varies depending on the amount
of motherboard RAM). ATI hasn't been very specific either so the
numbers below are a best guess.
|
32 MB actual video RAM |
64 MB actual video RAM |
128 MB actual video RAM |
256 MB actual video RAM |
| 256 MB motherboard RAM |
128 MB total useable |
128 MB total useable |
128 MB total useable |
256 MB total useable |
| 512 MB motherboard RAM |
128 MB total useable |
256 MB total useable |
256 MB total useable |
256 MB total useable |
| 1024 MB motherboard RAM or more |
128 MB total useable |
256 MB total useable |
512 MB total useable |
512 MB total useable |
|
| TC1 |
This video card implements NVIDIA TurboCache memory. It has a
total useable amount of RAM which is the sum of both the video RAM
actually on the video card plus RAM it borrows from the motherboard.
The total useable RAM may be limited by the amount of RAM on the
motherboard. The following table gives you the total useable RAM for
each combination of actual video RAM on the video card and amount of
RAM on the motherboard. The video card manufacturers tend to put out very
confusing descriptions of how much total useable RAM is available for a
given video card (presumably because it varies depending on the amount
of motherboard RAM). NVIDIA's TurboCache implementation has many
variations and their informatation hasn't been very specific either so
the numbers below are a best guess.
|
16 MB actual video RAM |
32 MB actual video RAM |
64 MB actual video RAM |
128 MB actual video RAM |
256 MB actual video RAM |
| 256 MB motherboard RAM (below the 512 MB minimum recommended by NVIDIA) |
64 MB total useable |
64 MB total useable |
128 MB total useable |
128 MB total useable |
256 MB total useable |
| 512 MB motherboard RAM |
128 MB total useable |
128 MB total useable |
256 MB total useable |
256 MB total useable |
256 MB total useable |
| 1024 MB motherboard RAM or more |
128 MB total useable |
128 MB total useable |
256 MB total useable |
512 MB total useable |
512 MB total useable |
|
| R1 |
This video card has a model which has 512 MB of
video RAM.
If you're considering gaming performance then the 512 MB model is only
somewhat faster than the 256 MB model under most circumstances.
|
| R2 |
This video card has
video RAM
which is either
64 bits or 128 bits wide. The 64 bit models
are very slow but are rarely much cheaper than the much faster 128 bit
models. Avoid the 64 bit models like the plague.
This video card can have anywhere from 128 to 512 MB of video RAM. The
128 MB models are rarely much cheaper than the 256 MB models and this
video card can derive some benefit from having 256 MB so it's better to
get the 256 MB model. If you're considering gaming performance then
there's very little benefit in getting the 512 MB model. Unfortunately,
there are some 512 MB cards which also have 64 bit video RAM which
slows the cards down a lot.
|
| R3 |
This video card has a model which has 512 MB of
video RAM.
If you're considering gaming performance then there's very little point
in getting more than 256 MB because this video card is too slow to
benefit much at all from more video RAM.
|
| PCIe2 |
This
GPU
supports PCI-Express 2.0. If this note doesn't appear for a PCI-Express
GPU then it supports PCI-Express 1.
|
GPU
This table column contains the name of the
GPU
used in this graphics card. You can use this name to search on the Internet
for more technical information about it. Some video cards can be made with
more than one GPU. 2 X means that it has two
GPUs.
GPU clock
This table column contains the
GPU
clock rate in MHz. The clock rate is the value for the standard
implementation of the video card. Some models may provide different clock
rates. If this video card commonly has a short list of possible GPU clock
rates then they are listed separated by ","s. If this video card commonly has
a range of clock rates then the table column contains the low end and high
end of the commonly used range separated by a "..". Some GPUs have separate
clocks for various parts of the GPU. The value listed in this column is the
main GPU clock rate. If any parts of the GPU operate at different clock rates
then they will be shown in the advanced GPU
hardware column. Many video cards automatically reduce this clock rate
when the video card is not heavily loaded to make the card easier to cool.
This table column reports the maximum clock rate.
GPU hardware
This table column lists major computational units of the GPU like hardware
T&L, pixel shaders, and vertex shaders. When the GPU contains more than one
of these units then it has the number of units, a ":", and the kind of unit.
If the unit(s) operates at a different frequency than the
GPU clock then it has an "@" followed by the
clock rate in
MHz.
For example, "12:PS3.0" means that there are 12 PS3.0 pixel shaders running
at the GPU clock rate. "8:VS3.0@550" means that
there are 8 VS3.0 vertex shaders running at 550 MHz. Hardware T&L does not
have any count or speed rating. It's presence is noted only.
These numbers are included only for comparisons within the same GPU family.
You cannot compare the numbers or clock rates of units between different
kinds of GPUs. For example, under most circumstances, one PS2.0 pixel shader
from ATI's R3xx GPU family was quicker than one PS2.0 pixel shader from
NVIDIA's NV3x family running at the same clock rate. ATI's did more
useful work at a given clock rate than NVIDIA's. You can't tell anything
useful by comparing the number or clock rate of PS2.0 shaders from those two
different families of GPUs. These numbers are included only for comparisons
within the same GPU family. ATI GPU names are "R#xx" or "RV#xx" where "#" is
the family number and the "xx" are various versions within the same family.
NVIDIA GPU names are "NV#x" where "#" is the family number and "x" are
various versions within the same family. Newer NVIDIA GPUs use a "G" instead
of "NV". Sometimes there are changes within a family like between the NV20
and the NV25.
Pixel hardware
- PS1.1 - Pixel shader model 1.1 (DirectX 8) - This is the very
first DirectX pixel shader model.
- PS1.3 - Pixel shader model 1.3 (DirectX 8.1) - This is a
special extended version of PS1.1 for the NVIDIA NV25 and NV28.
- PS1.4 - Pixel shader model 1.4 (DirectX 8.1) - This is a
special extended version of PS1.1 for the later versions of the ATI R2xx
and RV2xx family.
- PS2.0 - Pixel shader model 2.0 (DirectX 9) -
This is the original pixel shader included with DirectX 9.0. It was
introduced first in the Radeon 9700 Pro. This is considered to be the
first pixel shader model with the necessary color precision and
flexibility to generate high quality images.
- PS2.0a - Pixel shader model 2.0a (DirectX 9.0b) - This is a
special extended version of PS 2.0 for the NVIDIA NV3X
GPU
family.
- PS2.0b - Pixel shader model 2.0b (DirectX 9.0b) - This is a
special extended version of PS 2.0 for the ATI R4XX and RV4XX
GPU
family.
- PS3.0 - Pixel shader model 3.0 (DirectX 9.0c) - This removed
many of the remaining limits on programmability in pixel shaders. This is
an incremental step forward from PS 2.0 but is not as big a step forward
as PS 2.0 was over previous pixel shader versions.
Vertex hardware
- T&L - Hardware transforms and lighting (DirectX 7) - This is
hardware which allows the GPU to handle geometry transforms and lighting
rather than have it done by the CPU.
- VS1.1 - Vertex shader model 1.1 (DirectX 8) - This is the
very first DirectX vertex shader.
- VS2.0 - Vertex shader model 2.0 (DirectX 9) - This is the
original vertex shader included with DirectX 9.0.
- VS2.0a - Vertex shader model 2.0a (DirectX 9.0b) - This is a
special extended version of VS 2.0 for the NVIDIA NV3X
GPU
family.
- VS3.0 - Vertex shader model 3.0 (DirectX 9.0c) - This is an
updated version of VS 2.0. It adds (among other things) the ability to
access textures which can allow for more interesting effects.
Unified shader architecture hardware
- SM4.0 (DirectX 10.0) - Shader model 4.0 - This shader model
includes pixel shaders, vertex shaders, and geometry shaders. This
doesn't include a number of processing units because those are
dynamically allocated by unified shader architecture GPUs.
- SM4.1 (DirectX 10.1) - Shader model 4.1 - This is a minor
upgrade to shader model 4.0. This shader model
includes pixel shaders, vertex shaders, and geometry shaders. This
doesn't include a number of processing units because those are
dynamically allocated by unified shader architecture GPUs.
- SP - Stream processor (DirectX 10) - A stream processor is a
processing unit
which is shared between pixel shaders, vertex shaders, and geometry
shaders in a unified shader architecture. The number of stream processors
determines the total processing capacity of the GPU. The stream
processors are dynamically allocated to do pixel, vertex, or geometry
processing based on immediate demand.
Fill rate
This table column is the peak fill rate in gigatexels per second. This is a
measure (of sorts) of the speed of this video card at older multi-texturing
games. This value may be constrained by insufficient memory bandwidth. Newer
game speed depends much more on shader speed and various kinds of image
quality computations. So older video cards were designed to deliver lots of
gigatexels per second. But newer video cards are designed to deliver more
shader/image quality processing power with enough fill rate to spit the image
out. So for newer video cards, the fill rate measure just isn't that useful
anymore. As a result,
DirectX 10.0 or newer video cards aren't given a fill
rate value.
RAM size
This table column provides the amount of
video RAM
on the standard model of the video card in
megabytes.
Multiple numbers are listed if there are common models with different amounts
of video RAM.
Some low-end PCI-Express video cards can borrow RAM from the motherboard to
use as video RAM. NVIDIA's implementation is called
TurboCache.
ATI's is called
HyperMemory.
These cards have a total useable amount of RAM which is the sum of both the
video RAM actually on the video card plus the RAM borrowed from the
motherboard. The column value for this kind of video card only includes the
video RAM actually on the video card followed by "[*]". An entry in the
notes column provides the rules about how much
total useable RAM is available for each configuration.
RAM clock
This table column contains the total
video RAM
peak data rate in
MHz.
So if the video card has
DDR
memory with a clock rate of 100 MHz then this column contains 200 because
there are two transfers per clock. This is the value for the standard
implementation of the video card. Some models may provide different clock
rates. If this video card commonly has a short list of possible RAM clock
rates then they are listed separated by ","s. If this video card commonly has
a range of clock rates then the table column contains the low end and high
end of the commonly used range separated by a "..". Some video cards have
both single data rate (SDR) and
double data rate (DDR) models which actually have the same performance specs
because the DDR model has half the RAM width.
In that case, the SDR clock rate is followed by "(SDR)" and the DDR clock
rate is followed by "(DDR)". Many video cards automatically reduce this clock
rate when the video card is not heavily loaded to make the card easier to
cool. This table column reports the maximum clock rate.
RAM width
This table column contains the width of the
video RAM
in
bits.
Wider video RAM means faster video RAM. You can read more about
video RAM on this page.
Some video cards have both single data rate (SDR) and
double data rate (DDR) models which actually have the same performance specs
because the DDR model has half the memory width. In that case, the SDR memory
width is followed by "(SDR)" and the DDR memory width is followed by "(DDR)".
RAM speed
This table column provides the peak
video RAM
bandwidth in trillions (1,000,000,000) of bytes per second of the video RAM on the video card. I would rather
provide the bandwidth in GB/s
(2^30 bytes per second) but it seems that everyone (including old versions of
this web page) have opted for 10^9 instead of 2^30 when calling things
gigabytes per second when it comes to video RAM. By the way, in the rest of the
table a gigabyte of video RAM is still a "real" gigabyte (2^30) of RAM.
Hard
disk counting
methods apparently only apply to video RAM bandwidth numbers.
All binary grousing aside, larger numbers means that it has faster video RAM. If the
memory clock has a range of values then the
memory bandwidth contains a range of values. Note that if the video card
implements sharing of motherboard RAM over a
PCI-Express x16
slot (it implements
TurboCache or
HyperMemory),
then the value in this table column only provides the peak speed of the local
video RAM. It does not include any extra bandwidth for parallel accesses to
motherboard RAM. You can read more about memory bandwidth on
this page.
| Useful technical information |
|
|
