What
is the HD 2000 Series?
AMD's Radeon HD 2000 series is
currently broken up into 4 product categories,
the 2900 for enthusiasts, 2600 for mainstream
, 2400 for value and 2300 (for Mobility only at
this time). Within these specific categories there
are (or will be) several products such as Pro
and XT. Today we will be concentrating
on the 2900 XT however we will note many of the
differences between it and the 2600 series. (NOTE:
There will also be mobile variations of the products
above the 2300).
The 2900 XT originates from the
ATI design within the Xbox 360 GPU (Xenos), however
there are numerous significant differences, the
2900 series GPU is a 80nm part where as Xenos
is currently 90nm, the 2600 and 2400 series are
even more advanced using the 65nm process. The
transistor count for the HD 2900 is 700 million
and within the core there are 320 stream processing
units, when compared with the X1900 series which
had 48 vector and 48 scalar processing units this
is a significant architectural difference.
Comparing the 2900 XT to the recently
released 8800 GTS there is quite a difference
as the Nvidia counterpart has 96 Stream Processors,
however these are not directly comparable
to the 2900's stream processing units. The 80nm,
700m transistor XT has impressive clock rates,
the reference and initial retail cards will come
with a core speed of 740MHz and 512Mb of DDR3
running at 825Mhz on a new memory controller,
we will delve into that shortly.
Looking deeper into the stream
processing units we see that the 2000 series features
dynamic shader load balancing (handled by the
hardware scheduler) between vertex, pixel and
geometry shader operations and is fully DirectX
10.0 compliant which by default includes Shader
Model 4.0. Whereas the X1000 series could process
2 instructions per clock of 3+1 or 4+1 components
(Vector and Scalar) the HD 2000 series can process
5 instructions per clock with 5 components (Superscalar).
 |
The
texture units within the 2900 are capable of 64-bit
HDR textures, bilinear filtered at full speed
and 128-bit FP textures filtered at half speed.
They also feature a new compact 32-bit HDR shared
exponent texture format and of course 16x Anisotropic
Filtering which has been tweaked to provide theoretically
higher image quality compared to previous generation
Radeon hardware. Interestingly AMD have also made
a point of stating that the cards support Percentage
Closer Filtering (PCF) which is a feature we first
saw used by Nvidia about 4 years ago. (PCF is
used to optimise shadow rendering). This is quite
a strange time for AMD to promote this specific
feature, anyway we digress...back to the core
specifics...
In
terms of geometry performance the HD 2900 has
8x the vertex cache of the X1950 and all shader
processors can perform vertex and/or geometry
processing as required. As we mentioned earlier
in this section this GPU has its origins in the
X360 and one of the new features is the programmable
tessellation unit which has evolved from the technology
used in Xenos and provides geometry data compression.
In
a similar fashion to the X1000 hardware the HD
2000 series uses a ring bus as its memory controller.
The X1000 series used a hybrid design which took
the centralised/crossbar design used in previous
Radeons and added in two 256-bit rings which worked
in opposite directions:
The
2000 series moves to a fully distributed ring
bus design which is capable of 512-bit read and
write operations. With the 8 individual 64-bit
memory channels we have a controller which is
capable of over 100GB/sec in memory bandwidth.
 |
As
with previous generation ATI parts CrossFire is
also supported by the HD 2000 hardware. With the
later X1950 series cards we saw a move away from
the original external CrossFire dongle to an internal
bridge, the HD2900 also uses an internal connector
for communication between cards (much like Nvidas
SLI bridge) and additionally no master card is
required. End users can pick up 2 of the same
products, install them in a CrossFire enabled
motherboard and connect them together using the
bundled bridge (All retail cards we have seen
come with the connector). After installing the
driver they will then be good to go. As with earlier
CrossFire implementations the HD 2000 series feature
AFR, SuperTile and Scissor Rendering as well as
Super AA modes. These can be used at resolutions
up to 2560x2048 (60Hz). All of the above is pretty
much as expected, however there is one new featureset
of the HD2000 CrossFire which is the ability to
support more than 2 GPU's in the future.
The
2000 series feature new AA modes (8xMSAA and up
to 24x Custom Filter AA) with programmable sample
pattern and resolve filters which can be updated
by the drivers. It has to be said that the choice
of language used in the driver control panel doesn't
do anything to explain what the user may be choosing
for AA (Wide Tent????)
In
simple terms the new AA can be explained as follows.
Standard AA uses a box area in which to apply
the technology and there are many of these boxes
sitting beside each other on your screen. ATI's
new narrow tent method uses overlapping ovals
for the area to sample rather than adjacent boxes.
For Wide tent a larger oval is used which overlaps
further meaning more of the pixels on screen have
AA applied at any one time, resulting in a higher
level of overall AA.
 |
The
final major aspect of the HD 2000 series is Avivo
functionality. Avivo is a term AMD have been using
for some time however in the new products this
changes to Avivo HD. The Radeon HD 2900 series
features HDCP over Dual-Link DVI and HDMI. Bundled
with each 2900 XT (we have seen to date) is a
DVI to HDMI dongle which is ideal for those wishing
to connect their system to an HDTV. A new feature
is the ability of the Radeon HD 2000 series to
transfer audio as well as video through the DVI>HDMI
connector and into the TV, this becomes even more
impressive when we take into account that full
AC3 5.1 surround sound is supported (Dolby Digital
and DTS).
 |
The
HD 2000 series also features two levels of Video
decoder/acceleration technology. On the 2600/2400
cards we have a universal video decoder which
has full acceleration of HD/BD content. This should
allow for very low CPU usage as demonstrated by
the Geforce 8500/8600 series, however at this
time we have not received a sample of either Radeon
card to confirm the performance. The 2600/2400
also have the ability to perform various post
processing tasks such as De-interlacing and colour
correction. On the Radeon 2900, disappointingly,
the full acceleration of BD/HD is not present
and as a result the CPU usage is greater when
watching a high-definition movie. We will look
at the performance of the 2900 XT in Video Playback
later in the review.
Before
we go on to look at some of the retail cards it
is worth taking some time to note features that
the HD2000 series have that are not necessarily
in use today however can be exploited in the future.
First up, and probably most important for enthusiasts
is the ability for the Radeon to perform physics
work, this isn’t a new feature (Geforce
8 can also do this) however it is good to see
the option appear in this product. It will also
be possible to program the card to perform medical
and chemical visualisations as well as fluid dynamics
and volumetric work.
For
those wanting a quick summary of the card specifications/features
here it is:
