• 956 million transistors on 55nm fabrication process
• PCI Express 2.0 x16 bus interface
• 256-bit GDDR3/GDDR5 memory interface

• Microsoft DirectX 10.1 support

  • Shader Model 4.1
  • 32-bit floating point texture filtering
  • Indexed cube map arrays
  • Independent blend modes per render target
  • Pixel coverage sample masking
  • Read/write multi-sample surfaces with shaders
  • Gather4 texture fetching

• Unified Superscalar Shader Architecture

  • 800 stream processing units

    • Dynamic load balancing and resource allocation for vertex, geometry, and pixel shaders
    • Common instruction set and texture unit access supported for all types of shaders
    • Dedicated branch execution units and texture address processors
  • 128-bit floating point precision for all operations
  • Command processor for reduced CPU overhead
  • Shader instruction and constant caches
  • Up to 160 texture fetches per clock cycle
  • Up to 128 textures per pixel
  • Fully associative multi-level texture cache design
  • DXTC and 3Dc+ texture compression
  • High resolution texture support (up to 8192 x 8192)
  • Fully associative texture Z/stencil cache designs
  • Double-sided hierarchical Z/stencil buffer
  • Early Z test, Re-Z, Z Range optimization, and Fast Z Clear
  • Lossless Z & stencil compression (up to 128:1)
  • Lossless color compression (up to 8:1)
  • 8 render targets (MRTs) with anti-aliasing support
  • Physics processing support

• Dynamic Geometry Acceleration

  • High performance vertex cache
  • Programmable tessellation unit
  • Accelerated geometry shader path for geometry amplification
  • Memory read/write cache for improved stream output performance

• Anti-aliasing features

  • Multi-sample anti-aliasing (2, 4 or 8 samples per pixel)
  • Up to 24x Custom Filter Anti-Aliasing (CFAA) for improved quality
  • Adaptive super-sampling and multi-sampling
  • Gamma correct
  • Super AA (ATI CrossFireX configurations only)
  • All anti-aliasing features compatible with HDR rendering

• Texture filtering features

  • 2x/4x/8x/16x high quality adaptive anisotropic filtering modes (up to 128 taps per pixel)
  • 128-bit floating point HDR texture filtering
  • sRGB filtering (gamma/degamma)
  • Percentage Closer Filtering (PCF)
  • Depth & stencil texture (DST) format support
  • Shared exponent HDR (RGBE 9:9:9:5) texture format support

• OpenGL 2.0 support

Although the
cards look similar and follow a similar design philosophy, the Radeon
HD 4870 X2 differs from AMD’s previous dual-GPU based Radeon HD 3870 X2
in a number of ways.

As the above
diagram shows, the Radeon HD 3870 X2 featured a pair of GPUs–in its
case, RV670s–linked together via a PCI Express gen 1 bridge and a
CrossFire bridge interconnect with .9GB/s of peak bandwidth. In total,
the design offered an aggregate 6.8GB/s of total interconnect bandwidth
across the card.

The Radeon HD
4870 X2 is quite different, however. Of course, the card is built
around a new generation of GPUs, namely the RV770. The card is powered
by two of the very same GPUs used to power the Radeon HD 4870. They are
also linked together by the very same .9GB/s CrossFire bridge
interconnect, but the on-board switch used on the Radeon HD 4870 X2 is
PCI Express gen 2 based and hence offers twice the bandwidth. Also note
that the RV770 GPU features a new sideport interface that offers
another 5GB/s of bandwidth between the two GPUs, each way. The
combination of a new gen 2 switch and the sideport interface result in
a total aggregate interconnect bandwidth on the Radeon HD 4870 X2 of
21.8GB/s, which is more than 3X that of the 3870 X2.

We should
note, however, the additional bandwidth afforded by the new sideport is
not exploited by current software and applications. Representatives
from ATI have informed us that future applications should benefit from
the sideport, but they didn’t explain exactly how or when. We suspect
the sideport may enable AMD to develop new multi-GPU rendering modes,
or perhaps enable higher performance in certain stream computing
applications. Time will tell.
 

If you look
at the breakdown above, the Radeon HD 4870 X2 isn’t simply two Radeon
HD 4870s fused together on a single PCB. The GPUs on the Radeon HD 4870
X2 will be clocked at a minimum of 750MHz and the X2’s gigantic frame
buffer memory will be clocked 900MHz (3.6Gpb/s GDDR5 effective), which
is similar to the single GPU Radeon HD 4870.  But the X2 features
double the amount of frame buffer memory, which could help performance
at higher resolutions, and it’s also outfitted with a gen 2 PCI Express
switch.

The Radeon HD
4870 X2’s main features and benefits are outlined on the slide above.
As you can see, the cards offers up to 2.4 TFLOPS of compute power with
its 750MHz core GPU clock and 1600MHz (800 x 2) stream processors. Its
2GB (1024MB per GPU) is clocked at 900MHz, which results in a 3.6Gbps
GDDR5 data rate, and total memory bandwidth is 230GB/s (115GB/s x 2).
With all that horsepower under the hood, the card has a max board power
of 285 watts.
 

As the old saying
goes, "But wait! There’s More!" AMD has also informed us that they have
a bit of a surprise in the works. A Radeon HD 4850 X2 card that
utilizes less expensive GDDR3 RAM is also planned. The Radeon HD 4850
X2’s primaary tech specs are outlined in the slide above. With an
expected $399 MSRP, the Radeon HD 4850 could wind up being an
attractive option in the current 3D graphics landscape and may put even
more price pressure on NVIDIA. We hope to have the 4850 X2 in the lab
in a few weeks, and will reveal all of its juicy details then. 
Expected availabilit is sometime in September.

Upon initial
inspection, the Radeon HD 4870 X2 looks much like the Radeon HD 3870
X2, just with different PCB coloring. The card features a dark, black
PCB with an industrial looking image of bolts and the Radeon logo
emblazoned on the fan shroud. Flip the card over, and as you probably
expect, the complexity of the design shows through.

      
Palit Radeon HD 4870 X2

We
acquired cards from HIS and Palit for this article, both of which are
pictured here. In terms of their features and specifications, the Palit
and HIS Radeon HD 48070 X2 cards don’t differ from ATI’s reference
design, save for a couple of decals and their accessory bundles.

As
we’ve already mentioned, the Radeon HD 4870 X2 is powered by a pair of
750MHz RV770 GPUs, the same chips used on the Radeon HD 4870 and 4850.
The two GPUs are linked together on the PCB through a PCI Express 2.0
fan-out switch from PLX. That switch takes the 16 PCI express lanes
coming from the PEG slot and distributes them to both of the GPUs.

Like
the Radeon HD 3870 X2, the new 4870 X2 has only a single CrossFire edge
connector along the top of its PCB. It has only one because the other
connection is already utilized on the PCB as an interconnect between
the on-board GPUs.

      
HIS Radeon HD 4870 X2

Each
of the GPUs on the Radeon HD 4870 X2 has its own 1GBMB frame buffer,
for a total of 2GB of on-board memory. As you can see, the entire
assembly is quite large and results in a 10.5” PCB – a little longer
than a standard ATX motherboard. The cooler used on the card is much
like previous offerings, but there are some noteworthy elements to the
design. Each of the GPUs gets its own heatsink, as do the surrounding
switch and RAM. The squirrel-cage type fan draws air in through a hole
at the rear of the card, blows air across the heatsinks, and expels it
from the system through vents in the mounting plate.

Like the
other members of the Radeon HD 4800 series, the ATI Radeon HD 4870 X2
has a pair of dual-link DVI outputs, and an HD video output. One of the
DVI outputs can be converted to an HDMI output with audio using an
included adapter.

The Palit samples we received did not come in
full retail trim, so we can’t comment on the included bundle at this
time.  The HIS card, however, was retail-ready and included a drive CD
and manuals, a CrossFire bridge connector, a DVI-to-VGA adapter, a
DVI-to-HDMI adapter, an S-Video to composite adapter, an HD component
output dongle, a Molex-to-PCI Express power adapter and an HIS case
badge.