View Single Post
  #1  
Old October 2nd 05, 08:37 PM
external usenet poster
 
Posts: n/a
Default ATI R520 architecture

http://www.hardwareanalysis.com/cont.../article/1817/

Next week we'll see the introduction of ATI's R520 architecture in
the form of a whole new top-to-bottom generation of graphics cards
which will be dubbed the X1000 series. The X1000 series starts off with
the top-of-the-line Radeon X1800 XT and has the low-end X1300 closing
the ranks. ATI has put all their engineering effort towards a new
graphics accelerator architecture based on the 90-nm process. Key
features include SM3.0 support, a 512-bit ring bus memory controller
and support for their recently introduced Avivo platform. In these next
few paragraphs we'll be giving you an exclusive first look at what
ATI will be introducing next week.

R520 architecture

The focus of ATI's new R520 architecture has really been on
efficiency, making sure that all clock-cycles are put to good use and
creating minimal overhead. Unlike previous rumors indicating the
architecture had up to 32-pipelines, the actual number is 16-pipelines
or pixel-shader processors, with 8-vertex shader processors. Obviously
the pixel-shader processors have been upgraded to provide full support
for SM 3.0. But there's more; ATI added dynamic flow control to
streamline processing and multi-threading support, the new architecture
now allows for 512 pixel-shader threads to be outstanding. The
vertex-shader processors have also been updated; they're now able to
process two shader instructions per clock and handle up to 10-billion
instructions per second.




The large number of outstanding pixel-shader threads, 512, is actually
one of the strengths of the architecture as threads are spread out
across the pixel-shader processors and are executed in parallel. A
total of 16 pixel-shader processors are available on the X1800,
organized in groups of four, hence there's four shader cores, dubbed
the quad-pixel shader cores. The pixels are handled in small 4x4 pixel
groups, so the X1800 handles just 16-pixels per thread which makes the
thread size small and efficient. Each of the pixel-shader processors
can process per thread six different shader instructions on four pixels
per clock-cycle. Obviously branch prediction is used to minimize
overhead and this is not handled by the ALUs, but by dedicated logic
included with ever pixel-shader, further improving efficiency.




R520 architecture (block diagram)
http://hardwareanalysis.com/images/a...arge/11625.jpg

The pixel-shader engine itself is comprised of four quad-pixel shader
cores, a number of general purpose register arrays, sixteen texture
address units and the ultra-threading dispatch processor. Data gets
sent from the setup engine down to the ultra-threading dispatch engine
that distributes the load across the four quad-pixel shader cores. The
shader instructions are processed with 128-bit floating point precision
and make use of the general storage register arrays that serve to
increase bandwidth and provide local storage for instructions but also
allow for fast thread switching. Due to the dedicated branch execution
units per pixel-shader overhead is minimalized. Combined with the
dynamic flow control implemented in the ultra-threading dispatch
processor all the shader processors are constantly fed with data and
can execute in fewer clock-cycles. All in all this looks like an
efficient new architecture that can be manufactured cheaply on the 90nm
process, in combination with the fewer number of pipelines the die will
also be smaller than competing products from NVIDIA.

Shipping products

ATI will launch with immediate availability a number of graphics cards
based on their new architecture. The top-of-the-line Radeon X1800XT
will be available as of the 5th of November shipping at a 625MHz core
and 1500MHz memory frequency. Offered with 256MB and 512MB of GDDR3
memory it will be priced at $499 and $549 respectively. The Radeon
X1800XL will be available on October 5th and offers a 500MHz core and
1GHz memory frequency and comes with 256MB of GDDR3 memory with a price
tag of $449. All of the Radeon X1800 graphics cards utilize the full
16-pipelines, they're just different in terms of memory size and
clock speed or both in the case of the Radeon X1800 XL.




The top-of-the-line Radeon X1800 series will ship with a dual-slot
heatsink.

The mid-range Radeon X1600XT ships with either 128MB or 256MB of
memory, clocks in at a 590MHz core and 1.38GHz memory frequency and can
be bought starting November 30th for respectively $199 and $249. All of
the Radeon X1600 graphics cards will have to make do with just
12-pipelines. The low-end Radeon X1300 will be offered in three
different versions, the $149 Radeon X1300 Pro with 256MB of memory and
600MHz core and 800MHz memory clocks. The $99 and $129 Radeon X1300
with 450MHz core and 500MHz memory clocks and 128MB or 256MB of memory
respectively. And closing the ranks is the Radeon X1300 HyperMemory at
$79 with 32MB HyperMemory and 450MHz core and 1GHz memory clocks.




The mid-range X1600 and low-end X1300 series will ship with a
single-slot heatsink.

The Radeon X1300 Pro and X1300 will be available on October 5th and
sadly they will feature only 4 pipelines. Unfortunately no dates have
been given for the Radeon X1800 CrossFire Edition or the Radeon X1600
CrossFire Edition. What is clear, though, is that the Radeon 1300 will
not feature a CrossFire Edition card, it'll utilize the PCIe bus for
communication. Both the Radeon X1800 and X1600 CrossFire Edition will
support resolutions up to 2046x1536 pixels with a 70Hz refresh,
negating some of the limitations of CrossFire on the Xx00 platform.

That concludes our first look at the R520 architecture and the X1000
series cards that will be available at launch. As for performance
we'll have to wait for the actual launch. If ATI is to be believed
the Radeon X1800XT consistently outperforms the GeForce 7800GTX in all
games, sometimes by as much as 80%. I'm afraid that is pure marketing
speak however, as they've made bold claims such as these before. I
vividly remember when the Radeon X800XT PE was introduced they
circulated similar benchmark results showing it besting the GeForce
6800 Ultra by as much as 40%, which of course proved to be just wishful
thinking. The preliminary benchmarks we've provided earlier, run on a
pre-production card, will prove to be on the low side however. ATI has
already raised the clock speeds and has done some extensive work on
their drivers since then, effectively rendering them outdated. However
we don't doubt that the Radeon X1800XT will have better performance
than the GeForce 7800GTX, even if just by a small margin.

Sander Sassen.