ATI R520

ATI's new GPU chip R520 is internally called "Fudo". The design has been ready for quite some time now. One source is telling us "the chip is working, now ATI is designing the boards, and after that they have to fix the boards to get them working [laughs]".

The chip is produced using a 90nm process at TSMC and will have 300 to 350 million transistors.

We have confirmed the 24 pipelines rumors from several independant sources. However, these 24 pipelines are not comparable to 24 of today's pipelines - multiply by 1.3. So the 24 pipelines will have the performance of 32 "normal" pipelines.
The result of this is that the R520 will be two times as fast as the currently fastest X850s.

ATI will most probably present the chip in June at Computex, Taiwan.

The most important architectural change is support for the Advanced Driver Model of WGF 1.0 (Windows Graphics Foundation). This is the successor to DirectX9, often called DirectX Next. WGF will be included as standard in Microsoft Longhorn.
WGF offers new exciting features for game developers like "unlimited" shader length, Geometry Shaders, Much lower overhead than DirectX9 and many more.

Existing hardware will support the Basic Model of WGF once the manufacturer writes a driver for it. Features like resource virtualization etc. will only be available when using the Advanced Driver Model. R520 is the first GPU we know of to support it.

Obviously ShaderModel 3.0 is supported by R520 as well.

Memory configurations will be 256-bit memory interface (definitely not 512-bit) with 256 MB and 512 MB GDDR3. The internal registers of R520 support even 1GB memory size. With current GDDR3 prices we will definitely not see a 1GB design at least not in the near future and not for gaming cards. However, this shows that ATI is planning for the future.

The boards will be available in PCI-Express only. When asked about an AGP version, one source said "most probably not". Please note that this is not a definite "no". My guess is that if the market really demands an AGP version, it will be implemented via the Rialto bridge chip.

In the S3 part of this report we talked about OmniChrome supporting 10-bit-per-color display depths, so does R520. HDTV and HDCP output will be supported as well.

Clocks are not finalized yet, but expect something around 700 MHz core and 800 MHz memory (1.25ns).

The first boards will be VERY expensive people are talking $600-$700 range. It all depends on how the graphics memory prices develop in the next months and if TSMC can get good yields out of the 90nm process.

ATI SLI

Not much is known on this topic. Some people say the name is AMR "ATI Multi Rendering", others say MVP "Multi Visual Processing".
Our information differs from source to source. MVP and "neither AMR nor MVP" being the most popular answers.

ATI's technology is superior to NVIDIA's SLI in several ways:

While NVIDIA SLI splits the screen into two parts and one GPU renders each half, ATI's solution breaks the screen into little squares. Remember the chessboard effect you had when pipelines were bad with the softmod?
That's how it is going to work, just on a bigger scale, called "Super Tiling". Any GPU from R300 and up supports Super Tiling in hardware and they can be mixed when in SLI mode.
However, ATI most probably will not enable SLI for GPUs other than the R4xx and R520 series.

ATI's SLI is also not limited to two video cards. There will be one master card which presents the rendered data to the screen and coordinates the slaves which transfer their rendered data into the master's framebuffer. A physical PCB to link the video cards is not required, everything is done via PCI-Express bus.

Continue to Day 5.