New unified architecture

High-end Specifications Comparison
	Radeon HD 2900 XT	GeForce 8800 GTX	GeForce 8800 GTS 640MB
# of Transistors	700M	681M	681M
Core Clock Speed (MHz)	742	575	500
Stream Processor Clock Speed (MHz)	740	1350	1200
# of Stream Processors	320	128	96
Texture Units	16	32	24
ROPs	16	24	20
Texture fill-rate (Gigatexels/sec)	11.9	18.4	12.0
Memory Clock (MHz)	1650	1800	1600
Memory Interface	512-bit	384-bit	320-bit
Memory Bandwidth (GB/sec)	105.6	86.4	64
Memory Size	512MB	768MB	640MB

While R600 is technically a new design, the chip leverages many technologies already found in ATI’s last two GPUs: R520/580 from the Radeon X1000 series and the Xbox 360’s Xenos GPU. Looking over a block diagram of R600, you’ll immediately recognize many units from R5xx. For instance, to improve dynamic branching, ATI continues to break down the processing workload into a large number of small threads. These threads are then managed by the ultra-threading dispatch processor. Meanwhile from Xbox 360 ATI leverages their unified shading architecture.

For R600, ATI merely builds on this design, adding more powerful superscalar shader processors, a tweaked ultra-threading dispatch processor, the addition of a new tessellation unit (actually borrowed from Xenos) full DirectX 10 support, and a more robust, 512-bit memory interface. All this adds up to a GPU that’s been designed for the next generation in HD gaming, delivering very high levels of performance even when gaming at mega resolutions such as 2048x1536 and 2560x1600 with HDR+AA.

At the heart of all this is ATI’s 2nd-generation unified shader architecture. Consisting of 320 distinct, independent stream processing units, it’s quite impressive. Like NVIDIA’s G80 GPU, ATI has incorporated a scalar architecture for R600’s shading processors (the stream processing units). Only in ATI’s case, R600 can issue many more independent instructions in each shader processor due to its superscalar design. The Radeon HD 2900 can issue up to five scalar multiply-add (MAD) operations and one branch instruction to each shader processor per clock cycle. In comparison, each stream processor in G80 can dual-issue a MAD and MUL instruction.


In the block diagram above the stream processors are depicted as yellow squares in the center of the GPU. Attached to each group of five stream processors is a dedicated branch execution unit (the purple square) and general purpose registers which can be used to store input data, temporary values, and output data. Here’s a group of stream processors up close:

With more shading units onboard, it’s important to keep these shader processors fed with data. This is where the ultra-threading dispatch processor comes in. The ultra-threading dispatch processor acts as a traffic cop, it’s a central dispatch unit that is responsible for tracking and distributing thousands of threads simultaneously across the Radeon HD 2900’s shader processors. ATI won’t provide a specific max number of threads, but in comparison, R520/580 was limited to 512 threads.


The above diagram breaks down the ultra-threading dispatch processor. As you can see, ATI has provided separate command queues for each shader type: pixel, vertex, or geometry data. From there arbiter units determine which threads will be processed first, based on a variety of parameters. ATI provides two arbiter units per SIMD array, with dedicated arbiter units for texture and vertex fetches, allowing them to be scheduled independently of math operations. Threads that are already executing can be bumped at any time if a higher priority thread is pulled from the command queues. The temporary data is saved so the thread can be resumed later. If a thread is forced to wait for data, it is suspended and a new thread begins executing immediately. The suspended threads remain in the command queue until their requested data arrives. According to ATI, hundreds of threads can be queued up to make sure the SIMD arrays are never sitting idle.