GPU Core Competencies

With this generation of GPU, ATI’s beginning to see the payoff from the premature move to a 55nm die size last generation. While Nvidia languishes at 65nm, ATI is packing more silicon into a smaller space, and increasing efficiency at the same time. But, that’s not all ATI’s done. The new RV770 series GPUs feature a redesigned GPU core with an astounding 800 stream processors—the little silicon dynamos that handle everything from rendering soft shadows and bump maps to decoding H.264 video from Blu-ray movies.

By integrating 16KB of cache, bundles of 10 stream processors, and four dedicated texture units in so-called SIMD units, ATI has juiced much better shader performance out of the overall package. The 10 stream processors share can share information with other processors in the same bundle using their shared memory, which makes the new shader core much more efficient than previous designs. And, because the shader cores pump their output directly into dedicated texture units, there’s very little time lost between writing the output to texture memory.

By integrating the stream processors in modular units around the texture processing cores, the RV770 design minimizes latency and improves performance. Each SIMD unit is connected to four dedicated texture units with 480GB/sec of bandwidth between them. This was absolutely crucial to maintain performance, or the texture units, which render the actual pixels that are displayed, would become the bottleneck again.

Under the Hood: RV770 Unveiled

ATI packed the latency-sensitive silicon, like the stream processors and the basic texturing units in the center of the RV770s GPU. Surrounding that is are the memory controllers and L2 cache, and on the periphery of the chip rests the memory interface (GDDR5 on the 4870 and GDDR3 on the 4850), the PCI Express connection, the Crossfire controller, and the various display controllers for DVI, HDMI, VGA, and DisplayPort. And they packed all that on a 260mm^2 55nm die.

Sometimes Narrower is Better

There are two basic ways to increase memory bandwidth. You can increase the clock speed of the memory or you can transfer more data with every clock cycle by increasing the width of the memory bus. Like ATI’s previous-generation GPUs, Nvidia’s GTX 280 uses a 512-bit wide memory bus. The new GPU utilizes a narrower 256-bit bus, but it’s using new GDDR5 memory, which allows twice as many transfers per clock cycle as GDDR3. This gives ATI roughly the same memory bandwidth as the GTX 280 on a board with a cheap 256-bit bus and which transfers more data at lower clock speeds.

GDDR5 also uses fewer pins to connect the memory to the board. This reduces board complexity compared to DDR3, which is especially important when you consider the greatly reduced space available for connector pins on GPUs that use smaller process technology. By using a less complex 256-bit bus and cranking the clocks up on the GDDR5 memory, ATI should be able to able to bring decent memory performance in without harming yields for the GPU.

The high-end Nvidia GeForce GTX 280 runs its GDDR3 memory at a punishing 1100MHz and pushing an impressive 115GB/sec of bandwidth. Meanwhile ATI’s 4870 just ticks along at 900MHz, but runs at the same 115GB/sec. The net result is that the ATI card’s memory draws less power and generates less heat, while delivering the same level of performance as the more expensive card.

Running GDDR5 memory at speeds lower than GDDR3 memory with the same bandwidth is great, but the current low-end and mid-range ATI boards only support 512MB of total memory (the GeForce GTX 260 ships with 896MB of memory on a 448-bit interface and the GTX 280 ships with a full gigabyte). For the most part, performance doesn’t seem to suffer due to lack of memory, but that could change as graphically intensive games like Far Cry 2 and Fallout 3 are released this year.