Architecture

As mentioned earlier, the Radeon R9 290X is AMD's genuinely new high-end graphics chip, and it hence introduces a few new features. Codenamed "Hawaii," the silicon that drives the R9 290X packs a staggering 6.2 billion transistors and is built on the existing 28 nanometer silicon fabrication process, which we imagine should have achieved a good level of maturity by now (good yields). The very fact that AMD is now offering a product with a 6.2 billion transistor chip and 16 GDDR5 memory chips, let alone how well it ends up performing, shows that the GeForce GTX TITAN is a very high-margin product.



Internally, "Hawaii" features the same component hierarchy as "Tahiti," and is an upscale in many respects. While "Tahiti" features 2,048 stream processors, the number-crunching machinery of the chip across two shader engines, "Hawaii" features 2,816 stream processors across four shader engines. As you may have guessed, stream processors are not all that make up a shader engine. It has several other key components and by itself features a vast majority of the components that go into building a GPU. For instance, the stream processors inside a shader engine are arranged into groups of 64 called compute units (CUs).



In each of the four shader engines on "Hawaii," there are 11 compute units (CUs). A CU combines scalar and vector number crunching machinery into neatly organized blocks that are fed data and instructions by a scheduler told on how to go about number crunching by scalar and vector registers, and cushioned by caches at various levels.

Apart from compute units, a shader engine features a geometry processor, a key component tasked with computing geometry in a 3D scene. It combines geometry and vertex assemblers, and a tessellator, a component specialized in processing extremely complex geometric shapes. Since "Hawaii" features four shader engines and in turn four geometry processor, it features four independent tessellation units, double that of the previous generation.



Also inside the shader engine is the Rasterizer, a pre-rendering stage, and Render back-ends, a post-rendering stage. Each CU features 16 raster operations units (ROPs) for a total of 64 ROPs on "Hawaii," the highest number of ROPs on a GPU yet.

Processing loads to the four shader engines are managed by a Command Processor, a global data share that distributes data tagged along with the right instructions; and a unified 1 megabyte L3 cache which cushions data between the rest of the GPU and the memory controllers. 1 megabyte may seem infinitesimal compared to the amount of memory the R9 290X features, but the speed at which the cache serves its function as a scratchpad more than makes up for it.

"Hawaii" is the industry's first graphics processor to feature a 512-bit wide GDDR5 memory interface, double the width of "Cayman" and 33 percent wider than "Tahiti." The standard memory amount on the card is set at 4 GB, but don't be surprised if a few crazy AIBs go ahead and launch 8 GB variants in the future. The GPU core operates at 1000 MHz and the memory at 1250 MHz (5.00 GHz effective), which translates into a staggering 320 GB/s memory bandwidth.

"Hawaii" takes advantage of the PCI-Express Gen 3.0 x16 bus interface. Other miscellaneous components include specialized hardware for high-resolution video acceleration (UVD), a hardware video CODEC engine (VCE), specialized hardware for the TrueAudio positional audio processing pipeline, and the display controllers with support for up to six displays.

DirectX 11.2

We'd like to slip in a quick note about DirectX 11.2. DirectX 10.1 was only supported by AMD graphics cards, before the DirectX 11 era began, which greatly limited its adoption by game developers. Today's scenario is vastly different. Although NVIDIA's GeForce GTX 700 and the GTX TITAN don't support DirectX 11.2 and can only emulate certain effects over DirectX 11.0, AMD has a good shot at turning its DirectX 11.2 hardware support into a genuine advantage. AMD has vastly better developer relations than it did three years ago, and its DirectX 11.2-ready hardware drives the Microsoft Xbox One and Sony PlayStation 4. Microsoft could, on its part, encourage game developers to take advantage of DirectX 11.2 due to its single most interesting feature, Tiled Resources.

Analogous to OpenGL mega-textures, Tiled Resources is a feature game developers take to because it utilizes monolithic, large textures instead of countless small ones in a 3D scene by allocating portions of the texture to covering various objects on demand. This is a tangibly more efficient method of resource management and heralds a kind of virtual-memory system for GPUs. Proliferation of DirectX 11.2 doesn't just have lack of adoption by NVIDIA as a roadblock, there's also the OS limitation. DirectX 11.2 is not supported on Windows 7 and requires Windows 8.1, yet Windows 8 and its successor, Windows 8.1, are incredibly unpopular with PC gamers (at least our forum members) due to the idiotic user-interface that insults our intelligence.