The HD7900 series will feature the new Tahiti core which is the successor to the Caymen core of the HD6900 family. The Caymen core was based on a 40nm manufacturing process so it is fair to say that perhaps the biggest improvement to the Tahiti core is the reduction to 28nm process. This has enabled AMD to cram 4.31 billion (4,312,711,873) transistors into the new HD7970 core as opposed to the 2.64 billion on the HD6970 Caymen XT.
Comprising of what AMD calls 'Graphics Core Next (GCN)' architecture which is optimised for heterogeneous computing, the new Tahiti core has upto 32 CU's (compute units) with dual geometry engines along with 8 back-ends comprising of 32 colour ROPs per clock and 128 Z/Stencil ROPs per clock. Each CU has a Vector unit and Scaler Co-Processor and can execute instructions from multiple kernels.
The VLIW4 SIMD has evolved to the GCN quad SIMD which equates to similar computational power yet grants better performance, threading and simplification. The GCN also has a new Cache hierarchy with 16KB instruction Cache and32KB Scaler Cache shared between 4 CU's. The L2 Cache is accessible by each CU with it's own separate process to ensure coherency. These changes increase the GFLOPs by 140%.
Tesselation on the Tahiti core is also enhanced by increased Vertex re-use off-core which theoretically at least, means the 7900 series has 4 times the throughput of the 69XX series. The are improved ansioscopic filtering algorithms too which come at no performance cost allegedly. These new algorithm improvement will allow the use of Forward Rendering with 1000's of potential light sources and will also mean MSAA can be rendered easily.
Looking at the schematics above, we can identify the bulk of the GPU is taken up with the 32 CU's. Above these are the geometry engines. To the left and right are the 8 Render back-ends, 4 each side. There's 768KB of L2 cache along with 6x64bit dual channel memory controllers making a 384-bit GDDR5 memory interface capable of delivering up to 264GB/sec. All of technology coupled together is transferred through a fully compatible PCI Express 3.0 x16 bus interface. Have no fear though because the cards are backwards compatible with PCIe 2.0.
Comparing the HD7970 against it's closest competitor, the GTX580, the AMD card has 4x more processors twice as many texture units but less ROPs. No change with AMD clockspeeds being higher than the green team and with the HD7970 being clocked to a blistering 925MHz I worry about how much is left in the bank to allow us any worthwhile overclocking. As usual, we will investigate this later in the review in the overclocking section. A big change for the AMD card is the leap towards a 320-bit memory interface which AMD have traditionally played second fiddle to NVIDIA. The texture fill rate which has always been a strong point for AMD is increased to a mind boggling 118.4 billion/sec compared to the GTX580's miserly 49.4 while the HD7970 trails slightly in the pixel fill rate stakes ever so slightly. AMD have fired a warning shot across NVIDIA's bow too with a huge 3GB of GDDR5 which should put paid to any difficulties running this card at high resolutions with filters applied.
On paper then, it appears AMD have been squireling away to create a graphics card behemoth capable of besting NVIDIA's flagship single core GPU. I have little doubt that it would be faster but AMD would do well to keep an eye on the horizon as traditionally, the pendulum swings from one manufacturer to the other in a very short time. For now though, AMD have manufactured a card that is by far the most powerful card on test, on paper at least. We have seen many times in the past though that specifications mean very little if the GPU runs like a 3 legged donkey, how this transpires today we will see later in the benchmarks tests...