AMD’s Radeon HD 7970: 4.3 Billion Transistors of Pure Performance

Each of AMD’s GCN Compute Units accesses a dedicated L1 read/write cache. Each group of four units shares a 16KB instruction cache and a 32KB scalar data cache.

AMD’s new Graphics Core Next architecture signals a major shift from building GPUs that are primarily for graphics as well as some compute chores, to building a high-performance parallel-processing engine that’s capable of excellent compute performance without sacrificing graphics performance.

Tahiti is assembled from 32 GCN compute units, which translates to 2,048 stream processors, each of which is based on AMD’s new SIMD-plus-scalar architecture.

The only released images of the 7970s don't really tell us much about the commercial products: they look pretty much just like every other AMD reference design.