As we’ve mentioned, at the heart of AMD’s Radeon HD 7900 series graphics cards lies a totally new GPU codenamed Tahiti, which is based on a new architecture dubbed Graphics Core Next.
According to AMD, GCN was three years in the making and is designed to be scalable in ways AMD couldn't achieve before. The architecture is tuned for high graphics and compute performance and optimized for parallel and heterogeneous computing applications.
The cores in Tahiti have been completely redesigned from scratch. In its full configuration, the GPU features 2048 Stream Processors, arranged in 32 compute units, each with 64 vector units (4 x 16), and a 384-bit wide GDDR5 memory controller. The GPU has dual geometry engines and 8 render back ends than can process 32 color ROPs and 128 Z/Stencil ROPs per clock. The geometry engines and render back ends are also up to 50% more efficient than AMD’s older architectures. There is 768KB of read/write L2 cache, and L1 caches have been doubled over the previous generation as well. Not only is there more cache, but the caches have more bandwidth as well. In fact, AMD claims 2TB/s of bandwidth across all 32 caches on the chip, when operating at 925MHz.
The Tahiti GPU is manufactured using TSMC’s advanced 28nm process node and is comprised of a whopping 4.3 billion transistors, more than double the 2.1B transistors of the Cayman GPU used in the Radeon HD 6900 series. Dies size for the chip is 365mm2.
Graphics Core Next features a new instruction set, that’s not VLIW like previous generations and is more flexible and easier to program. The primary graphics pipeline and two compute pipelines can work independently of each other. Each compute unit can also execute instructions from multiple kernels simultaneously. The architecture is designed for increased instruction per clock per mm2 and is designed for high utilization, high throughput, and multi-tasking.
Radeon HD 6970 vs. 7970 Specification Comparison
As implemented on the Radeon HD 7970, the Tahiti GPU is clocked at 925MHz and offers up to 3.79TFLOPS of compute performance. There are 128 texture units (up from 96 in the Radeon HD 6900 series), which off a peak texture fillrate of 118.4GT/s and 32 ROPs; peak pixel fillrate is up slightly from the 6970 to 39.6GP/s.
Radeon HD 7970 boards will also feature 3GB of GDDR memory running at 1,375MHz, for an effective data rate of 5.5Gbps. The memory is attached to the GPU via a 384-bit wide memory bus, resulting in peak memory bandwidth of a 264GB/s, which is a huge leap over the 6970’s 176GB/s peak. Although the memory is clocked at a similar rate, the wider bus on the 7970 affords much more bandwidth. Board power is similar between the Radeon HD 7970 and 6970 at a peak of 250W.