Over the last 15 years, graphics processors have continually evolved and are on the cusp of becoming an integral part of the computing landscape. The first designs employed special purpose hardware with little flexibility. Later designs introduced limited programmability through shader programs, and eventually became highly programmable throughput computing devices that still maintained many graphics-specific capabilities.
The performance and efficiency potential of GPUs is incredible. Games provide visual quality comparable to leading films, and early adopters in the scientific community have seen an order of magnitude improvement in performance, with high-end GPUs capable of exceeding 4 TFLOPS (Floating point Operations per Second). The power efficiency is remarkable as well; for example, the AMD Radeon™ HD 7770M GPU achieves over 1 TFLOPS with a maximum power draw of 45W.
Key industry standards, such as OpenCL™, DirectCompute and C++ AMP recently have made GPUs accessible to programmers. The challenge going forward is creating seamless heterogeneous computing solutions for mainstream applications. This entails enhancing performance and power efficiency, but also programmability and flexibility. Mainstream applications demand industry standards that are adapted to the modern ecosystem with both CPUs and GPUs and a wide range of form factors from tablets to supercomputers.
AMD’s Graphics Core Next (GCN) represents a fundamental shift for GPU hardware and is the architecture for future programmable and heterogeneous systems. GCN is carefully optimized for power and area efficiency at the 28nm node and will scale to future process technologies in the coming years. The heart of GCN is the new Compute Unit (CU) design that is used in the shader array, which is described in detail in the following sections.
AMD’s GCN Architecture comes at a time of change for the industry. Graphics is an increasingly important part of the user experience, and a crucial component for SoCs that integrate CPUs and GPU side-by-side. The mandate for GPUs is expanding to include not just 3D rendering, but new general purpose, heterogeneous applications such as facial recognition, which are only feasible using the parallel performance of the GPU.
As a company, AMD is uniquely positioned with deep expertise and a long history of excellence in both CPUs and GPUs for the PC. GCN is a marriage of these domains, melding the reliability and programmability of traditional CPUs with the efficient parallel performance of a GPU. GCN is huge step forward, firmly placing AMD in the new era of heterogeneous computing, but without losing sight of efficiency or performance. The GCN Architecture encompasses innovations such as virtual memory, coherent caching and an elegant hybrid vector/scalar instruction set that are revolutionary. At the system level, GCN is the only discrete graphics architecture that is compatible with the x86 programming model, paving the way for future software and hardware integration.
Most importantly, AMD has carefully crafted an architecture that is a tremendous advance in programmability, but does not sacrifice performance or efficiency. Features such as a unified instruction stream and scheduling in the scalar pipelines enhance utilization and boost the achievable performance on real workloads. The first GPUs manufactured on 28nm were based on GCN and improved performance/watt and performance/mm2 by roughly 50% over the prior generation. The AMD Radeon™ HD 7970 GHz Edition achieves peak performance of over 1 TFLOPS double precision and 4 TFLOPS single precision, a testament to the underlying architecture. GCN will eventually revolutionize AMD’s entire product line, from tablets to supercomputers.