Advanced processor architectures (APAZ_07.ppt) Karel Vlček,
[email protected] Institute of Applied Informatics FAI, UTB in Zlin
Graphic processors (GPU) z z
z
z
2
Architectures of specialised processors exist from the middle of sixties of 20-th century The most frequently class are introduced, as the specialised processors, the graphic processors What similar properties, can one see at graphic processors (GPU), and so called central processors (CPU)? And what differences can we see, at graphic processors, and central processors? Karel Vlček
Pokročilé architektury procesorů
Similarities of GPU, and CPU z
z z
3
GPU, as well as CPU will have more then one kernels, where kernel is considered the link of execution units, and working registers consisted to pipelining configuration The GPU, as well as CPU are, in every kernel, a superscalar processor The GPU, as well as CPU has the complex hierarchy of memories Karel Vlček
Pokročilé architektury procesorů
Differences of GPU, and CPU z
z
4
Typical property of GPU are specialised units with high computational power (accelerators) for specific operations Accelerators can be considered as the complex execution units, which are used in the architecture, immediately, or by buses
Karel Vlček
Pokročilé architektury procesorů
Differences of GPU, and CPU z z
z
5
Differences of GPU from CPU will be in interconnection of cache memory The GPU, for example, will need not, compared to CPU, an implementation of ability of branching prediction, (for CPU it is necessary for computational power enhancement) The GPU will offer typical properties called as stream data processing, in the future Karel Vlček
Pokročilé architektury procesorů
GPU → Stream Data Processing (1) z
z z
6
Property Stream Data Processing describe the way of data processing, which asks big amount of operations with operands in fixed, and in floating point format Another asking for GPU is sufficiency of high speed memories with big capacity These asking to architecture of GPU are similar to algorithms of scientific and technique calculations Karel Vlček
Pokročilé architektury procesorů
GPU → Stream Data Processing (2) z
z z
7
The term Stream Data Processing is called if the way of architecture solving, it is putting stress to high degree of parallelism SIMD The big amount of operations are asked in floating point calculations There are not solved aims of memory allocation, synchronisation, as well as communication between processors Karel Vlček
Pokročilé architektury procesorů
Specific properties of GPU z
z z
8
The properties of GPU are actually evaluated by the aims of 3D graphics, (but not only by these properties) The asking of 3D graphics are called Unified Shades Architecture: These aims are: transformation of models, lighting, simulation of camera, rasterisation, texture making, and solving of hidden surface Karel Vlček
Pokročilé architektury procesorů
Solving of GPU properties (1) z
Original solving of GPU properties starts from sequential solving of partial aims:
Description of solution in Vertex
9
Solving of surface
Rasterisation
Karel Vlček
Manipulation with pixels
Memory operations
Pokročilé architektury procesorů
Solving of GPU properties (2) z
Enhanced solution of GPU architecture uses intensive exchange of data of partial aims, and memory:
Description of aim in the Vertex
Solving of surface
Rasterisation
Manipulation with pixels
Memory operations
Output data flow Memory
10
Karel Vlček
Pokročilé architektury procesorů
Evaluation of GPU computational power z
z z z
11
The computational power of GPU is evaluated by specially designed programs (Benchmarks) Programs are designed by independent organisations, which interests in this The GPU power is done not only the processor, but by applied subjects There are defined so called texels (TEXture ELements), which described asking for amount of operations better Karel Vlček
Pokročilé architektury procesorů
Co-operation of CPU and GPU
z
z z z
12
Different tests of GPU interests in: Co-operation of CPU and GPU in various conditions of displaying: parameter FPS (Frame per Second) Ability of fluent work of GPU during operation rendering of objects – the main activity Relation in the power GPU and CPU (so called bottleneck) – what processor „delays“? Heating and cooling Karel Vlček
Pokročilé architektury procesorů
Další vývoj GPU z
z z
z
13
Hlavní překážkou při zvyšování výkonu GPU je ztrátový příkon při vysoce paralelním způsobu zpracování dat Cesta se zvyšováním počtu výpočetních jader zde nemusí být tak účinná, jako u CPU Důvodem je jistě i to, že na paměti, které spolupracují s GPU, jsou kladeny stejně vysoké nároky, jako na výkonnou jednotku Příkon desky, kde je CPU a GPU je 1kW a více Karel Vlček
Pokročilé architektury procesorů
Literature: z z z z z
14
Dvořák, V.: Architektura a programování paralelních systémů, VUTIUM Brno, (2004), ISBN 80-214-2608-X Dvořák, V., Drábek, V.: Architektura procesorů, VUTIUM Brno, (1999), ISBN 80-214-1458-8 Drábek, V.: Výstavba počítačů, PC-DIR, s.r.o. Brno, (1995), ISBN 80214-0691-7 Mueller, S.: Osobní počítač, Computer Press, Praha, (2001), ISBN 807226-470-2 Pluháček, A.: Projektování logiky počítačů, Vydavatelství ČVUT Praha, (2003), ISBN 80-01-02145-9
Karel Vlček
Pokročilé architektury procesorů