Scan's TekSpek

Introduction

All personal computers (PCs) use some form of graphics to output the display on to a monitor or screen. Looking back over the last 15 years, there has seen significant development in what are termed 3D accelerators - dedicated graphics boards designed to render life-like images in real-time - and the market has been dominated by two companies: NVIDIA and ATI.

ATI, part of processor company AMD since 2006, released its latest graphics processing units (GPUs) in September 2009. Known as the Radeon HD 5800-series, this TekSpek explains what they are, which new features they bring to the table, and how they compete against NVIDIA's price-comparable GPUs.

Radeon HD 5800

Following on the back of the Radeon HD 4800-series, launched in June 2008, ATI has increased performance in practically every parameter by using a collection of cutting-edge technologies.

Bracketed in terms of price, ATI currently has two cards in the Radeon HD 5800 stable: the Radeon HD 5870 1,024MB ($399) and Radeon HD 5850 1,024MB. The two cards share the same basic architecture but the Radeon HD 5850's processing and, consequently, performance is pared-down to hit the lower price point.

Radeon HD 5870 and HD 5850 cards, currently available at most etailers, are being marketed by a large number of add-in board (AIB) partners.

Wider and faster - how?

Graphics workloads are inherently parallelisable, which means that programs can be broken down into instructions and run concurrently across a number of execution units. In an oversimplified sense, then, the greater the number of units, the faster the overall processing.

The Radeon HD 5800-series adds in greater processing power by packing in more of practically everything when compared to the previous-generation HD 4800-series, as shown by the table, below.

Adding in a larger number of processing units - called Stream Processors by ATI - means that both the Radeon 5-series cards have greater compute power when directly compared against the Radeon HD 4870, to the tune of 2.3x for the range-topping HD 5870.

Theoretical compute power is only one part of the GPU story, however. A GPU's memory-bandwidth is important because it enables the processing of effects once they've been 'computed', and having more bandwidth ensures that the GPU doesn't stall when the image-quality and resolution - usually selected in-game - are dialled-up.

What's the cost of extra performance?

In order to deliver extra performance from a wider, faster design, ATI has increased the transistor-count - the building blocks of computing silicon - by a factor of 2.2x, to 2.15bn.

Increasing sheer processing power this way means that heat becomes a real concern. ATI has used the latest 40nm process, manufactured by a fabrication company called TSMC, for the Radeon HD 5800-cards. Transistors based on 40nm are significantly smaller and more power-efficient than on 55nm. By adding more heat-generating transistors but reducing the per-transistor power-draw, ATI has managed to keep overall under-load board power below 200W.

Tellingly, idle power-draw has dropped dramatically, helped by better power regulation and improvements in chip fabrication.

Not just about hardware - DX11, OpenCL, DirectCompute, GPGPU

Rendering images is more than just about harnessing increased processing power. Modern graphics cards need to meet industry-wide standards such that developers can author code that will run on cross-company hardware.

Graphics-wise, the most common standards are DirectX and OpenGL. Known as APIs (Application Programming Interface), newer iterations provide a cleaner and tighter feature-set that enables better-looking effects with increased rendering speed.

Most games are based around Microsoft's DirectX. The latest iteration, DX11, will ship with the Windows 7 operating system on October 22, 2009, although users of Microsoft Vista will be able to download the update from DX10 to DX11 via Microsoft's servers a short while later.

DX11 main benefits are that it brings support for what's termed hardware tessellation - the ability for a GPU to create high-detail models from basic geometry, increasing efficiency - better multi-threading support for many-core CPUs that ship in the majority of systems today, and simpler access for GPGPU (General Processing on a Graphics Processing Unit) calculations.

The highly parallel, high-throughput design of modern GPUs enables them to perform certain non-gaming tasks at speeds far in excess of any CPU. Parallelisable workloads such as video-encoding and, in the professional space, complex calculations pertaining to oil and gas exploration/academia/medical world run especially well on a GPU, assuming that the application is well-coded. To this end, the Radeon HD 5800-series supports features such as high-speed double-precision computation, as well as newer standards such as OpenCL and DirectCompute, too.

Display choices

Another improvement of the new GPUs is support for a greater number of digital displays. Marketed by ATI under technology known as Eyefinity, both 5-series GPUs are internally plumbed to run up to six independent displays, although the majority of cards will feature three outputs, most likely DVI, HDMI, and DisplayPort.

The competition

Arch-rival NVIDIA has yet to launch a new range of cards, and industry watchers reckon that ATI has, at the very least, a three-month window of opportunity with the HD 5800-series.

However, at a GPU computing conference held in San Jose, California in late-September 2009, NVIDIA provided basic details for its competing products. Codenamed Fermi, it is practically impossible to determine whether retail models will be better GPUs than ATI's 5-series.

NVIDIA, it seems, will be deliberately architecting its GPUs to perform better in a GPGPU environment, backed up by the company's efforts in pushing an NVIDIA-only GPU-programming language known as CUDA, which helps developers and academics translate and port their computational problems on to the hardware via commonly-used C+.

Summary

The introduction of the Radeon 5870 and Radeon HD 5850 GPUs means that ATI has the fastest single-GPU graphics cards available. Greater processing is made possible by a more-efficient manufacturing process. Adherence to Microsoft's DX11 enables the GPUs to take advantage of newer gaming titles that use the API for better-looking effects which should, in some cases, render with increased speed.

ATI has a significant opportunity to leverage the architecture and derivate such that the bulk of its GPU catalogue, from top to bottom, is composed of 5-series parts before NVIDIA is able to launch its own parts, currently codenamed Fermi.