General computing on GPUs is actively promoted by the main manufacturers of graphics chips, who pay special attention to this issue. GPU computing is gradually integrating with regular software, used in everyday tasks, especially in image and video processing fields.
As we have already mentioned, RADEON HD 5870 possesses very high performance: up to 2.7 teraflops for single-precision computing and 544 gigaflops for double precision. It's much better than in previous solutions from AMD as well as competing cards from NVIDIA. Especially with double-precision computing. That's why it's a very attractive chip for GPU computing.
It has full hardware support for all features of DirectCompute11 and OpenCL 1.0 and some more. This GPU supports Memory Export, scatter operations up to 64 32-bit values per cycle, which is twice as many as in the prev-gen GPUs RV770/RV790.
Extra functions exceeding OpenCL 1.0 features include: 32-bit atomic operations, 32KB shared memory, 64KB global memory, global syncing, append/consume buffers.
AMD has always favored OpenCL support (open API for cross-platform applications for high-speed computing on CPU and GPU). OpenCL allows parallel development and CPU/GPU usage with the same commands from the same API. AMD was the first to release OpenCL for CPU, and it's very important for them to support CPU and GPU equally, as this company manufactures both.
We are pleased to say that the increasing number of usual programs get the GPU computing support. For example, multimedia software supporting ATI Stream includes such popular applications from Cyberlink as MediaShow 5, MediaShow Espresso, PowerDirector 8 (Version 8.1 and higher supports DirectCompute 11) and PowerDirector 7, as well as ArcSoft SimHDT plugin for TotalMedia Theatre and Roxio Creator 2010.
For example, PowerDirector 8 accelerates about 20 effects using AMD GPUs, and the latest solutions provides more than a twofold performance gains compared to CPU. Cyberlink is of the opinion that software development for DirectCompute alone is more efficient than the old approach with several APIs for generic computing on GPUs.
It goes without saying that AMD couldn't possibly ignore the issue of GPU physics computing, considering the success of their competitor here. Unfortunately, the so-called "open physics" is used in few games right now, and we don't see this number increasing. These open APIs (Bullet Physics and Pixelux) are not widely known to the public, although they have been used in several games.
These packages contain physics of soft bodies, deformed objects, destroyable objects, etc. As for now, AMD works only with these companies to add OpenCL support to their software. And we don't know when it's available in the market, to say nothing of ready games. They started the ball rolling, we'll see what happens.
One of interesting, but not very new features (just remember Matrox, who used to manufacture graphics cards for games supporting three monitors) is extended support for multi-monitor configurations. It may be useful to connect three or six monitors to a single graphics card in games and in professional applications. Here are some layouts, offered by AMD:
Video output in Cypress was overhauled to support up to six displays (using DisplayPort and a special modification of the card -- HD 5870 SIX) in various combinations. The number of supported monitors depends on a given configuration of the card, but it can reach six or three. The GPU has six integrated TMDS transmitters. In usual (not SIX) graphics cards four of them support two Dual Link DVIs, so it's impossible to plug six monitors to them, only three.
Multi-monitor configurations can work in clone and extended desktop modes. One big image can be composed of several monitors, it applies to desktops as well as to full-screen video and 3D applications (supported in Windows 7, Windows Vista, and Linux).
On the whole, there is nothing new in multi-monitor configurations, it support in applications that matters. AMD publishes a long list of games that support Eyefinity. Applications just need to work correctly with images of different aspect ratios, everything else works fine. In some cases it looks really good (you should try to play with such a system, photos do not make justice to it):
AMD also works with monitor manufacturers, with Samsung in particular. They manufacture special modifications of 23" monitors supporting 1920x1080, DisplayPort, DVI and VGA, as well as a very thin frame around the panel (just 7-8mm). However, 7-8mm must be doubled, as monitors are installed size by side. So we get a nasty frame that interferes with perceptual unity of the entire image.
On the whole, we cannot say that support for three or six monitors is such an important advantage. It's useful in some (rare) cases -- that's more like it. Will many users install three monitors instead of one large display (FullHD TV set, for example)? Are there that many users, who really need more than two monitors? It's just a useful feature, nothing more.
There is also something new in terms of HDMI support. We now have support for HDMI 1.3a with Dolby TrueHD and DTS-HD Master Audio, all audio formats of Blu-ray, AC3, and DTS. It's actually the first card that can transmit audio tracks in Dolby TrueHD and DTS HD format via HDMI to a receiver. That's not really necessary for a powerful gaming card, but it may be quite handy for future solutions for other price ranges.
Video stream in HDMI 1.3a can be in Deep Color and x.v.Color formats. It gives us a lot of simultaneously displayed colors via HDMI. A wide color range is supported: 12-bit and 10-bit color formats, wide-gamut xvYCC video signal.
Dynamic power management technology ATI PowerPlay has been slightly improved. A special control circuit in a GPU monitors its load and determines an optimal operating mode by controlling clock rates of a GPU, memory, voltages, and other parameters, optimizing power consumption and heat release. Voltages and frequencies (as well as fan speed) will be minimized under low 2D load. Under average 3D load all parameters will be set to medium. And when the GPU works at full capacity, voltages and clock rates will be set to maximum.
Power consumption of the new cards in the idle mode is reduced by lower GPU/memory frequencies and voltages as well as a special GDDR5 memory mode -- low power strobe mode. Hardware monitoring of VRM temperature is implemented to avoid its overheating. Multi-GPU configurations now have a special mode -- Ultra Low Power State (ULPS).
Windows 7 support
This operating system is one more excuse to show of some advantages, even if they are available not only in this operating system. AMD singles out high-quality support for Windows 7 in its drivers, as if they are not the same drivers for Windows Vista.
Along with acceleration of the new operating system to be released soon, AMD programmers raised performance in games. Compared to Vista, AMD obtained 5-20% performance gains in single-GPU and 10-50% in multi-GPU (CrossFire) configurations. The highest performance gains are marked in Far Cry 2 (10-20% in single-GPU and over 20% in multi-GPU modes), BattleForge (20% in the single-GPU mode) and Company of Heroes (50-60% in CrossFire). I don't understand why we cannot have the same results in Vista.
Hardware acceleration in Windows 7 is not limited to games. 3D acceleration is used for Aero Glass interface, Direct2D API. Besides, there is improved support for hardware-assisted video decoding -- DXVA. This GPU uses improved UVD 2.0, which decodes two high definition video streams simultaneously (1920x1080) and works simultaneously with Windows Aero acceleration. The new GPU accelerates not only video decoding, but also transcoding from one video format into another. Windows 7 has such a function, which can be accelerated by the new solutions from AMD.
So, we've examined all features of the new GPU from AMD in theoretical terms. The next part of the article is devoted to practice. In this part we are going to compare performance of the new Cypress solutions with that of the previous products from AMD and competing cards from NVIDIA. Will RADEON HD 5870 be limited by memory bandwidth in synthetic tests? Will it be twice as fast as the HD 4890? Read the next part of the review.
Write a comment below. No registration needed!