ATi Radeon R100 Series

Architecture

The first-generation Radeon GPU was launched in 2000, and was initially code-named Rage 6 (later R100), as the successor to ATI's aging Rage 128 Pro which was unable to compete with the GeForce 256. The card also had been described as Radeon 256 in the months leading up to its launch, possibly to draw comparisons with the competing Nvidia card, although the moniker was dropped with the launch of the final product.

The R100 was built on a 180 nm semiconductor manufacturing process. Like the GeForce, the Radeon R100 featured a hardware transform and lighting (T&L) engine to perform geometry calculations, freeing up the host computer's CPU. In 3D rendering the processor can write 2 pixels to the framebuffer and sample 3 texture maps per pixel per clock. This is commonly referred to as a 2×3 configuration, or a dual-pipeline design with 3 TMUs per pipe. As for Radeon's competitors, the GeForce 256 is 4×1, GeForce2 GTS is 4×2 and 3dfx Voodoo 5 5500 is a 2×1+2×1 SLI design. Unfortunately, the third texture unit did not get much use in games during the card's lifetime because software was not frequently performing more than dual texturing.

In terms of rendering, its "Pixel Tapestry" architecture allowed for Environment Mapped Bump Mapping (EMBM) and Dot Product (Dot3) Bump Mapping support, offering the most complete Bump Mapping support at the time along with the older Emboss method.^[3] Radeon also introduced a new memory bandwidth optimization and overdraw reduction technology called HyperZ. It basically improves the overall efficiency of the 3D rendering processes. Consisting of 3 different functions, it allows the Radeon to perform very competitively compared to competing designs with higher fillrates and bandwidth on paper.

ATI produced a real-time demo for their new card, to showcase its new features. The Radeon's Ark demo presents a science-fiction environment with heavy use of features such as multiple texture layers for image effects and detail. Among the effects are environment-mapped bump mapping, detail textures, glass reflections, mirrors, realistic water simulation, light maps, texture compression, planar reflective surfaces, and portal-based visibility.^[4]

In terms of performance, Radeon scores lower than the GeForce2 in most benchmarks, even with HyperZ activated. The performance difference was especially noticeable in 16-bit color, where both the GeForce2 GTS and Voodoo 5 5500 were far ahead. However, the Radeon could close the gap and occasionally outperform its fastest competitor, the GeForce2 GTS, in 32-bit color.

Aside from the new 3D hardware, Radeon also introduced per-pixel video-deinterlacing to ATI's HDTV-capable MPEG-2 engine.

R100's pixel shaders

R100-based GPUs have forward-looking programmable shading capability in their pipelines; however, the chips are not flexible enough to support the Microsoft Direct3D specification for Pixel Shader 1.1. A forum post by an ATI engineer in 2001 clarified this:

...prior to the final release of DirectX 8.0, Microsoft decided that it was better to expose the RADEON's and GeForce{2}'s extended multitexture capabilities via the extensions to SetTextureStageState() instead of via the pixel shader interface. There are various practical technical reasons for this. Much of the same math that can be done with pixel shaders can be done via SetTextureStageState(), especially with the enhancements to SetTextureStageState() in DirectX 8.0. At the end of the day, this means that DirectX 8.0 exposes 99% of what the RADEON can do in its pixel pipe without adding the complexity of a "0.5" pixel shader interface.

Additionally, you have to understand that the phrase "shader" is an incredibly ambiguous graphics term. Basically, we hardware manufacturers started using the word "shader" a lot once we were able to do per-pixel dot products (i.e. the RADEON / GF generation of chips). Even earlier than that, "ATI_shader_op" was our multitexture OpenGL extension on Rage 128 (which was replaced by the multivendor EXT_texture_env_combine extension). Quake III has ".shader" files it uses to describe how materials are lit. These are just a few examples of the use of the word shader in the game industry (nevermind the movie production industry which uses many different types of shaders, including those used by Pixar's RenderMan).

With the final release of DirectX 8.0, the term "shader" has become more crystallized in that it is actually used in the interface that developers use to write their programs rather than just general "industry lingo." In DirectX 8.0, there are two versions of pixel shaders: 1.0 and 1.1. (Future releases of DirectX will have 2.0 shaders, 3.0 shaders and so on.) Because of what I stated earlier, RADEON doesn't support either of the pixel shader versions in DirectX 8.0. Some of you have tweaked the registry and gotten the driver to export a 1.0 pixel shader version number to 3DMark2001. This causes 3DMark2001 to think it can run certain tests. Surely, we shouldn't crash when you do this, but you are forcing the (leaked and/or unsupported) driver down a path it isn't intended to ever go. The chip doesn't support 1.0 or 1.1 pixel shaders, therefore you won't see correct rendering even if we don't crash. The fact that that registry key exists indicates that we did some experiments in the driver, not that we are half way done implementing pixel shaders on RADEON. DirectX 8.0's 1.0 and 1.1 pixel shaders are not supported by RADEON and never will be. The silicon just can't do what is required to support 1.0 or 1.1 shaders. This is also true of GeForce and GeForce2.

R100

The first versions of the Radeon (R100) were the Radeon DDR, available in Spring 2000 with 32 MB or 64 MB configurations; the 64 MB card had a slightly faster clock speed and added VIVO (video-in video-out) capability. The core speed was 183Mhz and the 5.5 Ns DDR SDRAM memory clock speed was 183 MHz DDR (366 MHz effective). The R100 introduced HyperZ, an early culling technology (maybe inspired by the Tile Rendering present in St Microelectronics PowerVR chips) that became the way to go in graphic evolution and generation by generation rendering optimization, and can be considerend the first non tile rendering-based (and so DX7 compatible) card to use a Z-Buffer optimization. These cards were produced until mid-2001, when they were essentially replaced by the Radeon 7500 (RV200).

A slower and short-lived Radeon SDR (with 32 MB SDRAM memory) was added in mid-2000 to compete with the GeForce2 MX.

Also in 2000, an OEM-only Radeon LE 32MB DDR arrived. Compared to the regular Radeon DDR from ATI, the LE is produced by Athlon Micro from Radeon GPUs that did not meet spec and originally intended for the Asian OEM market. The card runs at a lower 143 MHz clock rate for both RAM and GPU, and its Hyper Z functionality has been disabled. Despite these handicaps, the Radeon LE was competitive with other contemporaries such as the GeForce 2 MX and Radeon SDR. Unlike its rivals, however, the LE has considerable performance potential, as is possible to enable HyperZ through a system registry alteration, plus there is considerable overclocking room. Later drivers do not differentiate the Radeon LE from other Radeon R100 cards and the HyperZ hardware is enabled by default, though there may be visual anomalies on cards with HyperZ hardware that is defective.^[5]

In 2001, a short-lived Radeon R100 with 64 MB SDR was released as the Radeon 7200. After this and all older R100 Radeon cards were discontinued, the R100 series was subsequently known as the Radeon 7200, in keeping with ATI's new naming scheme.

RV100

A budget variant of the R100 hardware was created and called the Radeon VE, later known as the Radeon 7000 in 2001 when ATI re-branded its products.

RV100 has only one pixel-pipeline, no hardware T&L, a 64-bit memory bus, and no HyperZ. It did however add HydraVision dual-monitor support and integrated a 2nd RAMDAC into the core (for Hydravision).

From the 3D performance standpoint, the Radeon VE did not fare well against the GeForce2 MX of the same era, though its multi-display support was clearly superior to the GeForce2 MX, however. The Matrox G450 has the best dual-display support out of the GPUs but the slowest 3D performance.

RV100 was the basis for the Mobility Radeon notebook solution.

RV200

The Radeon 7500 (RV200) is basically a die-shrink of the R100 in a new 150 nm manufacturing process. The increased density and various tweaks to the architecture allowed the GPU to function at higher clock speeds. It also allowed the card to operate with asynchronous clock operation, whereas the original R100 was always clocked synchronously with the RAM. It was ATI's first Direct3D 7-compliant GPU to include dual-monitor support (Hydravision).^[6]

The Radeon 7500 launched in the second half of 2001 alongside the Radeon 8500 (R200). It used an Accelerated Graphics Port (AGP) 4x interface. Around the time that the Radeon 8500 and 7500 were announced, rival Nvidia released its GeForce 3 Ti500 and Ti200, the 8500 and Ti500 are direct competitors but the 7500 and Ti200 are not.

The desktop Radeon 7500 board frequently came clocked at 290 MHz core and 230 MHz RAM. It competed with the GeForce2 Ti and later on, the GeForce4 MX440.