TimelineNVIDIA GeForce 256

NVIDIA GeForce 256

The graphics card that gave the industry a new acronym

1999 Graphics card NVIDIA Oct 11 release

Announced on 31 August 1999 and on shelves from 11 October, the GeForce 256 moved geometry maths off the CPU and onto the graphics chip. NVIDIA called it the world's first GPU, and the label stuck to the whole industry.

By the middle of 1999 the 3D accelerator market had settled into a rhythm. 3dfx owned the retail shelf with the Voodoo line, NVIDIA was pushing into the high end with the RIVA TNT2, and Matrox, S3 and ATI each held a corner. Every one of those cards did roughly the same job: rasterise triangles that the CPU had already transformed, lit and clipped. The GeForce 256, codenamed NV10, was the first mainstream part to take that geometry work for itself.

A big chip with a bigger claim

The numbers were substantial for a consumer part in 1999. NVIDIA built NV10 at TSMC on a 220 nm CMOS process, packing 17 million transistors onto a 139 square millimetre die. The core ran at 120 MHz, fed by either 32MB or 64MB of SDRAM on a 128-bit bus, and the card plugged into an AGP 4x slot, with PCI versions also offered. The SDR model shipped first at an MSRP of $199. A DDR version followed on 13 December 1999 at $249, raising the core to 150 MHz and memory bandwidth from 2.65 GB/s to 4.80 GB/s.

The name described the architecture, not a memory size. NV10 used a "256-bit QuadPipe Rendering Engine", four 64-bit pixel pipelines. In single-textured workloads the chip could put out four pixels per cycle; with two textures per pixel it fell to two, because each pipeline carried a single texture unit, exactly as the TNT2 had. What was genuinely new was Direct3D 7 compliance, cube environment mapping, dot-product bump mapping and, most importantly, a fixed-function transform and lighting engine.

Transform and lighting, in hardware

Transform and lighting is the stage where a 3D scene's vertices are converted from model space into the camera's view, then shaded according to the lights in the scene. For most of the 1990s that work fell to the CPU, which meant a slow processor could bottleneck a fast graphics card, and a fast processor was wasted once the accelerators caught up. Moving it into the GPU meant the card could sustain geometry throughput regardless of the host machine, and NVIDIA set its own bar for the new category at ten million polygons per second.

Because the same silicon could now do the professional work that had previously required specialist boards, NVIDIA spun off the Quadro line from the identical chip, differentiated by drivers and certification rather than hardware. It was the company's first serious entry into computer-aided design, and a template it would use for the next two decades.

The acronym that outlived the card

NVIDIA's marketing department did the most consequential work of the launch. The company defined a GPU as "a single-chip processor with integrated transform, lighting, triangle setup and clipping, and rendering engines that is capable of processing a minimum of 10 million polygons per second", then declared the GeForce 256 the world's first. Competitors objected, correctly noting that workstation chips had integrated some of those functions before, but the phrase was short, plausible and arrived at the moment 3D gaming went mainstream. Every graphics chip sold since has been called a GPU.

The software arrived late

Reviewers liked the hardware and doubted the feature. Hardware transform and lighting only pays off when a game engine uses it, and in late 1999 almost none did. The benefit showed up most clearly in OpenGL first-person shooters, with Quake III Arena the showcase title, and in benchmarks run on cheap CPUs: a Celeron 300A paired with a GeForce 256 could beat a Pentium II 300 running an older card such as the 3dfx Voodoo 2. On fast processors the picture was less flattering, which 3dfx and other rivals were quick to point out. Early drivers were slow and buggy as well, while 3dfx owners enjoyed a mature Glide path in most shipping games.

The result was a card that measured better than it played for its first year. Hardware transform and lighting only became routine after the GeForce 2 arrived and ATI shipped a T&L-capable Radeon.

Longer legs than it was given credit for

The GeForce 256 also carried hardware motion compensation for MPEG-2 video, the first implementation of its kind on an NVIDIA chip, later replaced by the company's HDVP video processor in the GeForce 2 GTS. That unit took part of the MPEG-2 decoding work off the CPU, a benefit that had nothing to do with frame rates.

Its real longevity came from the family it seeded. GeForce 2 MX and GeForce 4 MX parts kept the same fixed-function feature set in cheap machines for years, so developers kept a Direct3D 7 path alive long after the original card left the market. Games as late as Half-Life 2 and Star Wars: Empire at War could still run on one, the former through a Direct3D 7 compatible path that used a subset of Direct3D 9 to target the fixed-function pipeline. NVIDIA only retired the series in 2005: driver 71.84, released on 11 March 2005, was the last for Windows 9x and Windows Me, and 71.89 followed for Windows 2000 and 32-bit XP.

What it set in motion

NVIDIA's own retrospectives now treat the GeForce 256 as the start of the line that ran on through CUDA and the deep learning boom. The card itself is remembered for a definition rather than a benchmark record, and NVIDIA had replaced it with the GeForce 2 inside a year. Still, the idea it fixed in place, that geometry and shading belong on a dedicated processor rather than a general one, is the assumption every graphics chip since has been built on.

Sources

← See NVIDIA GeForce 256 on the gadget timeline