3D accelerator cards, known as GPUs, were only found in high-end workstations until the late 1990s, as they did not have the capacity to run video games and other applications on a desktop CPU alone.
Later extensions including MMX, 3DNow! and SSE provided significant performance boosts for games that supported it, but the GPU took over all SIMD vector tasks.
GPUs are poor at running arbitrary computing tasks, and communication via PCIe links is much slower than within GPU and CPU dies, leading to the increasing division of workloads between the two, such as Sony’s addition of a GPU to its PlayStation 3.
However, there are signs of a reversal of this trend, with Intel’s Larrabee and IBM’s Cell processor merging CPU and GPU characteristics, and DirectStorage API in DirectX adding CPU features to GPUs.
Perhaps the future lies inCombining CPU and GPU characteristics on a single die, as seen in Intel’s Larrabee and IBM’s Cell processor. However, developing for such a new kind of architecture can be challenging, as shown by Sony’s addition of a GPU to its PlayStation 3.
