Nvidia GeForce 8800
Posted by Bryan Catanzaro on November 8, 2006
Today is a big day in the GPU world, since Nvidia's releasing a brand new architecture! This one's sometimes referred to as G80, which will be productized into various shapes and sizes of GPU in the Geforce 8000 series.
This architecture is noteworthy in that it is the first GPU for the desktop with unified shaders (although the ATi GPU in the Xbox 360 had them a year ago). In other words, its processing elements are more general purpose and programmable than previous architectures, which had assemblages of more specialized units. The Geforce 8800, released today, has 128 of these shaders running at around 1.35 GHz. Nvidia is calling them "stream processors" in order to emphasize their more general purpose nature. Interestingly, they are scalar processors, which Nvidia claims offers more flexibility than the traditional vector processors used in GPUs.
The 128 stream processors are organized in hierarchical fashion: groups of 16 share an "L1" cache, and then 8 groups of shaders are connected via a crossbar to 6 rasterization groups, each of which contains an "L2" cache and a 64-bit channel to GPU memory.
Nvidia is touting the general purpose nature of this architecture, and claiming large advantages on many types of code over CPUs - like 10x on rigid body physics simulation to 200x on financial calculations. There are 2 separate drivers: one for games, and a new one for exposing the architecture to general purpose programming. The programming model looks to be a variant of C with threads.
This GPU is the first to support Microsoft's DX10 standard, which brings two important capabilities to the GPU: Virtual Memory, meaning that the graphics card's memory can now be virtualized by the OS, and Multitasking - in contrast to previous GPUs which assumed only one process was using the GPU at a given time. Also, a "higher" degree of compliance with 32-bit IEEE floating-point standards is claimed, although specifics weren't easy to find.
Benchmarks are impressive, as befits such a monstrous chip with almost 700M transistors. =)
This release continues the trend of GPUs becoming more programmable and intruding on traditional CPU territory. The stage continues to set up for a battle of titans over PC silicon between Intel, AMD/ATi and Nvidia.
Some good reviews:
This architecture is noteworthy in that it is the first GPU for the desktop with unified shaders (although the ATi GPU in the Xbox 360 had them a year ago). In other words, its processing elements are more general purpose and programmable than previous architectures, which had assemblages of more specialized units. The Geforce 8800, released today, has 128 of these shaders running at around 1.35 GHz. Nvidia is calling them "stream processors" in order to emphasize their more general purpose nature. Interestingly, they are scalar processors, which Nvidia claims offers more flexibility than the traditional vector processors used in GPUs.
The 128 stream processors are organized in hierarchical fashion: groups of 16 share an "L1" cache, and then 8 groups of shaders are connected via a crossbar to 6 rasterization groups, each of which contains an "L2" cache and a 64-bit channel to GPU memory.
Nvidia is touting the general purpose nature of this architecture, and claiming large advantages on many types of code over CPUs - like 10x on rigid body physics simulation to 200x on financial calculations. There are 2 separate drivers: one for games, and a new one for exposing the architecture to general purpose programming. The programming model looks to be a variant of C with threads.
This GPU is the first to support Microsoft's DX10 standard, which brings two important capabilities to the GPU: Virtual Memory, meaning that the graphics card's memory can now be virtualized by the OS, and Multitasking - in contrast to previous GPUs which assumed only one process was using the GPU at a given time. Also, a "higher" degree of compliance with 32-bit IEEE floating-point standards is claimed, although specifics weren't easy to find.
Benchmarks are impressive, as befits such a monstrous chip with almost 700M transistors. =)
This release continues the trend of GPUs becoming more programmable and intruding on traditional CPU territory. The stage continues to set up for a battle of titans over PC silicon between Intel, AMD/ATi and Nvidia.
Some good reviews: