Methods for semiconductor systems for graphics processing

Details Methods for semiconductor systems for graphics processing Methods for semiconductor systems for graphics processing

1999-03-08

2003-05-06

Zimmerman, Mark (Department: 2671)

Computer graphics processing and selective visual display system

Computer graphics display memory system

Graphic display memory controller

C345S501000, C345S530000, C345S538000

Reexamination Certificate

active

06559851

ABSTRACT:

FIELD OF THE INVENTION
The disclosed invention relates generally to computer systems. More particularly, the present invention relates to a dynamic random access memory (DRAM) frame buffer device, and system based on that device, which provides an architecture for performing accelerated two-dimensional and three-dimensional graphics rendering operations.
BACKGROUND
The Dual Pixel 3DRAM chip and graphics processing system is used to implement high performance, high capacity frame buffers. Certain aspects of the disclosed Dual Pixel 3DRAM chip, and graphics processing system based upon the Dual Pixel 3DRAM chip, are disclosed in U.S. Pat. No. 5,544,306, issued to Deering et al. on Aug. 6, 1996, which patent is incorporated by reference in its entirety into this disclosure as if it were fully set forth herein.
The disclosed invention presents an alternative to the use of external DRAM frame buffers. To meet near term performance objectives, it is tempting to use embedded DRAM for frame buffer memory, because it might be feasible to fit between 4 and 8 Megabits on a die with surface area remaining to implement an interesting amount of logic. However, in the same time frame, graphics-oriented computing products will require between 10 and 80 Megabits of frame buffer memory. Thus, between 2 and 10 embedded DRAM devices would be necessary to implement a frame buffer that would meet the requirements of graphics processing computing systems. While the fill rate for such a frame buffer would be very high, the cost would be prohibitive for a large segment of the computing market.
In processing two-dimensional and three-dimensional graphic images, texture mapping must be accelerated to match fill rate. However the distributed frame buffer described above makes this difficult to do efficiently. The easiest way to distribute texture data would be for each device to have its own copy of everything, however this method is a very inefficient use of embedded DRAM bits. An alternate approach would distribute texture data among the devices comprising a frame buffer such that data is not duplicated. This alternative would use embedded DRAM bits efficiently, but would also requires the routing of massive amounts of texture data between devices.
Using embedded DRAM to implement a texture cache on a single device might be more practical. Texture data would have to be paged in from system memory, which would work more efficiently if the texture data were compressed.
Embedded DRAM could also be used to implement a primitive FIFO between the setup unit and the rasterizer. This FIFO would allow geometry and setup processing to continue while big triangles are being rendered. It could also be used to tolerate the latency of paging and decompressing texture data in from system memory.
Region-based rendering architectures, such as Talisman, PixelFlow, or Oak's WARP 5, render a small portion of the frame buffer on the rendering controller and then transfer the final color values to external DRAM. The controller then renders the next region, and the one after that, until the entire frame is covered.
All of the bandwidth used for hidden surface removal and anti-aliasing remains entirely on the rendering controller, so fill rate is not limited by external bandwidth. All of the storage used for hidden surface removal and anti-aliasing needs only to be implemented for a small portion of the frame buffer and can be kept on the rendering controller.
The big disadvantage of region-based rendering is that all of the geometry for a frame needs to be sorted into regions and stored somewhere before rendering can begin. This requirement generally places an upper limit on the amount of geometry that can be rendered per frame. This limitation is unacceptable for many applications. Some region-based rendering architectures can still function somewhat correctly when given too much geometry, by writing depth and color values for regions to and from external DRAM. However, this implementation loses all of the benefits of region-based rendering, while retaining all of the disadvantages.
Mechanical CAD and other content creation applications cannot tolerate limits on geometric complexity. Such applications require the ability to smoothly trade off geometric complexity for frame rate. Thus neither embedded DRAM nor region-based rendering approaches provide adequate solutions to meet the performance demands and practical cost constraints of present graphics processing applications.
Another concern with respect to frame buffer design is the performance trade off between single-ported and dual-ported frame buffer memories. Dual-ported frame buffers have a dedicated display port which enables the render port to spend more of its time rendering. Typically, a dual-ported frame buffer comprised of video random access memory (VRAM) chips loses only approximately one to two percent of its fill rate to video transfer operations, because its video buffers are quite large. Frame buffers comprised of FBRAM chips (also referred to as 3D-RAM™ chips) lose approximately five to ten percent of their fill rate to video transfer operations, because their video buffers are smaller.
A single-port memory cannot render when it is reading pixel data for display, unless the port supports high speed, bidirectional signaling. If one compares single- and dual-ported memories where both render port bandwidths are identical, then the dual-ported memory will have both a higher fill rate and a higher cost. If one compares single- and dual-ported memories where the bandwidth of the single port is equal to the sum of the dual-port bandwidths, then the single-ported memory's fill rate is likely to be higher than the dual-ported memory, because the single-ported memory is more efficient. Thus, to the extent that bandwidth limitations are presently being relaxed due to the emergence high bandwidth input/output (I/O) capacities, a single-ported memory architecture promises more efficient frame buffer performance.
Dual-ported memories allow a smoother flow of pixels to the frame buffer. A single-ported memory is unavailable for rendering on a periodic basis while it reads bursts of display data. The rendering controller requires a larger pixel FIFO to smooth out pixel flow when interfacing with a single-ported memory. In a lower cost system, the renderer may be idle during such display bursts.
A single-ported memory is cheaper due to savings in die area, pins, packaging, testing, and power consumption. A single-ported memory has a significantly lower cost per bit of storage than a dual-ported memory of the same size. If the cost per bit is lower, storing non-displayable data in the frame buffer is easier to justify.
A dual-ported memory has a fixed display bandwidth. If the required display bandwidth is lower, then bandwidth is being wasted. If the required display bandwidth is higher, then the memory is not suited to the display requirements. A single-ported memory has the flexibility to trade off render bandwidth and display bandwidth. In a pinch, a single-ported memory can actually provide much higher display bandwidth.
The dedicated display port of a dual-ported memory is not used during horizontal and vertical blanking intervals, which means the display port is idle approximately twenty percent of the time.
A dual-ported memory dictates a fixed mapping of pixels and blocks to the screen. A single-ported memory can map pixel and blocks to the screen with much greater flexibility.
A dual-ported frame buffer memory only makes sense if the render and display ports are connected to different chips. If both ports are connected to the same chip, then a single-port memory, with equivalent bandwidth, would be more efficient for the reasons stated above.
A single-ported memory enables the building of lower cost systems, because the cost per bit of frame buffer storage is cheaper, and because the rendering and display chips can be merged into a single device.
Thus, a single-ported memory enables one to design lower cost, low-end systems than could be designed with

Affiliated with

Also associated with

Rating

Methods for semiconductor systems for graphics processing does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Methods for semiconductor systems for graphics processing, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Methods for semiconductor systems for graphics processing will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-3080184