Graphics Reference
In-Depth Information
memory storage and bandwidth, and its incompatibility with multi-sample
anti-aliasing.
The advantages of binned rendering are compelling, yet no current PC-class
GPU implements it.
18
The fundamental reason is that binned rendering differs
too much from the pipeline Direct3D and OpenGL architectures—we say that the
abstraction distance
is too large. Typically, excess abstraction distance results in
products with confounding performance characteristics (operations expected to be
fast are slow; those expected to be slow are fast) or subtle deviations from specified
operation. Practical problems that are encountered include the following.
•
Excess latency:
Previous binned rendering systems, such as the Pix-
elPlanes 5 system developed at the University of North Carolina, Chapel
Hill, have added a full frame time of latency.
•
Poor multipass operation:
Direct3D and OpenGL encourage advanced
multipass rendering techniques that, in a binned implementation, require
multiple two-pass operations per final frame. An example is rendering
reflection from a curved surface by 1) rendering the scene that will be visi-
ble in the reflection, 2) loading this image as a texture, and 3) rendering the
curved surface with the appropriately warped texture image. Some binned
rendering systems failed to support such operations; others supported them,
but performed poorly.
•
Unbounded memory requirements:
While binned rendering bounds
pixel storage
to that required for a single tile, the memory required by
the bins themselves grows with scene complexity. Neither OpenGL nor
Direct3D has scene-complexity limits, so a fully confirming implementa-
tion requires infinite memory (an obvious impossibility) or must introduce
complexity to handle cases for which finite bin storage is inadequate.
These complications have been sufficient to keep binned rendering out of
mainstream PC GPUs. But recent implementation trends, in particular the use of a
time-shared single compute engine to implement all pipeline shading stages, may
overcome some of the difficulties.
In 2008, Intel published a technical paper [SCS
+
08] describing a forthcoming
GPU, code-named Larrabee. While this product never shipped, for reasons we
will discuss shortly, it was a serious attempt to combine ideas from Intel CPUs
and competitive GPUs into a single compelling product. In this short section we
will analyze this hybrid GPU, first by comparing its implementation and architec-
ture with those of the NVIDIA GeForce 9800 GTX GPU and the Intel Core 2
Extreme QX9770 CPU, and then by considering the resultant strengths and weak-
nesses. We begin with implementation, referring to the Larrabee block diagram in
Figure 38.11, which is drawn to be consistent (where possible) with the NVIDIA
GeForce 9800 GTX block diagram in Figure 38.4.
In several key respects the Larrabee implementation more closely resembles
the implementation of the NVIDIA GeForce 9800 GTX
GPU
than it does that of
the Intel Core 2 Extreme QX9770
CPU
.
18. Inexpensive Intel GPUs have implemented binned rendering in the past, but none do
now.
