Graphics Reference
In-Depth Information
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
IU
SPU
VPU
Core
Instruction
unit
scalar
proc. unit
ALU
scalar
proc. unit
ALU
ALU
ALU
ALU
ALU
ALU
ALU
ALU
ALU
ALU
ALU
ALU
ALU
ALU
ALU
ALU
L1$
L1$
L1$
L1$
L1$
L1$
L1$
L1$
L1$
L1$
L1$
L1$
L1$
L1$
L1$
L1$
L2$
L2$
L2$
L2$
L2$
L2$
L2$
L2$
L2$
L2$
L2$
L2$
L2$
L2$
L2$
L2$
Ring bus
Memory interface
(details undisclosed)
GPU chip
PCIx interface
Texture evaluation unit
L1$: 1st level cache
L2$: 2nd level cache
Application
DRAM memory
(details undisclosed)
Figure 38.11: Intel Larrabee GPU block diagram. Compare this with the
NVIDIA
GeForce 9800 GTX GPU block diagram in Figure 38.4.
•
Many cores:
The Larrabee and GeForce 9800 GTX block diagrams are
both dominated by multiple processing cores, as are the implementations
of the GPUs themselves. While the number of cores in the Larrabee GPU
was never announced, it is known to be at least 16, matching or exceed-
ing the GeForce 9800 GTX, and greatly exceeding the four cores of the
Core 2 Extreme QX9770 CPU. Furthermore, like the GeForce 9800 GTX
cores, the Larrabee cores are designed to maximize performance per unit
of silicon area (and therefore overall GPU performance), whereas the Core
2 Extreme QX9770 CPU cores are designed to maximize performance per
core (at the expense of overall performance).
•
Wide vectors:
The Larrabee and GeForce 9800 GTX cores both include
wide SIMD units:
n
=
8 for the GeForce 9800 GTX (virtualized to
n
=
16)
and
n
=
16 for Larrabee. Both GPU cores provide hardware support for
predication, and both provide separate address circuits for each element,
allowing data to be efficiently
gathered
into a vector, then
scattered
back
to memory. Conversely, the Core 2 Extreme QX9770 CPU implements nar-
row vectors (
n
=
4) with no support for efficient scatter/gather or predica-
tion.
•
Texture evaluation:
Both GPUs support texture evaluation with dedicated,
fixed-function units; the CPU provides no support.
Balanced against these important similarities, the Larrabee and GeForce 9800
GTX implementations also differ in ways that betray Larrabee's CPU heritage.
These include the following.
•
Specialized, fixed-function hardware:
Except for its texture evaluation
unit, Larrabee omits support for the many GeForce 9800 GTX fixed-
function units, including vertex generation, primitive generation, frag-
ment generation (rasterization), work queueing and distribution, and pixel
operations.
