r***@yahoo.com
2005-10-30 19:57:25 UTC
http://www.extremetech.com/article2/0,1697,1877355,00.asp
IBM Talks Up Xbox 360 Processor
Power PC Plus
The Xbox 360 CPU chip contains three processor cores, the shared, 1MB
L2 cache, the crossbar linking the CPUs and L2, plus clocks, test and
debug circuitry and I/O.
The XCPU is a three-way symmetric multiprocessor; each core is
identical. The CPU clocks at 3.2GHz, and has a maximum front side bus
speed of 21.6GHz per second (divided into the two virtual channels, as
discussed earlier.) IBM designers also built in substantial clock
gating, to try to minimize power consumption. Entire pipelines can be
shut down until they're needed, so the Xbox 360 probably won't be
sucking too much power at idle-something essential for a consumer
electronics device.
The cores are built around an enhanced version of the Power PC
architecture. The total transistor count for the die is 165M
transistors. Contrast this with the 233M transistors in the Athlon 64
X2 Toledo die, which has two cores, but one MB of L2 cache per core.
The processor is built on IBM's 90nm Silicon-on-Insulator (SOI)
process, and is a hefty 168 square millimeters-large, but again, not
quite as big as AMD's Toledo core, which is 199 square millimeters.
(Xbox 360 CPU Core Block Diagram)
Loading Image...
Each core has a two-way, 32KB instruction cache with a 128 cache line
width. The data cache is similar, and is also 32KB. The L2 cache can
feed directly into a pre-decode block, which then moves instructions
into the instruction cache.
Perhaps most interesting is the VMX section. "VMX" is IBM's moniker for
their SIMD (single-instruction, multiple data) block, akin to Intel's
SSE instructions. The standard PowerPC has 32 VMX registers, but the
designers expanded the VMX register file to 128. There's also a
standard floating point unit, but it's likely that most of the floating
point will be handled by the VMX unit.
Note that the VMX unit can handle two simultaneous threads, and IBM
added some additional instructions to facilitate gaming and streaming
media. For example, there are now Direct3D-specific pack and unpack
instructions and floating-point dot product instructions. It's all
about accelerating 3D graphics, AI and physics, after all.
The CPU itself has 16 32-bit performance counters, which tie into the
Xbox 360 developer kit tools, and enables fine grain performance
analysis of Xbox 360 application code.
While you won't be able to buy an Xbox 360 processor, the Xbox 360
console will be shipping in the fourth week of November. Brown didn't
reveal what the ultimate power consumption would be, but did indicate
that work was continuing in improving overall efficiency. As the part
is revised over time, it's likely that overall power consumption will
decline.
Although the die size seems big, it's not as big as some current PC
processor designs, and is being built on a well-understood
manufacturing process. Still, it's not going to be a low-cost part, so
we can start to understand a bit why the Xbox 360 is going to be priced
higher than the original Xbox.
Given that this is a fully symmetric, multicore design, game developers
are going to have to be smart about threading their code. It's ironic
that an industry that's been focused on writing single event loop
titles since the early days will now rapidly develop a base of
experience in writing threaded code. It will be interesting to see just
how efficiently they'll make use of the new CPU.
IBM Talks Up Xbox 360 Processor
Power PC Plus
The Xbox 360 CPU chip contains three processor cores, the shared, 1MB
L2 cache, the crossbar linking the CPUs and L2, plus clocks, test and
debug circuitry and I/O.
The XCPU is a three-way symmetric multiprocessor; each core is
identical. The CPU clocks at 3.2GHz, and has a maximum front side bus
speed of 21.6GHz per second (divided into the two virtual channels, as
discussed earlier.) IBM designers also built in substantial clock
gating, to try to minimize power consumption. Entire pipelines can be
shut down until they're needed, so the Xbox 360 probably won't be
sucking too much power at idle-something essential for a consumer
electronics device.
The cores are built around an enhanced version of the Power PC
architecture. The total transistor count for the die is 165M
transistors. Contrast this with the 233M transistors in the Athlon 64
X2 Toledo die, which has two cores, but one MB of L2 cache per core.
The processor is built on IBM's 90nm Silicon-on-Insulator (SOI)
process, and is a hefty 168 square millimeters-large, but again, not
quite as big as AMD's Toledo core, which is 199 square millimeters.
(Xbox 360 CPU Core Block Diagram)
Loading Image...
Each core has a two-way, 32KB instruction cache with a 128 cache line
width. The data cache is similar, and is also 32KB. The L2 cache can
feed directly into a pre-decode block, which then moves instructions
into the instruction cache.
Perhaps most interesting is the VMX section. "VMX" is IBM's moniker for
their SIMD (single-instruction, multiple data) block, akin to Intel's
SSE instructions. The standard PowerPC has 32 VMX registers, but the
designers expanded the VMX register file to 128. There's also a
standard floating point unit, but it's likely that most of the floating
point will be handled by the VMX unit.
Note that the VMX unit can handle two simultaneous threads, and IBM
added some additional instructions to facilitate gaming and streaming
media. For example, there are now Direct3D-specific pack and unpack
instructions and floating-point dot product instructions. It's all
about accelerating 3D graphics, AI and physics, after all.
The CPU itself has 16 32-bit performance counters, which tie into the
Xbox 360 developer kit tools, and enables fine grain performance
analysis of Xbox 360 application code.
While you won't be able to buy an Xbox 360 processor, the Xbox 360
console will be shipping in the fourth week of November. Brown didn't
reveal what the ultimate power consumption would be, but did indicate
that work was continuing in improving overall efficiency. As the part
is revised over time, it's likely that overall power consumption will
decline.
Although the die size seems big, it's not as big as some current PC
processor designs, and is being built on a well-understood
manufacturing process. Still, it's not going to be a low-cost part, so
we can start to understand a bit why the Xbox 360 is going to be priced
higher than the original Xbox.
Given that this is a fully symmetric, multicore design, game developers
are going to have to be smart about threading their code. It's ironic
that an industry that's been focused on writing single event loop
titles since the early days will now rapidly develop a base of
experience in writing threaded code. It will be interesting to see just
how efficiently they'll make use of the new CPU.