Xbox 360 rides two beefy chips

Most of the gaming horsepower for the latest Xbox resides in two rather massive and complex chips, one a multicore processor from IBM Corp. (the XCPU X02046-003) and the other a graphics processor engine based on technology from ATI Technologies Inc. (GPU X02056-010).

The IBM CPU contains three individual processor cores, each with 32 kbytes of Level 1 cache memory. The three cores then communicate by way of a crossbar switch fabric to a shared 1-Mbyte Level 2 cache, with an external interface to the front-side bus used for off-chip communication.

Purportedly integrating 165 million transistors and based on a 90-nanometer process technology, the IBM chip is mounted flip-chip style to a 900-I/O BGA package. The sizable (and probably pricey) fine-fin heat sink and high proportion of package I/O dedicated to power and ground speak to the power-hungry nature of the CPU. Similarly, the forest of decoupling capacitors on the underside of the CPU mounting site indicates a speedy on- and off-chip switching environment where noise management is essential.

Most of the CPU's off-chip signaling is related to connection with the ATI-based graphics device, the GPU. A total of 39 differential-pair signals are visible on the top and bottom surfaces between the CPU and GPU, with more bus lines likely internal to the board layers. While the bus width has not been determined, a 10.8-Gbyte/second transfer rate in each direction (21.6 Gbytes/s total) has been widely disclosed.

The GPU is the second barrel in the silicon shotgun of the Xbox 360. Like the central processing unit, the GPU is a massively complex piece of silicon (235 million transistors, 90 nm) supported by an off-chip DRAM residing on the package topside with the GPU proper. With all this high-end processing, power concerns must again be addressed, this time with a somewhat lower-performance extrusion heat sink whose airflow is delivered through a plenum shared with the CPU heat sink assembly.

While hard to believe (and we couldn't fathom how to make the numbers work out), various observers have claimed an eye-popping 256-Gbyte/s transfer rate between GPU and in-package memory. Based on an examination of the memory die in depackaged form, it is difficult to identify more than a 256-bit bus (32 bytes). At this assumed bus width, the 256-Gbyte/s transfer rate would infer eight reads or writes per nanosecond for a 125-ps memory cycle time, an extraordinary figure if correct. Whatever the miscalculation or misinformation may be, it seems clear that there is plenty of local memory activity within the GPU complex before going off-chip.