by Anton Shilov
02/05/2013 | 10:23 PM
For the first time, reasonably detailed specifications of what is claimed to be the graphics processing unit for Microsoft Corp.’s next-generation Xbox have emerged on the web. The specs published by VGleaks web-site confirm that the GPU is based on AMD’s current graphics core next (GCN) architecture with some additional hardware and functionality. Unfortunately, the data does not reveal any special capabilities of the GPU.
As expected, the new Xbox GPU is heavily based on GCN and therefore has similar topology. The chip has 12 SCs (SIMD clusters) with 4 SIMDs (that feature 16 ALUs/stream processors as well as 256 vector general purpose registers (VGPRs), and 512 scalar general purpose registers (SGPRs).) and its own L1 cache, LSM (local shared memory), and scheduler. The SIMD instruction set is extensive, and supports 32-bit and 64-bit integer and float data types. It is believed that the capabilities of SIMDs and ALUs are similar to those of GCN. The chip also features 48 texture modules and 4 render back ends. With 768 stream processors on chip, the GPU should provide peak computer performance similar to AMD Radeon HD 7770 (Pitcairn).
According to the leaked data, the GPU of the code-named Durango system will fully support modern graphics capabilities, including full-scene antialiasing with 2x, 4x and 8x levels. While currently the talks indicate that both traditional multi-sample AA as well as compressed AA (coverage sample AA, surface sample AA) methods are supported, it is unlikely that bandwidth-hungry MSAA will actually be used by game developers.
The new GPU will also be equipped with 32MB of embedded SRAM (ESRAM) with 102.4Gb/s throughput and low-latency, which should speed up numerous memory-intensive operations as peak bandwidth of main memory is 68GB/s amid relatively high latency. The chip has special Move units that will move data between main and embedded memory.
Quite surprisingly, but the graphics solution for the Xbox Next has rather powerful general-purpose compute capabilities as well as robust caching system. In addition to dedicated 64KB LSM, 16KB L1 per SC, the chip will feature 512KB of L2 split into four domains.
Given the fact that the leaked data about the Durango GPU is pretty detailed, there is a high probability that it is indeed correct to a certain degree.
Microsoft did not comment on the news-story.