At the Hot Chips conference earlier this week, Microsoft showed off several slides detailing the SoC used in its upcoming Xbox One gaming console.
The Xbox One uses a System on a Chip (SoC) designed by AMD’s Semi-Custom Business Unit. The processor features eight “Jaguar” AMD CPU cores, an AMD GCN (Graphics Core Next) based GPU with 768 shader cores, an audio co-processor, and 32MB of on-chip eSRAM.
The SoC, measuring 363mm^2 is manufactured on TSMC’s 28nm HPM process. The chip can interface with 8GB of DDR3 main memory with bandwidth of 68.3 GB/s or utilize the on-chip SRAM which has bandwidth of 102GB/s. The embedded SRAM is in addition to the smaller L1 and L2 caches. The slides indicate that the GPU and CPU can at least access the SRAM, though it still remains frustratingly unknown if the SoC supports anything like AMD’s hUMA technology which would allow the CPU and GPU to both read and write to the same memory address spaces without having to copy data between CPU and GPU-accessible memory space. It may be that the CPU and GPU can use the SRAM, but the same memory spaces can not be shared, though that may be the pessimist in me talking. On the other hand, there could be something more, but it’s impossible to say from the block diagram spotted by Semi-Accurate at the Microsoft presentation.
With that said, the slides do reveal a few interesting figures about the SoC that were not known previously. The Xbox One SoC has 47MB of on-chip memory including 32MB eSRAM used by the CPU and GPU and 64KB of SRAM used by the audio co-processor. The chip’s GPU is rated for Microsoft’s DirectX 11.1 and above graphics API. Further, Microsoft rates the GPU at 1.31 TFLOPS, 41 Gigatexels-per-second, and 13.6 Gigapixels-per-second. Additionally, the GCN-based GPU is able to hardware-encode multi-stream H.264 AVC MVC video and hardware decode multiple formats, including H.264 MVC. The hardware encoder is likely being used for the console’s game capture functionality.
The audio processors in the Xbox One SoC use two 128-bit SIMD floating point vector cores rated at 15.4 GFLOPS and “specialized hardware engines” and “signal processing optimized vector and scalar cores.”
The final interesting specification I got from the slides was that the SoC is able to go into a low power state that is as low as 2.5% of the chip’s full power using power islands and clock gating techniques.
You can find all of the geeky details in these slides over at SemiAccurate.