Intel Broadwell Compute Gen8 GPU Architecture

Attention OpenCL programmers! Intel has released a detailed description of the Gen8 Broadwell GPU compute architecture, “The Compute Architecture of Intel® Processor Graphics Gen8“. Broadwell is a 14nm die shrink of Intel’s microarchitecture incorporating significant reworking of the Intel HD 5000-series (Iris) Gen 7.5 GPU found in Haswell including: (1)  throughput for 32-bit integer computation has doubled, (2)  native 16-bit floating-point support, (3)  the write bandwidth from GTI has doubled on some devices, (4) coherent shared virtual memory between CPU cores and Intel® Processor Graphics enabling seamless sharing of pointer rich data structures, (5)  increased L3 data cache capacity and improved local bandwidth between EUs  (Execution Units) and L3 data cache.

Another potential product design that instantiates the compute architecture of Intel® Processor Graphics Gen8. This design is composed of two slices, of three subslices each for a total of 48 EUs. (From The Compute Architecture of Intel® Processor Graphics Gen8)