While much of the focus for Skylake has been on the CPU side, Intel has also invested heavily in improving the iGPU side of things as well. As part of IDF 2015, Intel shared some details about what makes their iGPUs tick. Intel has long been improving their iGPUs and eroding the budget space for AMD and Nvidia.
With their Gen9/Skylake graphics, Intel largely makes iterative improvements. API support is one noticeable area of improvement with support improved to DX12/11.3, OpenCL and OpenGL 4.4. The architecture largely remains the same with the 3 domains of unslice, slice and subslice. For the unslice which is pretty much the GPU control unit, Intel introduced a number of improvements, most notably improved bandwidth to memory and better performance.
The biggest changes can be found in the slice, where all the Execution Units (EU) can be found. Each Alice is made up of 24 EUs which is the smallest iGPU Intel will ship in most cases. Here, Intel improved on MSAA performance for 2x, 4x and 8x while also adding 16x. While most iGPU gamers probably want higher fps rather than fewer jaggies, for the top end GT4 unit with eDRAM, 16x may be an option for say 720p.
Another major feature is the inclusion of lossless colour compression which is critical for the bandwidth starved iGPU. By compressing at up to a 2:1 ratio, 3-11% gains can be found in games as it increases the practical memory bandwidth for loading textures. Since the iGPUs have relatively low bandwidth and fast cache memory, moving textures from system ram quickly is critical. Both AMD and Nvidia have implemented lossless colour compression for their GPUs and it’s what allows the R9 380 and GTX 960 to have narrow 256 and 128-bit memory bus respectively. The lossless colour compression ties in with the improved ring, L3 and IMC bandwidth and latency. Other changes are an increase to the top GT4 SKU to have 72 EUs (3 slices), increased L3 cache available to the iGPU and improved pixel fill rate. This ties into the subslice which can now output up to 12 texels/CLK.
On the video front, Intel has improved their Codec support for encoding and decoding. HEVC (High-Efficiency Video Coding) 8bit, also known as H.265, encode and decode is supported and both VP8 and MJPEG decode added. HEVC 10bit is supported for decode but will require the use of the GPU and not dedicated hardware units. These additions should make streaming/recording HEVC 8bit on the iGPU much more palatable as you’re going to be losing a lot fewer frames compared to a software solution.
With Intel making such huge strides in the iGPU space, both AMD and Nvidia must be worried as Intel already dominates iGPU marketshare. AMD should probably update their APUs soon as well as they don’t yet support any form of large cache like HBM and Intel has caught up with important features like colour compression, API support and accelerated encoding/decoding. 2016 will be very interesting as AMD may finally catch up on the CPU front with Zen, HBM may be used as a cache and the GPU side continues to grow.