Intel Shows Off New Gen11 Graphics, Teases Xe Discrete GPU
Intel’s Architecture Day 2018 this past Tuesday wasn’t just a CPU show. The graphics market is poised to be a significant component of Intel’s strategy going forward, and the company’s Gen 11 solution looks like it’ll be a potent improvement over Skylake. These improvements are long overdue.
For most of the past twenty years, the phrase “Intel graphics” was a contradiction in terms if you cared about gaming. Starting in 2011, with Sandy Bridge, that began to change. There was a period of roughly five years where Intel’s own solutions were improving at a solid pace. From 2011-2015, IGP performance improved in real terms, meaning Intel’s GPUs got faster more quickly than games demanded additional GPU resources. There were still only a relative handful of titles that could be coaxed into running well, but the situation was improving by the year. And then it stopped. Neither Kaby nor Coffee Lake contained any additional 3D optimizations. After 3.5 years in the proverbial wilderness, Intel wants to change that.
The new Gen 11 GPU is Intel’s first TFLOP-class GPU hardware. It implements a tile-based renderer, presumably to take advantage of tiled rendering’s lower power consumption and increased efficiency. The GPU will contain 24-64 execution units and it packs a 4x larger L3 cache.
Other major features include a tile-based renderer that Intel will be able to enable or disable on a per render-pass basis. Memory compression has also improved, with a claimed best-case performance jump of 10 percent and a geometric mean average of four percent. The new Sunny Cove GT2 configuration intended for desktop will pack 64 EUs, compared with 24 EUs in Skylake’s GT2. That alone should be worth some significant performance improvements, subject to memory bandwidth constraints or other bottlenecks within the core.
Intel has also implemented a capability they call Coarse Rate Shading, similar to Nvidia’s Variable Pixel Shading. Coarse Rate Shading reduces the total amount of per-pixel shading work, based on either on the distance between the camera and the area in question or as a function of how close the area in question is to the center of the screen. By stretching the shading for a single pixel over a 2×2 block of pixels, Intel was able to improve performance by 10-20 percent in the demo we saw, depending on where the camera was on-screen. A 4×4 block size improved performance by even larger margins, though the impact on image quality was more visible at the distances players would actually play at.
HEVC decode units have been designed from the ground up, with two decoders and one encoder and support for up to 8K HEVC decoding. Adaptive Sync and HDR will both be supported by this hardware and by the Xe discrete GPU branding that Intel teased as part of the event. Intel demonstrated its new GPU capabilities by showing off Tekken in a side-by-side demo with Gen 9 (Skylake) on one side and Gen 11 (Sunny Cove) on the other. The systems were running Medium detail 1080p with an 85 percent scaling factor, putting the actual resolution around 900p. Performance was significantly better in the Gen 11 demo, despite the early state of drivers and feature support. While it’s completely unclear how Sunny Cove will match up against AMD’s own APUs in late 2019, the jump in EU count and associated changes suggest Intel will at least compete much more effectively with AMD than it has of late.