Intel’s Gen12 Xe graphics architecture is finally set to debut in late 2020 with the Tiger Lake-U CPUs. The Gen12 graphics featured on TGL will include up to 96EUs, 50% more than Ice Lake. Expect the to-end graphics to be on par with the GTX 1050 in certain titles. However, in this post, I’d like to discuss something else. The half-precision or FP16 performance of the Gen12 graphics processors. I’ve been checking all the leaked results and asking some people in the industry and turns out that one of the highlights of the Xe graphics architecture will be its half-precision performance.
Here you’ve got the Tiger Lake-U flagship with 96 Execution Units. Check the half-precision and single-precision compute scores. The former is notably higher than the latter. This is unusual because in SiSoft, all contemporary GPUs, both AMD and NVIDIA have a half-precision score slightly lower than their single-precision scores. Furthermore, the Gen12 Xe graphics parts have FP16 capabilities nearly on par with the NVIDIA GTX 970 and higher than the GTX 1050 Ti.
Keep in mind that this is only the case with the FP16 scores, the FP32 compute performance is notably higher on both NVIDIA GPUs.
So, what does this mean? The main takeaway is that Intel’s Xe architecture will work really well with mixed precision, as well as workloads that leverage half-precision compute. These days, it’s mainly machine learning and neural-network algorithms that utilize half-precision compute on a wide scale. This could mean that Intel is preparing its graphics architecture to excel in professional and Data Center environments first, and gaming seems to be a secondary priority. This would make sense, as in the latter both NVIDIA and AMD have a major hardware and driver advantage. It’ll take years before Intel’s Xe architecture is reasonably optimized for the bulk of mainstream games.