Intel provides more details about its video cards that will appear in early 2022. The top model gets 512 execution units and the GPUs are made on the 6nm process of TSMC. Intel also details its Xe Super Sampling technique for upscaling.
Intel shows the construction of its GPU for Alchemist video cards. The schematic representation shows the top model with 512 eu’s. Intel itself speaks of Xe cores, which each consist of sixteen 256-bit vector engines and sixteen 1024-bit matrix engines. An Xe-core is therefore equal to sixteen eu’s.
The construction of the Alchemist GPUs consists of render slices containing four Xe cores and four ray tracing cores. The GPUs can contain up to eight of these slices, good for 32 Xe cores and a total of 512 EUs.
Eight slices containing four Xe cores each
Production outsourced to TSMC
Intel confirms that the Alchemist GPUs are produced on TSMC’s N6 process. That’s an improved version of the N7 process, which is currently being used by AMD for Ryzen processors, Radeon video cards and the socs in the latest Xbox and PlayStation consoles.
The details will follow after the unveiling of the Intel Arc name earlier this week. That will be the counterpart of Nvidia’s GeForce and AMD’s Radeon. The first Arc video cards will appear in the first quarter of 2022. Intel announced the new details during its Architecture Day.
XeSS: Xe Super Sampling, also for third parties
Intel also talks about its own alternative to Nvidia’s DLSS and AMD’s FSR. Intel’s counterpart is called XeSS, which stands for Xe Super Sampling. This is a deep learning technique that uses information from previous frames when upscaling and takes the speed of movement into account. Like DLSS, XeSS is a temporal upscaling technique.
Intel claims that games rendered in 1080p resolution and upscaled to 4k using XeSS will look just as good as a native 4k resolution. The technique works on the basis of the DP4a and XMX instructions, which are integrated in the Xe GPUs. A XeSS sdk should be released for developers this month. All tools and the SDK are open source and can also be used by other GPU manufacturers. The technique works optimally with Intel hardware with Intel’s own XMX, but can also be used by other GPU manufacturers or with older Intel GPUs with the more widely available DP4a instructions.
Intel has also given a roadmap with some data about future GPUs. For example, Intel says the first three generations will all be built on iterations of the Xe HPG architecture. The fourth generation, codenamed Druid, will use a new GPU architecture. Intel does not yet provide a timeline and it is not yet known on which nodes the future video cards will be produced. Intel may switch to its own nodes in subsequent iterations, which recently received new names and a new roadmap.
|Intel Arc Roadmap|
|GPU architecture||Xe HPG||Xe2 HPG||Xe3 HPG||Xe ‘Next’|
|Release date||Q1 2022||Nnb||Nnb||Nnb|
Intel Arc GPUs for laptops and desktops