Days after revealing Arc, Intel’s own GPU graphics brand for consumers, particularly aimed for gamers and consumer use and now we’re getting a larger picture at the details behind Intel Arc and the first-generation to be released from the family, the GPU codenamed Alchemist. From Intel’s initial demo, aside from modern performance expectations, people were quite surprised when Intel showed off their own AI superampling technology: that tech was recently announced be called Xe Super Sampling or XeSS (pronounced X-E-S-S) and competes is expected to compete more with DLSS although FSR is still in the conversation for upscaling tech.
Just like AMD and NVIDIA, Intel has their own parallel processor designed to be scalable for their Xe graphics dubbed Xe-Core. An Intel Xe-core for the Alchemist architecture contains 16 256-bit vector engines and 1024-bit matrix engines together with basic load/store hardware along with an L1 cache. The Xe-core forms a Render Slice, with a total of 4 Xe-cores, 4 Raytracing units and multiple fixed-function hardware to complete the slice including rasterization pipeline, geometry pipeline, samplers and the pixel backends. The raytracing unit comprises of fixed-function hardware that handle raytracing related duties.
The Render Slice is joined by the global dispatch processor and the GPU memory fabric and is joined by an L2 cache. Intel Alchemist GPUs features 8 render slices, memory subsystem and a global dispatch with models in the family being made by disabling render slices. Adding everything up with 16 execution units per Xe-cores, 4 Xe-cores pero slice, with a total of 8 render slices, there is 512 total EU and a total of 4096 programmable shades in the largest Intel Alchemist GPU.
The Alchemist GPU will be made on TSMC’s 6nm node and Intel is claiming 50% performance-per-watt improvements over Xe LP made in Intel’s inhouse 10nm node using their SuperFin tech. This includes Iris Xe Max. It is expected that Alchemist will have a larger power rating which will allow it to run at higher clock speeds.
To power their AI functinalities, Alchemist GPUs 1024-bit Matrix backend, all 512 of it, allows the GPU to perform with excellent AI computations allowing their XeSS technology to take advantage of its computational ability alongside other deep-learning and real-time rendering techniques.
Intel states the Alchemist GPUs will make their debut in Q1 2021 with future releases already mapped ahead codenamed Battlemage, Catalyst and Druid which are the marketing codenames of their Xe2, Xe3 and XeNext GPUs that will arrive in the following years.
As of now, it is not officially stated if Intel will go with a partner release nor will Intel be releasing them in annual cycles to catch-up with its rivals.
