Intel details Arc Xe-HPG DG2 Alchemist architecture, each Alchemist render slice made of four Xe-cores and an equal number of ray tracing cores
Intel has detailed more information about Xe Arc during its 2021 Architecture Day today. The first generation Xe-HPG Alchemist discrete GPUs are fabbed on the TSMC 6 nm process.
At the heart of the Arc Xe-HPG DG2 GPU is what Intel calls as the Xe-core. An Xe-core is made of 16 vector engines (256-bit per engine) and 16 matrix engines (XMX, 1024-bit per engine). Four such Xe-cores each with their own L1 cache, sampler, and ray tracing units form part of a render slice apart from other components such as the geometry pipeline, rasterization pipeline, hierarchial depth buffers (HiZ), and pixel backends. The Therefore, the number of RT cores in Xe-HPG Alchemist is equal to the number of Xe-cores. The ray tracing units are capable of performing ray traversal, bounding box intersection, and triangle intersection functions while the geometry, rasterizers, samplers, and pixel backend offer fixed-function pipelines for graphics APIs such as DirectX 12 Ultimate.
Each render slice is also accompanied by a memory fabric with a large L2 cache and a global dispatch unit. The global dispatch and memory fabric with L2 cache scales according to the number of render slices. So, an Xe-HPG Alchemist GPU with eight render slices will feature 32 Xe-cores, 32 RT cores, 32 samplers, 32 L1 caches, one global dispatch, and one memory fabric with a common L2 cache. Intel said that Xe-HPG offers 1.5x better frequencies for the same voltage and 1.5x the performance per Watt compared to Xe-LP.
Intel once again confirmed the codenames of upcoming Xe-HPG architectures. While Battlemage and Celestial will be the successive iterations of the current design with obvious performance and efficiency improvements, Druid will see the introduction of a new Xe-HPG design.
Check out the slides from Intel's presentation below
Source(s)
Intel Architecture Day 2021 Keynote
