Intel details its Xe HPG “Alchemist” GPU architecture – Intel on TSMC 6nm

  

Intel’s Xe HPG architecture is designed to scale across to the high-end of the GPU market. Intel has already confirmed that they will be releasing a “family” of Xe HPG “Alchemist” SOCs under their new ARC brand, confirming that Intel has designed multiple Xe HPG chips. 

Intel’s Xe cores will be built with traditional vector engines, matrix engines, which can accelerate matrix operations (which are commonly used for AI), and dedicated Ray Tracing Units. Each Xe core compute block will feature 16 vector engines, sixteen matrix engines, a ray-tracing core, and a sampler. Four of these Xe cores are used to create an Intel XMX and share some resources and caches.  

  

Over the past few years, Intel has been working to improve their graphics drivers, paving the way to the launch of Xe HPG. This effort has reduced the CPU overhead of Intel’s graphics drivers, increased game performance and has dramatically increased the stability of the company’s integrated graphics solutions. 

Intel’s Xe HPG chips will support DirectX 12 Ultimate and can be used with the latest APIs and game engines like Vulkan, Unreal Engine and Unity. 

  

Based on the render below, Intel’s top Xe HPG “Alchemist” graphics card will feature eight Xe slices, each of which features four Xe cores that contain sixteen vector/matrix engines. (8 x 4 x 16 = 512) Intel’s high-end Xe HPG “Alchemist” GPU features 512 vector units, which is significantly more than any of Intel’s existing Xe graphics solutions. 

 

Intel has confirmed that they will be using TSMC to create their upcoming Xe HPG graphics chips, opting to use TSMC’s 6nm lithography node. With Xe HPG, Intel has promised a 1.5X increase in frequency over today’s Xe LP GPUs and a 50% increase in power efficiency. This will make Intel’s Xe HPG gaming architecture their most performant and most efficient to date. 

Using TSMC’s 6nm lithography node will mean that Intel’s GPU production will be limited by the manufacturing capacity of an external foundry, Intel decided not to use their own internal Intel 7 node to create their Xe HPG architecture, which is not a good sign for Intel’s foundry division. Why aren’t Intel using their own fabs? Are they not good enough? These are the questions that we are left asking, and they are not good questions to have. 

Only time will tell if Intel had gotten enough production capacity from TSMC to become a serious competitor to Nvidia or AMD within the GPU market. I guess we will see in early 2022. 

  

  
You can join the discussion on Intel’s Xe HPG “Alchemist” GPU architecture on the OC3D Forums.