At Architecture Day 2020, Intel’s chief architect Raja Koduri presented the roadmap for future dedicated and integrated graphics cards with Intel Xe architecture and further details. Let’s discover them.
Intel Xe Vs. NVIDIA and AMD: Xe-LP, Xe-HPG, Xe-HP, and Xe-HPC
Intel’s plan is to make 3 microarchitectures from the Xe architecture: Xe-LP, Xe-HP and Xe-HPC will have to place in order of performance, also the Xe-HPG, which do not appear in the slide above.
More specifically, Xe-LP will use a 10nm+ (“SuperFin”) manufacturing node and will be designed for high performance integrated graphics, as we will see in the 11th generation of laptop CPUs. We will also find them in dedicated graphics with entry-level performance such as Intel SG1 (for servers) or Intel DG1, which have already been sent to developers to familiarize them with the architecture.
Intel Xe-HPG, which will be used in its first high-performance dedicated graphics for games, will use an external manufacturing node, meaning it will be manufactured by another company.
It is expected that the Intel Xe HPG will be equipped with GDDR6 instead of HBM memory, or at least it is logical given the difficulties that HBM has caused AMD in the past on issues such as cost or availability. However, Xe-HP and Xe-HPC would use HBM2 after filtering the Xe-HP.
As expected, HPG will have dedicated real-time ray tracing hardware, in line with AMD’s efforts to implement this technology, which was originally introduced by NVIDIA with its GeForce RTXs at Big Navi.
In addition, Xe-HP will use a 10nm++ node (“Enhanced Superfin”) and will be assembled with EMIB, meaning it will not be monolithic chips but will be physically scalable by connecting new components together, as is possible with architectures such as AMD Zen. Basically they will offer chips with 2 times more surface area ( 2 tiles ) and 4 times more surface area ( 4 tiles ) than normal ( 1 tile ).
So far, we did not know what would be the target market for Xe-HP, but it seems that they will be focused on professional multimedia streaming, obviously, we are not talking about what someone can do at home, but about intensive tasks that are performed in corporate data centers. For example, they showed how an Intel Xe-HP “1 tile” could transcode 10 4K HEVC streams.
Finally, Intel Xe-HPC “Ponte Vecchio”, intended for high-performance computing, will have Foveros and CO-EMIB (more complex versions of EMIB) packages and will combine 4 regions: the base, produced in 10nm+, the computational base, produced in the new generation enterprise processes and external processes, the “Rambo Cache” (which connects different caches), produced in 10nm++, and the input/output cache, produced in an external process.
There are still a few months left to see the Intel Xe HPG in action, certainly the one that interests us all the most, as the graphics card market needs a boost and the competition needs to be revived. Even if these graphics don’t get there, we’ll be entertained with the knowledge of what the Intel Xe architecture is capable of when it comes to integrated graphics and whether it will take the lead over AMD’s Vega.