Intel GPU Instruction Set Architecture Gets First Massive Overhaul Since 2004 For Intel Xe GPU
Intel appears to be working full steam on its Xe GPU ambitions and getting ready to launch their first commercial product in 2020. According to a merge request posted on Gitlab, Gen12 (aka Tigerlake) will contain one of the most significant reworks to the Intel EU ISA since the original i965 chip released almost a decade ago. The foundation for Intel’s GPU efforts are clearly being laid and in due time, two will become three.
Intel’s EU (GPU) ISA getting first significant rework in 15 years as it prepares for Xe architecture
The Intel EU ISA is the instruction set for the execution unit (one block of an Intel architecture) found inside existing Intel GPUs and hasn’t had a major update since 2004 when the Intel i965 was introduced. This is because Intel focused primarily on iterative performance improvements and never had the need to focus on ambitious plans like Raja’s 2x performance promise. With Intel Xe now in the equation, however, the time has come to completely rework the EU ISA from the groundwork, laying the foundation for the Xe evolutionary path.
This merge request implements support for the Gen12 (Tigerlake) ISA in the i965/Iris compiler back-end.
Gen12 is planned to include one of the most in-depth reworks of the Intel EU ISA since the original i965. The encoding of almost every instruction field, hardware opcode and register type needs to be updated in this merge request. Gitlab
Intel will be introducing the brand new Xe Graphics Engine with Tiger Lake (Gen12) in 2020 along with brand new I/O technology and a new CPU Core architecture. Intel’s offerings for the year of 2019 have been fairly straightforward and didn’t really contain something that would get enthusiasts excited but it looks like 2020 might be the year to change all that. Raja Koduri of Intel has promised doubling performance with every generation for the foreseeable future and the journey begins with Tiger Lake.
Tiger Lake is very important for the company as a testing platform as well because the lessons they learn here will go on to be incorporated in their full-fledged discrete GPUs. Intel has already promised to include raytracing at the hardware level and since the company plans to shift everything to 7nm by 2021, its entirely possible we will see powerful Xe GPUs start to ship on the 7nm node from the start (2020 is for the first commercial release, which might or might not be gamer-centric).
Now the more skeptical of you might be thinking 10nm was delayed, so 7nm will be delayed as well, and while that is a logical conclusion its actually not accurate. The problem with 10nm was the use of non-EUV lithography and quad patterning. Since 7nm is based on EUV and will almost certainly not be quad patterned, the delay of 10nm should, in theory, have 0 effect on the 7nm timeline. Unfortunately, however, this does mean that we still have more than a year to go before things start to become fun again.