Intel has released another episode of its Talking Tech podcast, this time featuring Tom Peterson (TAP) the Intel Fellow heading up the graphics card department. In there, they talk in-depth about the next generation of Intel’s next-generation GPU architecture Xe2. Giving another breath into the possibility of there being a strong competitor in the market, as it aims to improve on its first Intel Arc GPUs released back in 2022, including the Intel Arc A770 we reviewed.
Unfortunately, the first generation had its problems primarily via driver issues and just not holding up to the competition, though this has improved over time. This leads to the promises given in the podcast that Xe2 can bring that now and become more competitive as a third choice for your gaming PC.
As the core technology that is looking to debut in the Lunar Lake and Battlemage releases, it will bring the second attempt in Intel’s graphics department. Considering TAP talked of the heavy investments, the big improvements better live up to the hype as Intel’s financial woes might not let it carry on with these endeavors if they don’t succeed.
Xe2 GPU architecture, Source Intel
Xe2 performance improvements, Source Intel
Xe2 core composition, Source Intel
Xe2 architecture scalability, Source Intel
Intel Xe2 architecture improvements
Primarily, the Xe2 generation was started as soon as the first Alchemist hardware designs were done. This gave the team a lot of time to configure this new design, creating a scalable architecture that can be easily used as both onboard and discrete graphics. Thanks to the cores now containing both the vector and matrix units with the cache for each. Creating a sort of Lego brick that can be combined with multiple of these to create a bigger die.
Along with the die, the discrete cards have the benefit of being able to host memory modules not shared with the system. Allowing it to use bigger confirmation with a faster memory setup, leading to better performance but also an improved efficiency with everything so close together.
TAP also explained where the effective performance increases are coming from. With the XMX instructions now in the stack, it allows vector arithmetic to be faster thanks to this new addition. Primarily helping the likes of AI processing but also general workloads. Then the Vector engine has been moved to SIMD16 (single instruction multiple data), the fundamental organization process for the computational pipeline.
Basically, the instructions behind organizing the workloads, do help with game compatibility as some have baked-in requirements for the higher version of the system. This allows it to get higher frame rates in older games, targeting the fallbacks of the Alchemist cards that struggled with older than DX12 games. It also now supports executing indirectly which is important in games, and now the GPUs can do so without emulation, leading to the 7-12x performance it boasts in its graphs.
All in all, we can expect to see a whole load of improvements from the cards. As Tom also mentions, now that they have the hardware they can test the Battlemage architecture in games and programs they couldn’t do the first time around with their initial Alchemist products. So hopefully it will lead to a working product out of the get-go instead of aging like a wine it had to utilize for Xe1.
