Intel reveals Xe2 "Battlemage" GPU architecture initially launching with Lunar Lake CPUs in low-power form

Intel released more details about its upcoming Xe2 GPU architecture codenamed “Battlemage” during the Intel Tech Tour even preceding the Computex keynote. While previous rumors regarding possible delays or even outright cancellation for Intel’s next gen GPU architecture now seem unfounded, it looks like Battlemage will not launch in GPU and iGPU forms at the same time. Intel plans to first introduce the low-power version with the Lunar Lake processors, then release desktop and HPC versions at an ulterior time. As such, Intel only showed estimated performance gains for the iGPU in the Lunar Lake chips.

With the Xe2 Battlemage GPU architecture, Intel is focusing on providing higher utilization, improved work distribution and reduced software driver overhead. When it comes to performance boosts over Xe Alchemist, intel promises 12.5X improved draw calls, 7X better compute dispatch, 4.1X mesh shader dispatch, just to mention the biggest gains. Still, aspects that make a bigger impact on the overall performance like vertex processing, tessellation, ray triangles and trace rays are improved by at most 2X.

Each Xe2 core features 8X 512-bit Vector Engines, 8X 2048-bit XMX Engines, 64b atomic ops support and 192 KB of shared L1 cache / SLM. Similar to Xe Alchemist, Xe2 Battlemage is highly scalable and will feature a certain number of cores depending on the product stack. The Lunar Lake iGPUs, for instance, will feature up to 8 Xe2 cores with 64 Vector Engines, 2 Geometry Pipelines, 8 Samplers, 4 Pixel Backends, 8 ray-tracing units and 8 MB L2 cache.

Intel upgraded all vector engines with SIMD16 native ALUs that also support SIMD32 ops, while the Xe Matrix Extensions support INT2, INT4, INT8. FP16 and BF16 operations. Processing speeds for FP16 operations have been boosted to 2048 OPS/clock and INT8 speeds can reach 4096 OPS/clock.

To reduce latency and improve hardware / software handshake, Intel is introducing render slices as the fundamental blocks for the rendering engine. These slices are scalable as well and can be stacked and optimized depending on the GPU type. Additionally, the Ray Tracing units have also been improved to include 3 Traversal pipelines, 18 box intersections and 2 triangle intersections.

Intel expects the Xe2 iGPUs on the Lunar Lake processors to be 50% faster than the Xe Alchemist models on Meteor Lake. Another improvement comes with the new Display Engine that supports eDisplayPort 1.5, which will increase power efficiency and improve adaptive sync capabilities with reduced judder on laptop displays. Furthermore, Intel is updating the Media engine with a dedicated 8 MB cache that should increase power efficiency for encode workloads, plus it introduces support for the new VVC (Versatile Video Coding) standard, allowing AV1 quality with 10% file size reduction.