"Battlemage" Architecture Preview

Both the Arc B580 and B570 are based on the "BMG-G21" a new monolithic silicon built on the TSMC 5 nm EUV process node. The silicon has a die-area of 272 mm², and a transistor count of 19.6 billion. Intel already had a very mature hardware-accelerated real-time ray tracing stack going in with "Alchemist," and this only evolves more with Xe2 "Battlemage." The new graphics architecture relies on higher utilization of hardware resources, improved work distribution among the various logic components of the GPU, and a generational reduction in software overhead (requiring CPU utilization to do things).


The new Xe2 core boasts of a generational IPC increase. There are four of these in a Render Slice, and five Render Slices on the "BMG-G21." Each Xe core has eight 512-bit vector engines, eight 2048-bit XMX matrix engines, support for 64-bit atomic ops, and a large 256 KB L1 cache. The ALUs now support SIMD16 and SIMD32 ops, while the XMX units support data types ranging from INT2, to INT4, INT8, FP16, BF16, and TF32. Intel's second generation Ray Tracing unit features a 50% to 100% increase in hardware throughput, thanks to three traversal pipelines, 18 box intersections, two triangle intersections, and a 16 KB native BVH cache per RT unit. All said and done, Intel is claiming a 70% increase in performance per Xe core, and a 50% increase in performance-per-Watt with "Battlemage."

The 5 Render Slices and 20 Xe cores talk to each other over a next-generation global dispatch unit, and an 18 MB L2 cache. The GPU features a 192-bit GDDR6 memory interface. The odd count of 5 slices and 192-bit memory bus suggests that the BMG-G21 features a sixth render slice and a wider 256-bit interface, but that will probably be used in a higher SKU in the near future.


Then there are updates to XeSS. Intel's super-resolution based performance enhancement on its own improves frame-rates by anywhere between 22% and 80%, depending on the game. In games with ray tracing, XeSS proves crucial in reducing the performance cost of ray tracing. XeSS is available in over 150 games as of now. The company released a minor update to XeSS 1, with the introduction of a new Compute Dispatcher backend for the XeSS-SR SDK, with individual interfaces for DirectX 11, DirectX 12, and Vulkan.


With this release, Intel is introducing the new XeSS 2 feature-set, which consists of XeSS-SR (super resolution), XeSS-FG (frame generation), and XeLL (low latency). XeSS 2 isn't an update to the SR in and of itself—the image quality at any given XeSS setting is the same, but rather, it is a collective of SR, FG, and LL. XeSS-FG is a frame generation technology that seeks to nearly double frame-rates at any given SR setting, or even at native resolution. It relies on optical flow re-projection, and motion-vector re-projection to generate an interpolated frame. In this sense, it is technologically closer to NVIDIA DLSS 3 Frame Generation than it is to AMD FSR 3 Frame Generation.


XeLL is a system latency reduction technology that works to reduce the time it takes for a game input to register on screen. XeSS-FG requires XeLL to be enabled. There's also a new driver-based low-latency mode Intel is introducing, which simply compacts the rendering queue to reduce latency. We will bring you a more detailed architecture deep-dive as we get closer to the product launch.

The complete slide-deck follows, clicking below loads images from the Intel presentation.