Intel Showcases Arc Alchemist GPU Performance With XeSS, Raytracing & Hybrid CPU Combos
This isn’t the first time we will be talking about Intel’s XeSS or Raytracing implementations but they did share some new details and performance metrics. Intel Arc Alchemist XeSS Technology (GDC 2022 Demo) Intel XeSS is a super-sampling technique that leverages Machine Learning to reconstruct a low-resolution frame into a high-resolution frame, running at a fraction of the cost of rendering at a higher resolution such as 4K. The core principle of XeSS is to take advantage of ML through the integrated XMX AI acceleration hardware that is featured on the Xe-Cores for Arc Alchemist GPUs. XeSS will be fully compliant with DirectX 12 and uses the Intel Vector Shading language-based NN (Neural Networking) running on Intel’s Arc SIMD architecture. Compared to native resolution (4K), XeSS 4K with XMX takes less than half the cost to render a scene. The technology is also DP4a compliant which means it can run on GPUs without AI or ML acceleration engines such as XMX but offer similar performance and image quality. In terms of image quality, Intel XeSS will eliminate all kinds of ghosting with minimal artifacts (such as shimmering) when compared to other upscaling methods such as TAAU. It also comes with its own built-in sharpening engine and will be replacing TAA. Intel has stated that XeSS can achieve much higher scaling ratios without compromising quality than Temporal Supersampling or Spatial upscaling. Intel also shared a brand new XeSS Rens demo which was running on an Intel Arc Alchemist GPU that was running at a fixed frequency. The demo was run on 1440p with raytracing and 4K with raytracing enabled. The demo was run at each resolution on 5 different XeSS presets that range from Ultra Performance, Performance, Balanced, Quality, and Ultra Quality. The Ultra performance mode offers up to 2.53x performance boost over native while Ultra Quality offers a 27% boost in performance over native at 4K resolution. Intel Arc Alchemist Raytracing Technology (GDC 2022 Demo) Next up, Intel talked about its raytracing approach and how they are making it better than AMD’s and NVIDIA’s approach. Intel will be bypassing SIMD divergency for hit shaders, textures and using a set of HW sorting for rays and threads to maximize uptime of each lane. This would allow Intel to essentially automatically accelerate raytracing on hardware. In a slide showcasing performance metrics on a pre-production Intel Arc Alchemist silicon, the GPU offers 0.775x performance in RayQuery (relative) vs DXR1.0. Intel states that this is a bigger perf hit than competitor 1 (NVIDIA) while competitor 2 (AMD) takes an even bigger hit. Intel also offers their own conclusion and fixes to why this performance hit is seen on their GPUs. 12th Gen Hybrid Core Technology (GDC 2022 Demo) Moving away from GPUs for a bit, Intel also discussed how they can leverage their hybrid design introduced in 12th Gen Alder Lake CPUs within game engines. Intel and IO Interactive have been optimizing both the GPU and CPU sides of things. It is stated that developers can leverage back-ground tasks such as AI acceleration, Character Animation, Physics, Collisions, audio-processing, and more, leaving the performance cores with their leading single-threaded performance to be available for the more demanding tasks. We finally have a performance evaluation shared by Intel, comparing just the P-Cores with a P-Core and E-Core enabled configuration in the Hitman 3: Dartmoor benchmark: Intel states that the 12th Gen Hybrid architecture allows for up to 4% faster average frame rate, 12% faster 90%tile, and 2 just 2% slower 10%tile FPS. The biggest improvement is within the physics framework which sees a wait time reduction of 55%, 90% reduction by 1.7ms, and a 25% reduction in overall frame time. The faster frames do slow down a bit due to the requirements to wait on the render thread. These are just the initial details and we are looking forward to the launch of Intel’s first Arc Alchemist GPUs on 29th March.
