Intel Arc Alchemist, with Xe-Core and XeSS is a challenge to Nvidia and AMD on gaming GPUs

Intel announced in recent days Intel Arc, the commercial brand of GPU Car HPG (High Performance Gaming) – codename Alchemist – at the base of the dedicated video cards that will face the market with the Nvidia GeForce and AMD Radeon solutions starting from first quarter 2022.

During the Architecture Day 2021 Intel told us not only about Xe HPG but also about XeSS, abbreviation that stands Xe Super Sampling, a technology that like Nvidia’s DLSS and AMD’s FSR wants to guarantee gamers an image quality similar to the native one but with much higher performance, especially where ray tracing is enabled.

Xe HPG, with Alchemist a new chapter in the history of Intel

Intel explained to us, through the mouth of its chief GPU architect Davide Blythe, that HPG il concentrato di Car LP, HP Car e HPC: from the first it took the graphic efficiency, from the second the scalability and from the third the calculation efficiency.

The founding block of Xe HPG is not the Execution Unit (EU) we have been used to for some time, but Xe-core, a whole new unit with arithmetic blocks, caches and units of load and store, that is everything necessary to manage traditional loads and also those of AI. We talk about 16 Vector Engine and as many Xe Matrix Engine called XMX, assigned to AI tasks such as XeSS which we will talk about later.

Within what Intel calls “Render Slice“we find quattro Xe-core con XMX, but also the necessary to render the graphics in real time, namely the fixed functions such as Geometry Pipeline, Rasterization Pipeline, Samplers and Pixel Backends, all optimized to take advantage of Microsoft’s DirectX 12 Ultimate API, but also Vulkan.

Then in the Render Slice there are four units that deal with the management of calculations related to ray tracing (ray transversal, bounding box intersection, triangle intersection). According to Intel, in the flagship HPG GPUs we will see up to 8 of these Render slices, flanked by a large L2 cache.

Intel also explained to us that it has worked hard to ensure the energy efficiency and, at the same time, the operating frequency of the graphics chip. As a result, the optimization of every part of Xe-core affected not only performance, but other aspects as well. Intel isn’t ready to reveal everything yet, but the company claims it does having improved performance per watt and frequency by 50% compared to Xe LP.

Incidentally, Intel has made it known that the Xe HPG GPUs will be manufactured by TSMC with the N6 process. Alchemist GPUs are in the sampling phase and will debut in Q1 2022. In the future, but the company has not clarified when, Alchemist will be followed by Xe2 HPG, code name Battlemage and then from Xe3 HPG “Celestial”. It also appears in the roadmap Druid, but in this case we speak generically of “Xe Next Architecture”, which could imply a more radical revision of the project.

XeSS, a competitor for Nvidia DLSS and AMD FSR

Designing a well-made GPU (or at least that’s what Intel says) is an important step, but to make it work at its best you also need some optimized drivers. An ill-conceived driver can in fact sink even the most powerful GPU in the world.

The US firm has worked along this past year for re-architect your driver, going to work on the heart, in particular on the compiler and the memory manager. “We improved throughput in CPU-bound titles by 80 percent and game load times by 25 percent,” said Lisa Pearce, Intel’s head of graphics software development. In parallel, Intel worked to “pick up the pace” and release optimizations for various upcoming titles quickly and on time.

Intel’s driver team is also working hard to better support core APIs, DirectX 12 Ultimate and Vulkan, as well as popular game engines such as Unreal Engine (Intel claims its dedicated GPUs are running the UE 5 already today). and Unity. At launch, Intel GPUs, for example, will support technologies such as mesh shading, sampler feedback e ray tracing via hardware.

There will also be the rest of the features that other GPU manufacturers offer, such as the ability to stream games, a modern user interface to act on various parameters and obviously also spend in overclocking.

Among the innovations in the pipeline there is alsoXe Super Sampling, XeSS. Lisa Pearce explains that this is an easily integrated API that uses deep learning to synthesize images with a quality very close to native quality. “It works by reconstructing subpixel details from adjacent pixels as well as from previous frames with motion compensation.”

The reconstruction work done by a trained neural network to ensure “high performance and high quality”. Intel showed us a comparison demo made with the Unreal Engine of a native 4K rendering versus a 4K rendering with XeSS (1080p upscaled), in which Intel’s technology not only guaranteed. an image practically identical to the native one, but up to twice the performance.

XeSS it will also work on GPUs from other manufacturers, albeit with slightly more overhead than Intel GPUs due to the use of DP4a instruction over XMX. The XeSS SDK, in order to give developers a way to get familiar with it, will be available from this month in the XMX version while the DP4a version will debut within the year.

Too early for specs and performance

For now, Intel is not yet ready to enter into the merits of the product lineup, technical specifications, their performance and prices. As previously pointed out, we have to wait until the first quarter of 2022 to see them on the market and probably from here to then the company will rattle off more and more information to create attention and expectation on its return to the world of dedicated video cards.

On the subject of Intel Architecture Day 2021, we suggest reading the article dedicated to Alder Lake microprocessors coming to the client – notebook and desktop market:Intel Alder Lake, here’s how the high-performance hybrid processor coming this fall works.