Intel XeSS: among the novelties of the Architecture Day 2021 there is also the rival of the DLSS

Intel XeSS: among the novelties of the Architecture Day 2021 there is also the rival of the DLSS

Intel XeSS

At Intel Architecture Day 2021, Raja Koduri and some of Intel's leading designers provided details on two new x86 core architectures, Intel's first high-performance hybrid architecture, codenamed "Alder Lake", and "Sapphire Rapids" , the next generation Intel Xeon Scalable processor for data centers, as well as on new IPUs (Infrastructure Processing Units) and new graphics architectures, including Xe HPG and Xe HPC microarchitectures, as well as Alchemist and Ponte Vecchio SoCs.

These new architectures will be the foundation of upcoming high-performance products and set the stage for a new era of Intel innovations aimed at responding to the growing global need for more computing power: in this regard, Raja Koduri has talked about the importance of advancements in architectures that aim to meet this demand.

The new Intel Efficient-core microarchitecture, previously not a codenamed "Gracemont Cove" is designed to improve throughput efficiency and ensure scalable multithreaded performance for modern multitasking. The microarchitecture and a targeted design allow the Efficient-core to operate at low voltage, thus reducing the overall power consumption, while creating the power margin necessary to operate at higher frequencies: in this way Efficient-core is able to increase performance when there are more demanding workloads.

As for Performance-core, formerly known by the codename “Golden Cove”, it is designed to prioritize speed and further reduce latency by increasing the performance of single-threaded applications. Workloads are supported by increasingly complex code and require greater execution capabilities. This new microarchitecture therefore offers a significant increase in performance for all uses and better support for applications supported by large code.

Let's move on to the graphics component: Xe HPG is a new dedicated graphics microarchitecture, specially designed to accommodate higher performance for gaming and media creation workloads. This microarchitecture is the foundation of the Alchemist System on Chip (SoC) family, and the first related products are expected to hit the market in the first quarter of 2022 under the Intel Arc brand. Xe HPG also introduces a new Xe-core, focused on computing, programmable and scalable.

Then there is Intel XeSS (Xe Super Sampling), which takes advantage of Alchemist's integrated XMX AI acceleration to offer a new upscaling technology that allows you to obtain high fidelity images, while reducing the load on the GPU. This technology uses deep learning to synthesize images that approach the quality of native high-resolution rendering, making sure that games that would only be playable at lower quality settings (or lower resolutions), can run smoothly at settings and resolutions. of superior quality. The main features are as follows:

Reconstruction of subpixel details from adjacent pixels as well as from previous frames with motion compensation: this reconstruction is performed by a specially trained neural network to offer high performance and excellent visual quality , with a performance increase of up to 50% AI-based super-sampling on a large set of hardware, including integrated graphics cards, leveraging the DP4a instruction set Several game developers are working on XeSS, and the SDK for the version The initial XMX will be available to ISVs this month, with the DP4a version available by the end of 2021. XeSS is therefore Intel's answer to Nvidia's DLSS and AMD's FidelityFX Super Resolution. The technology from the Santa Clara giant will also work on GPUs from other manufacturers, although performance will be worse as DP4a instruction will be used instead of XMX. By the end of the month, Intel will make the XMX version of the SDK available to developers, so they can start working with the algorithm, implement it in their titles, and make the most of it when the cards are released. The version with DP4a will instead be available within the year.

XeSS is Intel’s answer to Nvidia DLSS, and it has one big advantage

On the heels of the Intel Arc announcement earlier this week, Intel has revealed more details about its upcoming GPUs and the technologies they’ll feature — in particular, XeSS. XeSS is a new supersampling feature on upcoming Intel Arc GPUs, and it looks to combine the best of competing technologies from AMD and Nvidia.

Functionality-wise, XeSS works a lot like Nvidia’s Deep Learning Super Sampling (DLSS). It uses a neural network that has been trained on images of a game to reconstruct details from neighboring pixels, and critically, it leverages previous frames to track motion through a scene. This temporal information is what sets DLSS apart, and now it looks to have a competitor in the form of XeSS.

Intel XeSS quality comparison.

Intel demoed a scene upscaled from 1080p to 4K, nearly matching the quality of native 4K. The demo was running on an Intel Alchemist graphics card, which is the code name Intel is using for its first-generation Arc products.

Intel claims that XeSS can provide up to a 2x performance increase over native rendering. That sounds great, but Intel didn’t show XeSS running in any games or announce any titles that will support the feature (though Intel said “several” game developers are already engaged with XeSS).

The difference between DLSS and XeSS

XeSS looks functionally similar to DLSS, but it has one key difference. Intel announced that it will release two XeSS modes. One will leverage Intel’s dedicated XMX, or Xe Matrix Extension, cores to handle the artificial intelligence (A.I.) supersampling. That’s similar to how Nvidia leverages its Tensor cores in RTX graphics cards to upscale with DLSS.

The other mode is where things get interesting. It will use DP4a instruction, which is used for A.I. operations on recent Nvidia graphics cards and recent Intel integrated graphics. Intel claims there’s a “smart” performance and quality trade-off for the DP4a version. Regardless, what it means is that XeSS will still work on hardware that doesn’t have XMX cores.

Intel XeSS rendering pipeline demonstration.

That brings one of the major benefits of AMD’s FidelityFX Super Resolution (FSR) to XeSS. FSR supports just about any graphics card, while DLSS works solely on Nvidia RTX graphics cards. By utilizing two different software development kits (SDKs), Intel is able to bring support for a wide range of hardware while still utilizing A.I.-assisted upscaling on dedicated hardware.

That’s something we wanted to see out of Nvidia following the launch of FSR. XeSS looks like it will provide the best of both worlds, giving users with dedicated hardware a proper A.I.-assisted supersampling tool without restricting the feature to a certain platform. That said, we still need to wait on game support and independent testing to see how it stacks up.

Intel said that the XMX version of the XeSS SDK will be available later this month. The DP4a version will arrive “later this year.” Unfortunately, Intel didn’t provide any more details on the release window for the DP4a version. Regardless of the version, Intel says XeSS can easily fit in existing rendering pipelines.

A look inside Xe HPG

XeSS uses dedicated XMX units to handle the A.I. supersampling, and there are 16 XMX units on each Xe Core. Moving forward, Intel is moving away from the term “execution units,” which defined the number of cores on previous versions of Xe graphics. Instead, it will use “Xe Core” to show what GPUs have inside.

Each Xe Core features 16 XMX units and 16 vector engines, and there are four Xe Cores on each render slice. On the slice, each Xe Core also gets a dedicated ray-tracing unit for DirectX 12 and Vulkan ray tracing, as well as a shared L2 cache. This slice is the building block of cards using the Xe HPG architecture, and Intel is able to add or remove slices to meet performance targets.

Intel Xe HPG render slice model.

Currently, Intel is able to use eight slices on a graphics card, totaling 32 Xe cores and 512 vector engines. Previously, reports suggested that Intel’s flagship Alchemist card would use 512 execution units, and that still appears to be the case. The only thing different now is the naming.

Overall, Intel says that the Xe HPG architecture provides a 1.5x increase in frequency over Xe LP, as well as a 1.5x increase in performance per watt. These numbers come from Intel’s internal benchmarks, however, so we’ll need to wait until we have some time with a card to draw any conclusions.

Still, our first look at Xe HPG and XeSS is exciting. For decades, Nvidia and AMD have been the only two options for graphics cards. Intel is not only releasing a graphics card for gamers but one that comes with all the fittings modern GPUs need.

We’ll have to wait to see how Alchemist cards perform when they launch, but it would be nice to see a blue wave in a sea that’s been dominated by red and green for too long.

Editors' Recommendations

Powered by Blogger.