Intel enters the GPU game with its Xe-HPG (DG2)

by

You can also be interested in these:

Intel promised the launch of the Intel Xe-HPG as a GPU optimized for gaming a few months ago. Time has passed and with the AMD RX 6000 and Nvidia RTX 3000 dominating the market, we know almost nothing about it.



Intel has a curious history with GPUs in the last twenty years. Starting from the failed Intel i740 that was anticipated over and over again with the revolutionary Lockhead Martin’s Real3D technology. Also the cancellation of the Intel Larrabe that became the now discontinued Intel Xeon Phi. To our point, Intel’s success in graphics processors in the last two decades has been inversely proportional to the success of its CPU business.

With the Intel Xe architecture, not only Intel aims to gain performance in its integrated GPUs, but also to be able to compete against NVIDIA and AMD in the gaming market with their own graphics cards.

The Intel Xe Architecture

During the Architecture Day 2020 details of the Intel Xe graphics architecture were seen and several product releases were announced, including the Intel Xe-HPG, a GPU oriented for gaming. This means that Intel officially made a declaration of war a few months ago against AMD’s Radeon and NVIDIA’s GeForce.

The ability of the Intel Xe architecture to reach various markets is due to the modularity of its architecture where different units can be added and removed according to their needs and the target market. This allows Intel to adapt its GPU according to the application in several different versions, where the architecture is common, but units are added and removed according to the final application and therefore to the market each version serves.

Xe-HPG: Intel DG2 architecture for gaming

While the Intel Xe HP and HPC are based on a configuration of various GPUs operating in parallel and built using chiplets, with technologies such as EMIB and Foveros, the Intel Xe-HPG on the other hand is nothing more than the classic GPU designed for a lifetime. This is mounted on a graphics card and has its own VRAM. So the Intel Xe-HPG like the Xe-LP concentrates all its technology in one chip, but we are going to get 2 different flavors of it:

The first configuration has a GPU with 512 EUs and 16 GB GDDR6 of VRAM. This amount of EU equals 4096 ALUs in FP32, which is the same as an RX 5000 or an RX 6000 with 64 Compute Units in the case of AMD. In the case of NVIDIA it would be equivalent to an RTX 2000 with 64 SM and an RTX 3000 with 32 SM.

The GPU has support for Ray Tracing and Variable Rate Shading, but everything indicates that the equivalent to NVIDIA’s Tensor Cores in the form of “Matrix Units” will not be seen in the first version of the Intel Xe-HPG. Especially considering that the support of DirectML functions are optional at the moment.

As for the second GPU, this one is simpler, since it only has 128 EUs. Given its low configuration, we do not know if we are going to see it in the form of a graphics card populating the low-end range or instead we are going to see it in the form of second-generation Intel Xe Max, especially when Intel’s Alder Lake processors will come out, both in laptops and in desktop. So the GPU with less specs could be a replacement for Intel Xe-LP.

Delays on the Xe-HPG architecture

The current Intel Xe-LPs on the market are not the definitive form of the Intel Xe, as they are a preview version called Intel DG1. The performance is just not enough to compete against NVIDIA and AMD. Despite the jump in the first Intel Xe with respect to the Intel Gen 11 GPUs, not to be confused with the CPUs, it would still be lower in terms of “cores” and clock speed than the NVIDIA and AMD architectures.

Despite the fact that an architecture may be the same as another on paper and diagrammed, in reality, engineers do not have much time to make profound changes when it comes to working in the short term, but rather they gradually make them according to planning and go adding improvements and replacing items over time in orderly planning. Many of the optimizations and changes that Raja Koduri and his team were unable to add to Intel DG1 have been added to Intel DG2.

The reason why we have not seen an Intel Xe-HPG throughout 2020 is because then it would have been based on Intel DG1 and it would have been a real performance fiasco. Intel cannot take any false steps in such a competitive market and especially the chief architect, Raja Koduri, who was shot out of AMD after the AMD Vega fiasco.

The second reason for the Xe-HPG delay would be that part of the Intel Xe workforce would have moved to help finish Intel’s twelfth-generation CPUs under the Alder Lake architecture. Let’s not forget that the main market for Intel is CPUs and it is where they make the most money and AMD has become very strong in recent years.

Taking into account that the design of the Intel Xe has needed to be tweaked in order to increase its performance, this has placed the GPU outside the 2020 launch date and the first half of 2021. For what Intel has been able to move part of its workforce to finish Alder Lake throughout 2020. Either that there are already working samples of Alder Lake, so the silicon implementation of the new CPU was already completed a few months.

The future looks promising for Intel

Just a few days ago, Raja Koduri surprised us with a test of the Intel Xe-HPG, finished and in working order, through his Twitter account. The fact that the GPU already runs the 3D Mark test that measures the performance of the card making use of the Mesh Shaders, which are an integral part of DirectX 12 Ultimate, is a great advancement. Mesh Shaders, also called Primitive Shaders, require a series of functionalities and characteristics in the GPU not available in the current Intel DG1 (Xe-LP) model.

The fact that the GPU is finished does not mean that it will be released tomorrow or next month, processors usually have a pre-build period that lasts about six months before mass production. At the moment Alder Lake is more advanced and does not come out until September, so logically we should wait until the end of 2021 to find out when the Intel Xe-HPG will be available.

More stories like this