Intel Arc: DG2 graphics card Alchemist with 6 nm production

Share your love

Intel has revealed a few additional details about its upcoming Arc graphics cards, which should make gamers’ hearts beat faster from the first quarter of 2022. In addition to availability, performance and prices are still marked with big question marks.

Just a few days ago, Intel announced the delay to 2022 together with the newly invented brand name “Arc”.

The first gaming graphics cards use the GPU called Alchemist, followed by Battlemage, Celestial and Druid. Just as creatively as Intel switched from “Gen” for Generation to “Arc” for Architecture, the chips are sorted from A to currently D.

The “Alchemist” GPUs are compatible with DirectX 12 Ultimate. This means that, in addition to the performance-saving variable rate shading and the previously rarely used mesh shaders, you can also accelerate ray tracing in hardware – this is what sets you apart from the integrated Xe graphics and the rarely used DG1 offshoot.

Xe-Core with vector and matrix engine

(Image: Intel)

Intel has also turned a few screws in the architecture. Instead of execution units, there are now 16 vector and 16 matrix engines (“XMX”) in the smallest organizational form of Xe, the Xe core. The vector units correspond to the EUs and each create eight 32-bit operations per cycle, i.e. fused multiply add instructions that are used to calculate the theoretical throughput. The XMX is four times as wide, but Intel has not yet confirmed which data formats it can handle in the gaming version Alchemist – it is likely that in addition to INT8, there is at least FP16. The HPC version in Ponte Vecchio, like Nvidia’s Tensor Cores, can use the TF32 (“Tensor-Float”) and BF16 (“Brain Float”) formats, as well as FP16 and INT8, which are intended for AI training.

The XMX engines will be used in 3D games with what is known as Xe super sampling. These are Intel’s competitors to Nvidia’s DLSS, a technology that uses a pre-trained neural network to upscale and reconstruct images in order to increase performance, especially in ray tracing titles, without impairing the image quality too much. Xe-Super-Sampling should take into account image information both temporally and spatially surrounding pixels in order to output the final image.

Xe SS: AI-supported upscaling with reconstruction

(Image: Intel)

A Xe core also gets instruction and L1 caches, which can also be reconfigured as scratchpad local memory via software, similar to Nvidia’s current GeForce graphics chips. Fixed load / store units are also provided.

Four Xe cores are combined in a so-called render slice and are given fixed function blocks for rasterization, advance depth tests, geometry processing, texture sampling (4×8), media encoding and decoding as well as raster output stages (2×16).

Ray tracing units with traversal

(Image: Intel)

In addition, there is a ray tracing unit per Xe-Core, again similar to Nvidia, which not only checks the box and triangular intersections of the rays being tracked, but also takes on what is known as ray traversal, i.e. the pacing of the ray through the acceleration structure.

In a DG2 alchemist, eight such render slices work together with a total of 512 ex-excution units, 256 texture samplers and 128 raster output stages as well as 32 ray tracing units, plus a level 2 cache that scales with the number of render slices .

Intel has now officially confirmed that the DG2 chip “Alchemist” with 6 nanometer technology is manufactured by the Taiwanese chip manufacturer TSMC. The process already uses extremely short-wave UV rays to illuminate the structures.

Xe HPG: 50 percent higher clock rate, 50 percent more efficient than Xe LP

(Image: Intel)

Among other things, the manufacturing technology allows the clock rate to be increased by 50 percent compared to the DG1 graphics card, promises Intel. But the architecture itself, the transistors used and the design of the circuits also play a role. This is not explicitly mentioned, but the GD1 cards ran at around 1.5 GHz – half of them on top and you are at around 2.25 GHz, a clock rate with which the first benchmark results for the DG2 cards have already been viewed in a database . The efficiency, i.e. the performance per watt, should also increase by 50 percent.

Exactly on which day the cards will hit the market, what they will cost and how readily they will be available, is just as unclear as the final performance. Because one thing is clear: Success also depends on the quality of the drivers.


Article Source

Read Also   How to create a folder in Telegram
Share your love