During his speech, Mr. Koduri took out of his pocket the largest and most powerful graphics processor X ^e HP, which combines four crystals on one substrate (there will be simpler versions with one and two crystals ). Each of them includes 512 computing units (EU, Execution Units) and, most likely, two HBM2e multilayer memory modules.

Then, the head held the test to show excellent scalability architecture X ^e and how to increase the number of crystals on the substrate leads to multiple sizing performance. Working on a single crystal, the accelerator provides up to 10,588 teraflops of FP32 floating point at single precision. When two crystals are involved, performance scales almost perfectly to 21,161 teraflops (1.999x). The flagship quad-die solution delivers 3.993x the performance boost to 41.908 teraflops.

The distribution of tasks between GPUs in technologies such as SLI and CrossFire is much more difficult, and scaling from additional accelerators usually gives players again in the region of 50-80% at best. However, in compute workloads, tasks are often independent and can scale perfectly. Therefore, in professional tasks, Intel’s flagship GPU with four crystals will clearly show itself well. And here is how things will be at Intel with graphics scaling – independent tests will show.

For comparison: the GeForce RTX 2080 Ti graphics card in FP32 operations is capable of providing up to 14.2 teraflops of performance. But that’s not all: Mr. Koduri also mentioned that the X ^e HP based on four crystals is capable of providing performance in petaflops. In other words, using tensor blocks, Intel’s new GPU will be incredibly fast in machine learning and artificial intelligence tasks where high precision is not required.