Intel introduces Graphics Xe for data centers

After years of waiting and delays, last week Intel announced that the fourth generation of the Xeon Scalable, codenamed “Sapphire Rapids,” is finally ready for some sort of launch on January 10, 2023. Now, more pieces are joining the server puzzle in a form of a family Max oriented to the data center, where, among other things, the graphics project “Ponte Vecchio” comes to the fore.

Just as with Arc graphics cards on both consumer and WorkstationAs an aside, the Intel Xe architecture is the basis for the new computation cards, but in the form of Xe-HPC rather than Xe-HPG. Here, the company targets data centers and supercomputers with PCI Express interface cards and so-called OAM modules.

In complete design, it’s a graphics circuit with 128 Xe cores, compared to the 32 found in the consumer-oriented ACM-G10 in the consumer-oriented Arc A770. It is accompanied by up to 128 GB of high-bandwidth memory (HBM) and up to 64 MB and 128 MB of L1 and L2 cache, respectively. Many Xe cores also mean much more ray tracing– Course (m. ray tracing unitsRTU) and Intel further emphasize that this is useful for heavier rendering.

For the variant connected via PCI Express, 56 Xe cores and 48 GB of HBM memory are implemented, with the ability to connect multiple computing cards to each other using Intel Xe Link bridges. Here, Intel is catching up with Nvidia in terms of power supply and implementing the new 12VHPWR standard with the associated 16-pin connector.

The Max family doesn’t just include graphics. Processors also belong to the series. There, Xeon Max is the news, there is also on the basis Imminent “Sapphire Rapids”. As expected is chipConvenient design, with four smaller circles containing a maximum of 15 centers each. This gives a total of 60 cores, but in the case of the Xeon Max, it comes to 56 active cores at most.

The Xeon Max is the first x86-based processor with onboard HBM, something that Intel believes should provide an adequate performance boost in loads that include, for example, simulation, machine learning or database management.

With the Max family, Intel is taking another step towards realizing the scalable model in which the company has invested so heavily. Partly with a layout that can be scaled from integrated graphics in processors to advanced computing cards for data centers, partly by bringing together members of the Max family. Processors and graphics circuits should work optimally together, a common arrangement not quite unlike Nvidia’s concept with its ARM-based “Grace” and “Hopper” graphics architecture.

However, Intel itself compares mainly the Max family with the EPYC 7003 series from AMD “Milan-X”, in which the Zen 3 architecture is combined with a stacked cache. According to the company’s own performance numbers, the performance is about twice as high, as a result of switching to HBM memory. However, we still have to figure out how well Intel is equipped to face AMD’s next “genoa” with up to 96 Zen 4 cores.

Read more about workstation and data center graphics: