As expected, Intel has to make an announcement in the server area shortly before the end of the year. However, this is not the new Xeon generation based on Ice Lake, but an Intel server GPU. This has also been announced in the recent past, because like the integrated graphics unit of the Tiger Lake processors and the dedicated Iris-Xe-Max-GPU, it is based on the Xe-LP architecture.

With the initially meaningless server GPU, Intel is targeting the market for video and game streaming on the Internet. High growth rates are expected in this area in the coming years and in addition to Google Stadia, NVIDIA, Sony and Microsoft have also started corresponding services.

The server GPU now presented goes by the name of H3C XG 310. It is actually not a GPU, instead there are three Xe-LP GPUs on the PCI Express card. In addition to the hardware, Intel offers the associated software ecosystem, which consists of the necessary drivers, APIs and developer tools. Intel relies on a mixture of open source and licensed software.

For rendering, the focus is initially on services that offer Android games in the cloud. The games are rendered on the execution units of the GPU via Android in Container (AIC). An interface called Intel Cloud Rendering (ICR) receives the inputs of the Game Service Streamer – the rendered frames of the AIC are brought together and an audio and video stream is created from this, which is then delivered via the Game Service Stream.

On the video stream side, the GPUs or the server GPU support AVC, HEVC, MPEG2, VP9 (encode and decode) and AV1 (decode) as expansion cards. Intel can also offer a corresponding FFMPEG package with corresponding hardware acceleration for Xeon processors and the server GPU.

While the Iris-Xe-Max-GPU with 4 GB LPDDR4 on a 128 Bit wide memory interface, the Xe-GPU offers the server GPU an 8 GB LPDDR4 memory. Four of these GPUs are located on the XG 310 from H3C, which is a 3/4 Length – Full Height PCI-Express card that supports PCIe 3.0 and 16 Lanes is connected. The card or GPU is cooled passively and via the air flow from the server. The additional power supply is provided via an 8-pin connector. Four of these cards can be used per server / Xeon processor.

In a demo, Intel showed an Android RPG or the streaming of the same 60 Instances each with 30 FPS. Initially, only one XG 310 with four GPUs is used. With two XG 310 cards 120 instances / streams can be executed. Depending on the game and server configuration, 100 to 160 users can work on one server with a Xeon processor and two servers at the same time -GPUs play. This is of course particularly interesting for the operators of cloud streaming services.

One of the first customers for Intel’s server GPU is Tencent in China.

Another outlook on the Xe roadmap

Intel also gave another outlook on the Xe roadmap. The Xe-LP architecture is already used in the integrated and dedicated GPUs for notebooks. The server GPU is now the third area in which it is used.

In the coming year, the next generation of dedicated graphics solutions is expected with the Xe-HPG architecture, aimed at gamers and desktop systems. Intel has already announced that the first samples of this GPU will work in the laboratory. Xe-HP is the data center variant that is already delivered to partners in the form of samples. We are still working on Ponte Vecchio based on the Xe-HPC architecture, but we have already received the first revision of all chip sets for Ponte Vecchio from production.