Intel has offered up fresh details on its new line of datacenter GPUs, codenamed Arctic Sound-M (ATS-M), first announced in February.
At Intel Vision 2022, the company revealed there will be two ATS-M variants; one 150W card for peak performance, and a 75W card for maximum density.
Based on Intel’s Xe-HPG microarchitecture (the same that underpins the firm's Arc chips), the new server GPUs boast 4 Xe Media Engines, up to 32 Xe cores and a built-in AI accelerator. The company remains tight-lipped about the rest of the specification.
Arctic Sound-M for visual cloud
Intel’s new datacenter GPUs are built specifically for visual cloud, a collection of workloads and use cases that seek to deliver high numbers of pixels for purposes ranging from cloud gaming, to virtual desktop infrastructure and AI inference.
During the day one keynote, the lineup was characterized by Raja Koduri, the brain behind Intel’s GPU efforts, as “a media supercomputer on a single chip”.
According to Koduri, Arctic Sound-M chips are capable of handling more than 30 full-HD video streams, 40 cloud gaming streams, and 62 virtual desktops, and hit as many as 150 trillion AI operations per second.
Presumably, the new datacenter cards will synergize with other Intel silicon too, offering performance advantages unavailable with mix-and-match configurations courtesy of Intel’s Deep Link technology suite.
The company’s new Arc client cards, for example, can shift power dynamically between the GPU and CPU, depending on the demands of the specific workload.
Intel was also eager to emphasize the importance of the open software stack supporting Arctic Sound-M, based on oneAPI standards, which should enable developers to build easily for the platform.
The new datacenter cards are set to launch inside servers from fifteen vendors – including the likes of Dell, Supermicro and HPE – in Q3 2022.