Intel’s CEO lays out two tracks for its Xeon chip development through 2024 with implications for servers, supercomputing, and AI.
At this year’s Intel’s investors’ day meeting with Wall Street analysts, CEO Pat Gelsinger revealed new road maps for Xeon CPUs and Xe GPUs that you would typically expect to see reveals at an IDF show that stretches through 2024.
Most notable about the Xeon news is that for the first time, Intel is bifurcating the processor line into two microarchitecture types, Performance and Efficient. If that sounds reminiscent of the Alder Lake hybrid architecture currently used in client CPUs, it is but it’s very different implementation.
Adler Lake uses two types of cores; the performance core used to do the computing, and the efficient core, used to do small background tests that don’t require a high-performance core. For the Xeon, Performance and Efficient cores will be two separate processors.
The P-core chips will be used primarily for high performance-per workloads, like AI, HPC, and databases. The E-core processors are designed to maximize power efficiency for highly-parallel and less demanding workloads.
By not mixing the two types of cores on a single chip as they did with Adler Lake, Intel’s customers can better optimize P-core and E-core systems based upon their needs. For compute intensive tasks, they can go with a lot of P-core servers, and the company with less demanding tasks, such as VMs, can deploy E-cores.
Xeon P and E won’t be on the same motherboard since they will live on different machines but can be in the same data center and even in the same server rack.
The next generation of Xeon processors is known as Sapphire Rapids, and is due out later this year. Following Sapphire Rapids, Intel will deliver Emerald Rapids in 2023 and Granite Rapids in 2024. The second-generation of all new Xeon processors using efficient cores is called Sierra Forest and will arrive in 2024.
On the supercomputer side, Intel announced a new architecture called Falcon Shores that will combine a Xeon CPU and Intel Xe GPU into a single Xeon processor. Intel claims Falcon Shores will provide five times greater performance-per-watt, five times greater compute density in an x86 socket and five times greater memory capacity and bandwidth when compared to Intel’s current generation of products. But don’t expect to see it until at least 2024.
GPU-based audio encoder
Here’s a new use for GPUs: Intel announced Arctic Sound-M, which it said is the first GPU in the industry to include a hardware-based AV1 encoder for media processing.
The GPU will be focused on four areas: video transcoding, cloud gaming, virtual desktop infrastructure (VDI), and media AI analytics. In a separate video presentation, Raja Koduri, head of Intel’s Accelerated Compute Systems and Graphics Group, said media transcode needs, along with cloud gaming, real-time media analytics, and virtualization, are all being addressed by several discrete chips today.
“There is no seamless hardware-software media solution that collectively addresses quality density and latency requirements until now,” he said.
For video transcoding, Intel said Arctic Sound-M will be capable of eight simultaneous 8K streams and more than 30 simultaneous 1080p streams. For virtual desktop infrastructure, the GPU will be able to provide more than 60 virtualized functions, and in cloud gaming, it will have the ability to stream more than 30 game sessions at once. For media AI analytics, Koduri claims the GPU will be able to perform 150 trillion operations per second.