Intel has introduced new processors with high-bandwidth reminiscence (HBM) geared towards high-performance computing (HPC), supercomputing, and synthetic intelligence (AI).
The merchandise are often called the Xeon CPU Max collection and GPU Max collection. The chips are primarily based on present expertise; the CPU is 4th Era Xeon Scalable, aka Sapphire Rapids, and the GPU is Ponte Vecchio, the information middle model of Intel’s Xe GPU expertise.
The distinction is that the 2 processors include HBM on the processor die fairly than relying solely on customary DRAM. HBM is significantly sooner than DDR4 or DDR5 reminiscence and sits on the processor die proper subsequent to the CPU/GPU core with a excessive velocity interconnect, fairly than on reminiscence sticks like DDR reminiscence.
“When you take a look at the general workloads within the HPC and AI area, there is a huge variety of workloads,” stated Jeff McVeigh, vp and common supervisor of supercomputing at Intel. “Historically, there’s been two routes up this summit. One is the CPU route and the opposite is the GPU route, they usually every have their very own obstacles. And our aim is to actually go ahead and tackle them holistically.”
The main focus of CPU Max and GPU Max is round maximizing the bandwidth, maximizing the compute, and maximizing the capabilities and prospects that they provide for addressing the breadth of workloads, stated McVeigh.
CPU Max
CPU Max is available in three server configurations. The primary is with out DRAM, so the one reminiscence within the system is 64GB of HBM on the CPU Max chip. That is how Japan’s Fugaku supercomputer, for a while one of many quickest supercomputers on this planet, operates. In a two-socket system, that’s 128GB of reminiscence, which McVeigh stated “for a lot of purposes and workloads is enough.” On this use situation, purposes can run unchanged.
The second configuration known as HBM flat mode, which mixes HBM within the CPU package deal with customary DDR5 reminiscence sticks within the system. With each HBM and DDR software program must be optimized to maneuver knowledge between these completely different reminiscence areas.
The third configuration is HBM caching mode, the place the HBM acts as a cache for the DDR reminiscence within the system. On this mode, no software program code adjustments are required. “You would possibly need to do some tuning to make the most of that very massive cache that you just now have, however you do not have to whenever you get quick advantages,” stated McVeigh.
GPU Max
GPU Max additionally is available in three configurations; the 1100, 1350, and 1550 fashions. The 1100 is a 300-Watt double-wide PCIe card with 56 Xe cores and 48GB of HBM2e reminiscence. A number of playing cards may be linked through Intel Xe Hyperlink bridges.
The opposite two configurations make use of the Open Compute Challenge (OCP) accelerator module, often called OAM, which is a sooner various interface to PCIe playing cards.
The 1350 GPU is a 450-Watt OAM module with 112 Xe cores and 96GB of HBM. The 1550 GPU is a 600-Watt OAM module with 128 Xe cores and 128GB of HBM.
The PCI Categorical card is nice to be used in a regular server even workstation methods, however the OAM modules are actually oriented for increased density environments, stated McVeigh. He stated Intel has received a lot of system designs being developed by OEM and system answer suppliers, who will convey out servers with OAM beginning in 2023.
Intel is constructing a supercomputer for the Argonne Nationwide Lab with CPU Max and GPU Max processors that, when it goes on-line in 2023, will exceed 2 exaFLOPs of efficiency. That’s double the velocity of Frontier, the present chief within the supercomputer race. McVeigh stated it’s the mix of the 2 processors that makes it occur.
“You’d argue, nicely, we have offloaded every little thing to the GPU, we do not want that highest finish CPU, proper? We do not want the HBM reminiscence, proper? Fallacious. We will achieve vital efficiency enchancment by turning on the HBM that is built-in throughout the CPU as a result of there’s nonetheless a whole lot of code that runs on the CPU, even when we have offloaded a few of our bigger kernels off to the GPU,” he stated.
The brand new processors have already begun transport to preliminary clients, together with Argonne. The Max Collection is slated to launch in January 2023.
Copyright © 2022 IDG Communications, Inc.
