Intel has delivered most of the Aurora supercomputer to the Argonne National Lab, and it promises over two exaFLOPS of performance.
Although Intel's business has suffered all across the board in recent years, its high performance computing (or HPC) processors have perhaps lost the most ground due to back-to-back delays. However, Intel might be on the path to recovery with its latest announcements at the ISC High Performance 2023 conference, as the industry titan disclosed the Aurora supercomputer was nearing completion after years of development hell and that progress in developing upcoming HPC chips was steady and going as planned.
Aurora nears completion with over 10,000 blades shipped
When the Aurora supercomputer was first announced in 2015, it was supposed to be the first supercomputer to have one exaFLOP in computing power; for reference, one exaFLOP is equal to one million teraFLOPS or about 12,000 RTX 4090s. Aurora was originally slated for a 2020 launch, but that didn't happen because Intel's 10nm process kept getting delayed. This allowed AMD to power the first exaFLOP supercomputer, Frontier, but this delay actually hasn't been a complete disaster for Intel as Aurora's maximum performance has been upgraded from one exaFLOP to more than two.
The Aurora supercomputer is composed of "blades" (also called nodes), which are essentially computers in their own right with each blade holding two 40-core Sapphire Rapids CPUs and six Ponte Vecchio GPUs. There are 10,624 blades in Aurora which means the supercomputer will have 21,248 CPUs and 63,744 GPUs when it's complete. Intel claims in "real-world science and engineering workloads," Ponte Vecchio GPUs are significantly faster than AMD's MI250X and Nvidia's A100 and H100 GPUs.
With over 10,000 blades already delivered to Argonne National Laboratory, Aurora is on track to be powered on later this year and will have more FLOPs than Frontier. FLOPs aren't the only thing that determines the overall performance of a supercomputer, but it does give Aurora an advantage.
Intel's HPC roadmap promises a return to regularly scheduled launches
Now that Intel has finally gotten its Intel 7 (formerly 10nm) processors out the door, it can finally focus on upcoming chips that use the Intel 4 (formerly 7nm), Intel 3, 20A, and 18A process nodes. Intel hopes to be able to achieve a yearly release cadence for its Xeon CPUs and a bi-yearly cadence for its HPC GPUs, and the result is a pretty crowded roadmap.
By the end of 2025, Intel plans on launching four new Xeon CPUs. Emerald Rapids is slated for a launch late this year and will feature Raptor Cove P-cores (the same present in 13th-generation Intel Core CPUs). Granite Rapids will succeed Emerald Rapids in 2024 and will utilize the Intel 3 process and Multiplexer Combined Ranks (or MCR) memory with up to 8,800MT/s. Meanwhile, Intel's first E-core Xeon, Sierra Forest, is supposed to launch in 2024 prior to Granite Rapids and will also use the Intel 3 process. Afterward, Clearwater Forest seems scheduled for launch in 2025 but there are few details beyond that.
Intel's GPU roadmap changed earlier this year as the company canceled the Max Series Rialto Ridge GPU and Flex Series Lancaster Sound, which were supposed to launch this year. Intel has decided to instead focus on hastening development on Falcon Shores and Melville Sound, which will likely launch in 2025 and 2024 respectively. Alongside those cancelations, Intel also said Falcon Shores would no longer offer an "XPU" option with both CPU and GPU cores, but still has plans for an XPU at a future date.
Beyond the healthy amount of releases Intel plans up to 2025, the company's 2023 roadmap is the first in a very long time to feature more on-track chips than delayed chips. The only significant changes from the 2022 roadmap from early last year are the canceled GPUs. This doesn't mean Intel won't delay anything within the next couple of years, but it's a very good sign for a company that has been weakened by delays upon delays.