The data must flow

Intel is diversifying its GPUs to work with architectures other than x86 chips, which could be a step toward making the chip maker a manufacturing-first company.

Intel's upcoming GPU called Ponte Vecchio will be used in supercomputers with ARM-based chips made by company SiPearl.

Those computers will be in production in 2023, Philippe Notton, CEO of SiPearl, told The Register. SiPearl's ARM-based CPUs, called Rhea, also has RISC-V controllers.

There is interest among high-performance customers in Europe for this type of system, Notton said.

Intel a few days ago announced it is putting its Ponte Vecchio GPU in high-performance computers with SiPearl's Rhea CPU, but didn't mention the CPU is based on an ARM Neoverse V1 design.

Intel did not return request for comment on Intel GPU compatibility with ARM processors in PCs and servers.

Ponte Vecchio is also being paired with Xeon Scalable chips code-named on the upcoming Aurora supercomputer, which will deliver two exaflops of performance.

The mega GPU delivers 45 teraflops of FP32 performance. It has close to 100 billion transistors, and is based on a stacked design in which up to 128 graphics Xe graphics cores are closely linked up in a mesh. Detailed specs about the chip are available at The Next Platform.

SiPearl is also developing CPUs to meet the European Union's goal of making indigenous chips and building an exascale supercomputer based on homegrown chips by 2023.

Keeping pipelines flowing

Intel and SiPearl have optimized a low-latency communications pipeline for the CPU and GPU, Notton said. The Compute Express Link interconnect specification facilitates communication between CPUs, GPUs, and other accelerator chips.

Intel is looking to top Nvidia in the GPU, while ARM are looking to catch up with x86 in high-performance computing systems. The claimed world's fastest supercomputer, Fugaku, runs on ARM CPUs with vector extensions. Seven of the remaining Top 10 supercomputers in the Top500 list have Nvidia GPUs with either x86 chips or IBM's Power processors.

Intel's interest in making its GPUs compatible with ARM, and working with SiPearl, could have multiple benefits. Intel could receive more subsidies and funds from the EU to make advanced factories so chips – be it ARM CPUs or its own GPU – can quickly and easily available to regional customers.

Governments are shelling out billions to chip manufacturers to establish fabs in the region, and Intel is investing €80 billion to build factories in Europe. Intel has opened up its manufacturing facilities to also make ARM and RISC-V chips.

Intel factories in Europe would strength the European chip supply, Notton said, adding that it would give his company another option to acquire chips at reasonable cost.

"We don't have any choice right now. It has to be made in Taiwan," Notton said. 

SiPearl and Intel are also bridging the software gap between ARM processors and the Intel GPU, Notton said. Intel's OneAPI provides a one-stop-shop for developers to write applications for any chip type or architecture.

"What we are going to do with OneAPI, you will have compliance between OneAPI and ARM architecture," Notton said.

ARM was the only available option for SiPearl to build a supercomputing chip as it licensing options were not available for x86, and RISC-V isn't mature, Notton said. He noted ARM's strong presence in European supercomputing scene with systems like Barcelona Supercomputing Center's Mont Blanc-1, which was introduced in 2015 and considered one of the world's first ARM supercomputers.

But SiPearl is also cranking up work on open-source RISC-V as ARM's European heritage is slipping, Notton said. It was originally a British company, then got acquired by SoftBank, then came Brexit, and it is in the process of being acquired by Nvidia. ®

https://www.theregister.com/2021/10/29/intel_gpu/