AMD is getting personal with chips as it sets sail to make products more to the liking of its customers.
The chipmaker detailed a modular chip future in which customers can mix and match non-AMD processors in a custom chip package.
“We are focused on making it easier to implement chips with more flexibility,” said Mark Papermaster, chief technology officer at AMD during the analyst day meeting late last week.
AMD will allow customers to implement multiple dies — also called chiplets or compute tiles — in a tight chip package. AMD already uses tiles, but is now welcoming third parties to make accelerators or other chips to be included in 2D or 3D packages alongside its x86 CPUs and GPUs.
AMD has been customizing chips for many years, most notably for gaming consoles made by Sony and Microsoft, but using mostly internal technology.
Customers want flexibility to meet specific organizational requirements, and are looking to integrate accelerators inside chips for applications that include AI and automotive, AMD executives said during the show.
“We’re going to make it much easier to add third-party IP as well as customer IP to that chiplet platform,” said Lisa Su, AMD CEO, during the conference.
AMD needed to expand its chiplet platform for technologies it acquired from Xilinx, which makes FPGAs and AI accelerators, and Pensando, which makes networking chips.
“We’ve gotten a lot of positive customer engagement so far when you think about hyperscalers, when you think about 5G in automotive,” Su said, adding, “these are big opportunities where people want to customize and we want to be their partner of choice.”
Most of AMD’s chips , including its flagship PC and server chips, are built with the x86 instruction set architecture, while Pensando and Xilinx are mostly based on Arm. Xilinx and AMD are also members of the open-source RISC-V Foundation, which has created an open-source chip architecture to rival x86 and ARM.
“We’re happy to be ISA-agnostic here,” Su said.
Intel and Nvidia have also shared their chiplet plans. Intel has opened up its x86 architecture for licensing, and has a chiplet strategy where it allows customers to put together Arm and RISC-V cores in a single package.
Intel’s goal is to draw more customers to its Foveros packaging, which will bring more business to its factories. Intel is investing billions to build new factories using its latest nodes.
“I think for a company the size of AMD, it’s really about designing and optimizing their own chiplets to give them a manufacturing advantage, but I’m not sure they really need to support these other standards or use third party chiplets. They can get better stuff designing themselves,” said Linley Gwennap, principal analyst at TechInsights.
AMD is basing its chiplet strategy on TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) packaging technologies, which is also backed by Nvidia and Apple. Nvidia is welcoming third parties to develop cores that connect to its CPUs and GPUs using the proprietary NVLink interconnect.
The chiplet strategy is more relevant to manufacturing, with Intel offering foundry services around its chiplet strategy. AMD and Nvidia are representatives for TSMC in its manufacturing competition with Intel, Gwennap said.
AMD’s custom chip strategy will revolve around the new Infinity Architecture 4.0, which is the interconnect for dies in the chip package. The proprietary Infinity fabric will be compatible with the CXL 2.0 interconnect.
The Infinity interconnect will also support the UCIe (Universal Chiplet Interconnect Express) to link up chiplets in a package. UCIe is already backed by Intel, AMD, Arm, Google, Meta and others.
“The implementation of the fourth-generation AMD Infinity architecture also allows a unified, coherent, shared memory across host and external devices,” Papermaster said.
AMD’s server chips are already being used by Meta and cloud providers that include Amazon, Microsoft and Google. Meta is building its metaverse infrastructure with AMD Epyc chips, and all the major cloud providers are offering virtual machines with AMD’s chips.