The Korea Herald


AMD’s next-gen server CPU standard signals ties with Samsung

By Son Ji-hyoung

Published : June 13, 2022 - 15:00

    • Link copied

A visual concept image of a processor chip (GettyImages) A visual concept image of a processor chip (GettyImages)
The latest announcement of US logic chip designer Advanced Micro Devices’ next-generational processor standard for data center servers indicates that a collaboration with South Korean tech giant Samsung Electronics is imminent.

This comes as AMD’s chip architecture in the making is to be compatible with the groundbreaking Samsung-backed memory standard Compute Express Link, in order for the chip platform to go full throttle.

AMD hinted at heterogeneous computing by opening up to industry standards such as CXL for its latest fourth-generation Infinity Architecture, one of the pillars comprising AMD’s CDNA 3 architecture-based data center server products.

The CDNA 3 architecture, which AMD aims to launch in 2023, is touted as showing five times stronger performance than the current CDNA 2 standard, making the server chipset more suitable for processing massive data workloads for the metaverse and artificial intelligence.

AMD Chief Technology Officer Mark Papermaster said during AMD‘s Financial Analyst Day last week that the new data center chip architecture “allows a unified, coherent, shared memory across a host of external devices” to pursue efficiency in data processing, adding the architecture will support CXL 2.0 and is extensible to CXL 3.0 standard.

This draws attention to a range of chip players in the CXL ecosystem for heterogeneous memory systems, including Samsung Electronics.

The heterogeneous memory system is considered an alternative to the conventional local memory-based system, which Samsung claimed to have limited memory capacity scaling to less than 10 terabytes. The CXL standard adopts the concept of chip resource disaggregation and pooling, so that a processor can have access to a memory pool that enables data flow using not only local memory but also remote memory.

Samsung has unveiled memory modules that allow an expanded short-term memory capacity to be shared and interconnected with processor chips, teaming up with memory expander controller maker Montage Technologies and other CXL consortium partners.

Samsung’s announcement has shown that it is ready to deploy the most advanced CXL-based memory devices, the 512-gigabyte CXL Memory eXpander 2.0, that are poised for sampling in the third quarter this year.

The upcoming product is believed to quadruple memory capacity and lower the system latency by 80 percent compared to what Samsung had unveiled for the first time in the industry last year.

“The semiconductor industry will definitely need a greater memory scalability and capacity,” said Kim Yang-paeng, a researcher at the Korea Institute for Industrial Economics and Trade.

“It is great to see the next-generational architecture coming forward with the industry still in the works to adopt the most-advanced double data rate 5 memory standard (for DRAMs).”

This comes in line with Samsung’s increasing focus on memory solutions targeting enterprise server clients amid server chip supply shortages, contributing to its quarterly sales to an all-time high for the three straight quarters. Shinhan Investment analyst Choi Do-yeon noted that the server chip supply bottleneck has persisted since the third quarter of 2021, dwarfing macroeconomic uncertainties in the memory business.

Samsung declined to comment on the road map for the launch of its memory module product with AMD.