Samsung on Wednesday unveiled an updated roadmap for its most advanced chip nodes meant for AI chips. The South Korean tech giant also introduced a new turnkey service that leverages its multiple chip business areas to entice companies such as Nvidia and AMD to use its foundry for their AI chips.
The announcement marks the shifting focus of Samsung Foundry, the company's contract chip-making business unit, from processors for mobile devices to chips for AI and high-performance computing (HPC).
Also: Samsung's chip division wants the top spot back: What's happening?
Samsung Foundry's AI sales have increased by 80% over the past year, and it has made significant strides in diversifying its customer base and application areas amid evolving market demand, the company noted. The tech giant aims to bring in over 50% of its foundry revenue from outside of mobile, the company also said.
At Samsung Foundry Forum, Samsung's annual foundry event themed this year as "Empowering the AI Revolution," held in San Jose, the company showed off its new 2-nanometer (nm) and 4nm process nodes, called SF2Z and SF4U, respectively.
SF2Z incorporates what is called a backside power delivery network (BSPDN) to a conventional 2nm node (Samsung's SF2), where the power rails are placed on the backside of the wafer. Current chips have all the components in a chip such as memory, logic, and power rails on the front side of the wafer. Other contract chipmakers are also preparing their own BSPDN technologies -- Intel has PowerVia and TSMC has Super PowerRail -- that they are also planning to adopt for their 2nm or under node chips.
Also: Samsung to form HVAC joint venture with Lennox to expand sales in North America
According to Samsung, putting the power rails on the backside enhances the power, performance, and area (PPA) and voltage drop. SF2Z is aimed at HPC and AI chips and will roll out in 2027. Samsung has previously said that SF2 will launch in 2025.
Samsung was the first to start production of a 3nm gate-all-around (GAA) node in 2022. At the forum, Samsung noted that its GAA process was maturing in terms of performance and yield and a second-generation 3nm node called SF3 would launch later this year. The company plans to adopt GAA to its 2nm node launching next year.
SF4U, meanwhile, is a variant of Samsung's 4nm, SF4 process that offers PPA boosts using optical shrink, where an existing die designed by the customer is scaled down to fit into the newer node. This saves money, as no major architectural changes in the chip's design is required to migrate into the more advanced node. SF4U is launching in 2025, while Samsung already offers SF4 to customers. The South Korean tech giant also reiterated that its 1.4nm node (SF1.4) will launch in 2027.
Samsung also unveiled its new turnkey foundry platform, Samsung AI Solutions. This platform is offered together by Samsung's three business units in its chip division, Foundry, Memory, and Advanced Package (AVP).
Also: Samsung Galaxy Ring: Features, price, availability, and everything else we know
According to the company, the platform integrates the "unique strengths" of each of these business units that will let Samsung offer solutions for customers tailored to the specific requirements for their AI chips. Overall, it will provide more bandwidth in a compact form factor, reduce power consumption, and improve signal integrity, Samsung said.
Because Samsung is the only chipmaker that can offer memory chips to go with the customer's chip, manufacture the chips, and package them, the company said its offering streamlines supply chain management and reduces time to market for the customer. The company said Samsung AI Solutions offers a 20% improvement in total turnaround time for customers. Samsung AI Solutions will be offered with co-packaged optics (CPO) in 2027, the South Korean tech giant said.
Samsung's decision to add more variants to its process nodes makes sense for AI, a general term that in fact encompasses a wide variety of chips designed for different tasks and scales. Some are made for ChatGPT and other large language models. Some are for vision and image processing for use in drones and display devices. Others are AI accelerators for data centers.
According to market research firm Omdia, the global foundry market is expected to grow 18.1% per year on average from 2023, when it was worth$103.55 billion, to 2027, when it is expected to be worth$201.28 billion. Growth in cutting-edge nodes of 3nm and under will have the most marked growth at 92.3% per year on average over the time period, the research firm forecasted.
Samsung is the world's largest memory chip producer and the second-largest contract chipmaker. Its Memory Business is attempting to win Nvidia as a customer for its HBM3E chips, which are used together with the GPU maker's AI accelerators.