Samsung introduces new turnkey foundry service, advanced nodes for AI chips


Samsung on Wednesday unveiled an updated roadmap for its most advanced chip nodes catered 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, or contract chip production, service for their AI chips.

The announcement marks the shifting focus of Samsung Foundry, the company’s contract chip-making business unit, to chips for AI and high-performance computing (HPC) rather than processors on mobile devices.

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 was making significant strides in diversifying its customer base and application areas amid evolving market demand, the company noted. The tech giant aims to have over 50% of its foundry revenue brought in outside of mobile, it also said.

At Samsung Foundry Forum, its annual event for foundry themed this year as Empowering the AI Revolution, held at 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 chip makers are also preparing their own BSPDN technologies __ Intel calls it PoweVia and TSMC refers it to as Super PowerRail which 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 instead enhances the power, performance, and area (PPA) and voltage drop. SF2Z is aimed at high-performance computing (HPC) and AI chips and will roll out in 2027. Samsung already said previously that SF2 will launch in 2025 prior to the forum.

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 will launch later this year. GAA will also be adopted to its 2nm node launching next year, the company said.


Samsung is adding various variants to its 2nm node, with a particular focus on AI. Image: Samsung

SF4U, meanwhile, is a variant of its 4nm, SF4 process that offers PPA boosts using optical shrink, where an existing die design by the customer is scaled down to fit into the newer node. This saves them in cost as no major architectural changes in the chip’s design is required for them 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 and was also preparing for chips below 1.4nm.

Samsung also unveiled its new turnkey foundry platform dubbed 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 allow Samsung to offer solutions for customers tailored for 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 it is the only chipmaker that can offer memory chips to go with the customer’s chip, manufacture these chips as well as package them, Samsung said this 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) __ where even the optics are packaged __ in 2027, the South Korean tech giant said.

Samsung’s decision to add more variants to its process nodes makes sense as a strategy for AI, a general term that in fact encompasses a wide variety of different 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% a 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’s forecast noted.

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.

Source link