Samsung on Wednesday unveiled an up to date roadmap for its most superior chip nodes catered for AI chips. The South Korean tech large additionally launched a brand new turnkey service that leverages its a number of chip enterprise areas to entice firms reminiscent of Nvidia and AMD to make use of its foundry, or contract chip manufacturing, service for his or her AI chips.
The announcement marks the shifting focus of Samsung Foundry, the corporate’s contract chip-making enterprise unit, to chips for AI and high-performance computing (HPC) quite than processors on cell units.
Samsung Foundry’s AI gross sales have elevated by 80% over the previous 12 months and it was making vital strides in diversifying its buyer base and software areas amid evolving market demand, the corporate famous. The tech large goals to have over 50% of its foundry income introduced in exterior of cell, it additionally mentioned.
At Samsung Foundry Discussion board, its annual occasion for foundry themed this 12 months as Empowering the AI Revolution, held at San Jose, the corporate confirmed off its new 2-nanometer (nm) and 4nm course of nodes, referred to as SF2Z and SF4U, respectively.
SF2Z incorporates what is named a bottom energy supply community (BSPDN) to a traditional 2nm node (Samsung’s SF2), the place the ability rails are positioned on the bottom of the wafer. Present chips have all of the parts in a chip reminiscent of reminiscence, logic, and energy rails on the entrance aspect of the wafer. Different contract chip makers are additionally getting ready their very own BSPDN applied sciences __ Intel calls it PoweVia and TSMC refers it to as Tremendous PowerRail which they’re additionally planning to undertake for his or her 2nm or beneath node chips.
In accordance with Samsung, placing the ability rails on the bottom as an alternative enhances the ability, efficiency, and space (PPA) and voltage drop. SF2Z is geared toward high-performance computing (HPC) and AI chips and can roll out in 2027. Samsung already mentioned beforehand that SF2 will launch in 2025 previous to the discussion board.
Samsung was the primary to begin manufacturing of a 3nm gate-all-around (GAA) node in 2022. On the discussion board, Samsung famous that its GAA course of was maturing by way of efficiency and yield and a second-generation 3nm node referred to as SF3 will launch later this 12 months. GAA can even be adopted to its 2nm node launching subsequent 12 months, the corporate mentioned.
SF4U, in the meantime, is a variant of its 4nm, SF4 course of that gives PPA boosts utilizing optical shrink, the place an present die design by the client is scaled down to suit into the newer node. This protects them in value as no main architectural adjustments within the chip’s design is required for them emigrate into the extra superior node. SF4U is launching in 2025 whereas Samsung already provides SF4 to clients. The South Korean tech large additionally reiterated that its 1.4nm node (SF1.4) will launch in 2027 and was additionally getting ready for chips under 1.4nm.
Samsung additionally unveiled its new turnkey foundry platform dubbed Samsung AI Options. This platform is obtainable collectively by Samsung’s three enterprise models in its chip division, Foundry, Reminiscence, and Superior Package deal (AVP).
In accordance with the corporate, the platform integrates the “distinctive strengths” of every of those enterprise models that may enable Samsung to supply options for purchasers tailor-made for the particular necessities for his or her AI chips. General, it is going to present extra bandwidth in a compact type issue, cut back energy consumption, and enhance sign integrity, Samsung mentioned.
As a result of it’s the solely chipmaker that may supply reminiscence chips to go together with the client’s chip, manufacture these chips in addition to bundle them, Samsung mentioned this streamlines provide chain administration and reduces time to marketplace for the client. The corporate mentioned Samsung AI Options provides a 20% enchancment in complete turnaround time for purchasers. Samsung AI Options will likely be provided with co-packaged optics (CPO) __ the place even the optics are packaged __ in 2027, the South Korean tech large mentioned.
Samsung’s resolution so as to add extra variants to its course of nodes is sensible as a technique for AI, a normal time period that in reality encompasses all kinds of various chips designed for various duties and scales. Some are made for ChatGPT and different massive language fashions. Some are for imaginative and prescient and picture processing to be used in drones and show units. Others are AI accelerators for knowledge facilities.
In accordance with market analysis agency Omdia, the worldwide foundry market is predicted to develop 18.1% a 12 months on common from 2023 when it was value $103.55 billion to 2027, when it’s anticipated to be value $201.28 billion. Progress in cutting-edge nodes of 3nm and beneath can have probably the most marked development at 92.3% per 12 months on common over the time interval, the analysis agency’s forecast famous.
Samsung is the world’s largest reminiscence chip producer and the second-largest contract chipmaker. Its Reminiscence Enterprise is making an attempt to win Nvidia as a buyer for its HBM3E chips, that are used along with the GPU maker’s AI accelerators.
