At the SAFE Forum (Samsung Advanced Foundry Ecosystem), Samsung outlined its roadmap for future semiconductor nodes. The company previously introduced the Exynos 2600, the first smartphone chip built on the 2nm SF2 node.
Following this, Samsung plans to launch the SF2P node, which aims to reduce power consumption by 26% while increasing clock speeds by 15%. However, the node itself plays only a partial role in achieving these improvements. According to Shin Jong-shin, VP of the Design Platform Development Team at Samsung Foundry, over half of these enhancements are attributed to DTCO.
DTCO, or Design-Technology Co-Optimization, is an innovative approach that integrates manufacturing methodologies with chip design. Traditionally viewed as separate processes, combining them allows for more optimized designs, resulting in faster performance, reduced power usage, and potential cost savings.
Following SF2P, the SF2P+ node is expected to enter mass production between 2027 and 2028. Additionally, Samsung is developing the SF2X node, which will be tailored for artificial intelligence hardware.

Samsung is also focused on increasing the amount of SRAM on its chips. SRAM provides the registers and caches for CPUs and GPUs. While SRAM is significantly faster, it occupies more space than DRAM; for instance, a standard SRAM cell, which stores one bit, requires six transistors, whereas a DRAM cell utilizes only one.
The Nvidia Rubin GPU, produced using TSMC’s N3 node, includes 128MB of on-die SRAM. Meanwhile, Samsung and Groq are collaborating on a large language model accelerator featuring over 500MB of SRAM, manufactured on the older 4nm node. Additionally, the Samsung-supported Rebellions showcased the REBEL-100 AI accelerator, currently being produced on Samsung’s 4nm technology.
Post-2nm, Samsung plans to transition to 1.4nm nodes, starting with the SF1.4 node set for mass production in 2029. This will be succeeded by the SF1.4+ node in 2030. Initially, SF1.4 was targeted for 2027, but the timeline has since been pushed back.
Source | Via