Since 2020, TSMC's SRAM memory has stopped shrinking. With the A14 node, due in production in 2028, the foundry's 6T memory cell is expected to fall to about 0.017 µm², a 19% reduction from the roughly 0.021 µm² that has held since the 5-nanometre N5 process. After about eight years of stall, memory would shrink again at a pace closer to logic.
The problem has been known for some time. While transistor density in logic circuits has grown with each generation, SRAM, the fast memory used as cache inside chips, has not kept up. As a result, memory takes up a growing share of the silicon in many AI accelerators: in Nvidia's H100, SRAM reaches about 10% of the die area, rising to between 12% and 15% with the L1 cache of the multiprocessors. A smaller memory cell therefore feeds directly into the cost and density of AI chips.
Image: TSMC — https://esg.tsmc.com/en-US/articles/366
According to the technical sources cited, TSMC will present at the IEDM conference the smallest 6T cell yet, below 0.017 µm² on the 2028 node. SemiAnalysis describes an SRAM about 25% denser with A14, a figure that captures the recovery after years of stagnation, while noting that designs use more and more SRAM as memory bottlenecks grow.
On timing, TSMC positions A14 as its next leading-edge logic process, delivering full-node power, performance and area scaling, aimed at AI computing and at improving on-device AI in smartphones. The company says development is progressing smoothly, with yields ahead of schedule. Industry reports suggest pilot production is accelerating and that volume production could arrive in the first half of 2028 rather than the second.
Image: TSMC — https://www.tsmc.com/english
The first beneficiaries will be the large customers that have always accessed TSMC's most advanced nodes. According to industry reports, Apple is expected to be the first customer, likely with A-series chips. For AI chipmakers the news matters more than the size reduction alone: a smaller cell means more cache in the same silicon area, which translates into compute capacity and cost per unit.
The next test is whether the announced shrink holds up in mass production, when the process runs on real volumes rather than in the lab.
No comments yet. Be the first!