TSMC has revealed some extra particulars about its upcoming N2 and N2P course of know-how at its European Expertise Symposium 2023. Each manufacturing nodes are being developed with high-performance computing (HPC) in thoughts, so, they function plenty of enhancements designed particularly to enhance efficiency. In the meantime, given the performance-efficiency focus that almost all chips purpose to enhance upon, low-power purposes may also make the most of TSMC’s N2 nodes as they’ll naturally enhance performance-per-watt in comparison with predecessors.
“N2 is a good match for the vitality environment friendly computing paradigm that we’re in at the moment,” stated Yujun Li, TSMC’s director of enterprise improvement who’s answerable for the foundry’s Excessive Efficiency Computing Enterprise Division, on the firm’s European Expertise Symposium 2023. “The velocity and energy benefits of N2 over N3 over your entire voltage provide ranges as proven may be very constant, making it appropriate for each low-power and high-performance purposes on the identical time.”
TSMC’s N2 manufacturing node — the foundry’s first manufacturing nodes to make use of nanosheet gate-all-around (GAAFET) transistors — guarantees to extend transistor efficiency by 10-15% on the identical energy and complexity, or decrease energy utilization by 25-30% on the identical clock velocity and transistor rely. Energy supply is without doubt one of the nook stones in the case of enhancing transistor efficiency and TSMC’s N2 and N2P manufacturing processes introduce a number of interconnects-related improvements to squeeze some extra efficiency. Moreover, N2P brings in bottom energy rail to optimize energy supply and die space.
Preventing Resistance
One of many improvements that N2 brings to the desk is super-high-performance metal-insulator-metal (SHPMIM) capacitor to reinforce energy provide stability and facilitate on-chip decoupling. TSMC says that the brand new SHPMIM capacitor provides over 2X increased capability density in comparison with its super-high-density metal-insulator-metal (SHDMIM) capacitor launched a number of years in the past for HPC (which increased capacitance by 4X when in comparison with previous-generation HDMIM). The brand new SHPMIM additionally reduces Rs sheet resistance (Ohm/sq.) by 50% in comparison with SHDMIM in addition to Rc by way of resistance by 50% in comparison with SHDMIM.
Yet one more solution to scale back resistance within the energy supply community has been to rearchitect the redistribution layer (RDL). Ranging from its N2 course of know-how, TSMC will use a copper RDL as an alternative of at the moment’s aluminum RDL. A copper RDL will present an identical RDL pitch, however will scale back sheet resistance by 30% in addition to minimize down by way of resistance by 60%.
Each SHPMIM and Cu RDL are elements of TSMC’s N2 know-how that’s projected for use for top quantity manufacturing (HVM) within the second half 2025 (presumably very late in 2025).
Decoupling Energy and I/O Wiring
Using a bottom energy supply community (PDN) is a one more main enchancment that might be featured by N2P. Common benefits of bottom energy rail are well-known: by separating I/O and energy wiring by shifting energy rails to the again, it’s doable to make energy wires thicker and subsequently scale back by way of resistances within the back-end-of-line (BEOL), which guarantees to enhance efficiency and minimize down energy consumption. Additionally, decoupling I/O and energy wires permits to shrink logic space, which implies decrease prices.
At its Expertise Symposium 2023 the corporate revealed that bottom PDN of its N2P will allow 10% to 12% increased efficiency by decreasing IR droop and enhancing signaling, in addition to decreasing the logic space by 10% to fifteen%. Now, in fact, such benefits might be extra apparent in high-performance CPUs and GPUs which have dense energy supply community and subsequently shifting it to the again makes an ideal sense for them.
Bottom PDN is part of TSMC’s N2P fabrication know-how that may enter HVM in late 2026 or early 2027.
