Data centres powering modern computing could consume up to 8% of global electricity by 2030, and cooling their chips to liquid nitrogen temperature (77 Kelvin) could slash power use more than sevenfold—but only if engineers can solve a new problem that emerges in the cold. At cryogenic temperatures, transistors switching at extremely low voltages become unreliable. A single atomic defect can randomly flip a transistor on or off, causing logic faults. Existing reliability models assume many defects average out, but at low voltage and low temperature, individual defects dominate probabilistically. This forces designers to use overly conservative safety margins that waste power, blocking the 25-fold performance-to-power improvement that DARPA has targeted for cryogenic computing. This project will develop a new probabilistic modelling framework—called Power-Fault Co-Optimisation—that treats faults as rates per million operations rather than worst-case guarantees. By measuring individual defects at cryogenic temperatures, building compact physical models of their probabilistic behaviour, and integrating those models into commercial chip-design tools, the team aims to identify the lowest safe operating voltage for each circuit. If successful, the approach could unlock ultra-efficient cryogenic processors for data centres, cutting their energy footprint dramatically. The same probabilistic design methodology could also improve low-power electronics for satellites, medical implants, and other applications where heat and energy are tightly constrained.
View original technical description
High performance computing (HPC) and corresponding CMOS chips used in large data centres today face two major challenges: (1) Energy: Achieving a low-energy, low-carbon, greener footprint is critical. It has been predicted that total energy consumed by data centres could increase 15 times by 2030, accounting for up to 8% of global electricity demand. It is crucial to develop new technologies that can drastically reduce HPC energy consumption by orders of magnitude. (2) Performance: Due to the difficulties in continuing the conventional scaling, particularly for power reduction, HPC improvement has dramatically slowed from 32 times per decade in the 1990s to only 1.5 times per decade, as up to 80% of transistors on the chip cannot be used simultaneously due to chip overheating. Power efficiency of logic technologies is limited by VDD, which cannot be scaled much below 0.8 V at room temperature, due to the large Vth and off-leakage imposed by the large sub-threshold swing (SS) of MOSFETs. Liquid nitrogen cooling can drastically improve the performance-to-energy ratio for greener data centres. The significantly reduced SS at cryogenic temperatures (CT) =77 K enables significant Vth and VDD scaling to ~0.1 V and 0.3 V, respectively, and delivers >7.5 times power reduction. DARPA recently called for ground-breaking innovations in cryogenic logic technologies and set the target of 25 times performance-to-power ratio (PPR) improvement. However, nanoscale devices operating at a low VDD near Vth are susceptible to logic and timing faults from false transistor switching on/off, induced by charging and discharging of a single defect. The risk of false switching is amplified by the steeper I-V characteristic at CT. Existing device fault and reliability modelling cannot be used under low VDD at CT where the probabilistic process dominates, because they are based on conventional deterministic methodologies. These models use the accumulation or average of multiple defects to set the failure criteria. They are overly pessimistic and force the designers to use the worst-case methodology and unnecessarily large design margins, which hinders the achievement of the 25X Cryo-CMOS optimization targets. Therefore, a novel probabilistic criterion of Fault Rate in Per Million Operations (FPMO) should be used to identify the lowest VDD at CT, instead. This project aims to break the bottleneck and fill in the knowledge gap of cryogenic power fault co-optimisation (PFCO) under ultra-low-VDD. SPICE-level compact modelling of probabilistic and quantum phenomena in defects and their correlations with random faults in devices and circuits will be developed, for achieving ultra-low power and ultra-low fault cryogenic HPC. The project is built on the holistic integration of three key research pillars and objectives to demonstrate the probabilistic PFCO technology and EDA tools at defects, devices, circuits, and system modelling levels: 1) Measure and characterise different probabilistic defects/faults at CT. 2) Develop probabilistic physical and compact device/fault modelling. 3) Demonstrate the new probabilistic Power and Fault Co-optimisation strategy. The integrated pursuit of the four WPs aims at delivering a major breakthrough deployable for HPC at low VDD and CT, through probabilistic fault modelling and Device-Technology-System co-optimisation for achieving unprecedented PPR, which is transformable for many other applications in low power design and green ICT technologies. The results will be implemented in cryogenic process-development-kits (PDK) and commercial design tools, and disseminated to leading industry and academia partners through collaborations with ARM, IBM, Synopsys, IMEC and Semiwise.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know