Search for: All records

Enhanced Zn anode kinetics and reversibility are achieved at a high ZUR by guiding Zn²⁺plating underlying the SnO_1.17interphase with a regulated (101) orientation, surpassing those achieved by inducing Zn(002) plating overlying the interphase. 

The growth rate of the GPU memory capacity has not been able to keep up with that of the size of large language models (LLMs), hindering the model training process. In particular, activations—the intermediate tensors produced during forward propagation and reused in backward propagation—dominate the GPU memory use. This leads to high training overheads such as expensive weight update costs due to the small micro-batch size. To address this challenge, we propose SSDTrain, an adaptive activation offloading framework to high-capacity NVMe SSDs. SSDTrain reduces GPU memory usage without impacting performance by fully overlapping data transfers with computation. SSDTrain is compatible with popular deep learning frameworks like PyTorch, Megatron, and DeepSpeed, and it employs techniques such as tensor deduplication and forwarding to further enhance efficiency. We extensively experimented with popular LLMs like GPT, BERT, and T5. Results demonstrate that SSDTrain reduces 47% of the activation peak memory usage. At the same time, SSDTrain perfectly overlaps the I/O with the computation and incurs negligible overhead. Compared with keeping activations in GPU memory and layerwise full recomputation, SSDTrain achieves the best memory savings with negligible throughput loss. We further analyze how the reduced activation memory use may be leveraged to increase throughput by increasing micro-batch size and reducing pipeline parallelism bubbles. 

Finite element analysis provides visual insights into conductive path evolution in a SiO₂-based memristor. Electrochemical impedance spectroscopy experimentally validated the theoretical findings by interpreting with an equivalent circuit. 

The overarching goal herein is to identify the factors dominating the performance of a‐IGZO‐based memristors. Despite the highest on/off ratio, greater than 10⁴with a preferred minimal set/reset bias achieved from a‐IGZO‐based memristors, it is observed that the switching performance and stability/reliability of the devices is significantly dominated by theV_O^··density and metallization material, depending on their reactivity with IGZO. As the first governing factor, ensuring optimalV_O^··concentration in the switching layer IGZO (V_O^··/O_O^xratio 24.3% in this study) is crucial to obtain the tractable formation and rupture of conduction filament. Neither higher nor lowerV_O^··density than the optimized results in detrimental reliability issues, which may be ascribed to an uncontrollable filament in an abundant vacancy environment or a weak conducting path, respectively. As the second governing mechanism determining the memristor performance and reliability, it is suggested that metallization materials need to be carefully selected based on the thermodynamic redox potential and interfacial stability of the metallization material with IGZO. Metallization materials with larger reduction potential and interfacial stability are found to yield higher switching on/off ratio and greater device performance reliability.