Himax (Taiwan)

Large language models are limited in deployment by GPU memory and inference latency. We present Minima, a production compression pipeline that learns where and how to structurally compress a Transformer and turns that compression into real serving gains. Minima trains a lightweight convolutional predictor to estimate layer- and patch-level sensitivity, applies a mixture of Tucker, tensor-train, and tensor-ring decompositions to low-sensitivity regions, performs a short healing fine-tune, and executes the resulting operators with custom Triton and CUDA kernels. The reduced memory footprint enables speculative decoding with a small draft model and a larger verifier. On Qwen3-32B at an 8k-token context window, Minima reduces peak VRAM from 64 GiB to 40 GiB. For a single active request, throughput increases from 40 tokens per second (baseline) to 50 tokens per second (Minima) and 75 tokens per second (Minima with speculative decoding). Under 50 parallel requests, throughput is 34, 44, and 53 tokens per second respectively, showing that Minima remains effective under high concurrency even when speculative decoding gains compress. We position Minima relative to recent tensor-network, low-rank plus quantization, and cross-layer sharing methods, and argue that it is a practical step toward more aggressive structural compression via shared tensor backbones with tiny per-layer adapters.

Abstract Oxidation is a key failure mechanism for environmental barrier coatings (EBCs) deposited on ceramic matrix composites (CMCs). Current generation EBC/CMC systems incorporate a silicon (Si) bond coat between the CMC substrate and the EBC topcoat. Oxidation occurs at the interface between the Si bond coat and the EBC topcoat due to diffusion of water vapor and oxygen creating a thermally grown silica (SiO2) scale. As the SiO2 scale increases in thickness, the adhesion strength of the EBC topcoat becomes compromised leading to delamination and spallation. This work investigates the oxidation kinetics of an EBC/CMC system with an Si bond coat utilizing a natural gas/oxygen (NG/O2) burner rig at atmospheric pressure. An oxidation model is presented to predict the oxidation kinetics of the EBC/CMC system. The model exhibits good agreement with experimental data obtained in the NG/O2 rig up to 250 hours after which the oxidation kinetics deviate from the predicted behavior. A reason for the observed change in oxidation kinetics is presented along with directions for future work.

You hear a song but can’t quite place it. You pull out your phone and use a music recognition app to identify the song immediately. That process is hashing a short snippet of signal and identifying where it comes from in a database of hashes from millions of songs in seconds. Can that approach be adapted to help with picking and checking tops on well logs? The short answer is yes! HXMX Tops Fingerprinting, short for Hash Match Tops Fingerprinting, does exactly that by matching hashes with those in a database to pick tops on well logs.This abstract introduces Tops Fingerprinting, outlining its potential to revolutionize the traditional methods of well log analysis. Not only does this approach promise to significantly reduce the time required for picking tops—from hours to mere seconds—but it also offers objective and quantifiable definitions of tops, identifies missing sections, and provides comparative metrics between wells within a given interval. Such advancements hold immense promise for enhancing the efficiency, accuracy, and richness of geological analysis.This presentation will share proof-of-concept work done on DJ Basin well logs, in collaboration with the Innov8X program at the Colorado School of Mines.

As the current of integrated circuits (ICs) gradually increase within the high-performance application, power consumption also increases, which in turn raises the overall temperature of the ICs. To make sure these high-power consumption chips could operate within a certain temperature range. The thermal managements on chips are becoming increasingly important. This research aims to discuss the thermal impact of Chip-on-Film (COF) packages on the temperature of chips. This study develops a simulation model to simulate the junction temperature of IC and analyze the temperature distribution profile in COF package. The complete mechanical COF package design is imported into a simulation software for thermal analysis. Each mechanical COF package design corresponds to varying Inner Lead Bonding (ILB) trace design. It means that each IC in different application product requirements corresponds to a different ILB trace design on each COF package. This study implements the different ILB layout designs on COF package and transfers these ILB trace layout from two-dimension file into three- dimensional simulation model. By generating three-dimensional structures, they can be imported into ANSYS thermal software to simulate the junction temperature of IC and temperature profile on total COF package for each ILB trace design. The heat transfer is enhanced by optimizing the ILB trace routing on the COF package to reduce the IC temperature. Furthermore, infrared thermometers will be used to measure the temperature of IC on different COFs, and the results will be compared with the simulation results. Based on results, the design of different trace routing on the COF package has an impact on IC temperature. The optimized ILB trace design can effectively reduce the junction temperature of IC. This study shows the results that the design of ILB trace contact(B) is more helpful than the design of ILB trace w/o contact(A) for improving IC temperature performance. In addition to optimizing the design of ILB traces, attaching heatsink on the COF package is also an effective method for reducing junction temperature of IC. In the current application, the heatsinks could be attached to the top side and to the bottom side of COF package. The size of the heatsinks are vary depending on the specific application requirements. By using the same method, this study also generates two different sizes of heatsinks in common within dimensions of 26mmx12mm and 30mmx15mm and simulate the junction temperature of IC to evaluate to temperature reduce on IC. These two heatsinks are attached separately to the top side and bottom side in simulation model. By developing the three-dimensional model in thermal analysis, we could get the junction n temperature of IC at first and evaluate the various feasibility of thermal management.

Abstract Selenium (Se) is an essential micromineral that is incorporated into the active center of selenoproteins which play important roles in immune regulation, thyroid function, antioxidant defense, and glucose homeostasis. Se deficiency is commonly observed in patients of chronic illness, cancer, and severe trauma; Se supplementation to this group has been shown to provide clinical benefits including reduced mortality rates and better prognosis. However, there is a lack of consensus to determine Se adequacy and deficiency in terms of health and disease in the context of the local Taiwanese population. To obtain the normal Se range of the general adult Taiwanese by which to assess an individual’s Se status, studies reporting on plasma and serum Se values in Taiwanese adults from databases and available data from government-accredited clinical laboratories were gathered. After assessment for suitability against predefined criteria including sample size, subject characteristics, method of Se measurement, Se levels, and year of publication, six sources were included for subsequent mathematical calculations. The combined mean Se levels from a total of 5,508 subjects were 182.8 ± 124.3 µg/L. The range was estimated to be 58–307 µg/L, and Se levels between 100–270 µg/L were considered the normal range in Taiwanese adults, whereas Se levels < 100 µg/L were suggestive of deficiency and Se supplementation was recommended.

After tooth extraction, alveolar ridge loss due to resorption is almost inevitable. Most of this bone loss occurs during the first six months after the extraction procedure. Many studies have indicated that applying socket-filling biomaterials after extraction can effectively reduce the resorption rate of the alveolar ridge. The purpose of this study was to investigate the clinical efficacy of the application of a hydroxyapatite/β-tricalcium plus collagen (HA/β-TCP + collagen) dental bone graft in dental sockets immediately after tooth extraction, so as to prevent socket resorption. The study was conducted on 57 extraction sockets located in the mandible and maxilla posterior regions in 51 patients. HA/β-TCP + collagen was inserted into all of the dental sockets immediately after extraction, and was covered with a flap. Follow-up was performed for three months after extraction, using radiographs and stents for the vertical and horizontal alveolar ridge measurements. A minimal alveolar bone width reduction of 1.03 ± 2.43 mm (p < 0.05) was observed. The height reduction showed a slight decrease to 0.62 ± 1.46 mm (p < 0.05). Radiographically, the bone height was maintained after three months, indicating a good HA/β-TCP + collagen graft performance in preserving alveolar bone. In conclusion, the HA/β-TCP + collagen graft demonstrated adequate safety and efficacy in dental socket preservation following tooth extraction.

Texture compression is an important technique to reduce memory storage and increase rendering speed in graphic processing units (GPUs). To ensure fast decoding and retain random memory access characteristics, the industry standard DXTC compression finds an approximate line segment in color space for each 4×4 block. The idea has been extended to multiline solutions by segmenting the 4×4 block into several partitions and adopted by the new format BPTC. BPTC can provide the highest quality than other texture compression standards. However, the limited partition combinations used in BPTC would smooth the texture detail in some cases. This work introduces an adaptive color grouping and selection algorithm to relax the texture smoothing issue. Simulation results show that the proposed algorithm can improve the average PSNR by 0.78 dB for those blocks.