Marina Zapater

Marina Zapater, José L. Risco-Martín, Patricia Arroba et al.

Data Centers are huge power consumers, both because of the energy required for computation and the cooling needed to keep servers below thermal redlining. The most common technique to minimize cooling costs is increasing data room temperature. However, to avoid reliability issues, and to enhance energy efficiency, there is a need to predict the temperature attained by servers under variable cooling setups. Due to the complex thermal dynamics of data rooms, accurate runtime data center temperature prediction has remained as an important challenge. By using Gramatical Evolution techniques, this paper presents a methodology for the generation of temperature models for data centers and the runtime prediction of CPU and inlet temperature under variable cooling setups. As opposed to time costly Computational Fluid Dynamics techniques, our models do not need specific knowledge about the problem, can be used in arbitrary data centers, re-trained if conditions change and have negligible overhead during runtime prediction. Our models have been trained and tested by using traces from real Data Center scenarios. Our results show how we can fully predict the temperature of the servers in a data rooms, with prediction errors below 2 C and 0.5 C in CPU and server inlet temperature respectively.

Qunyou Liu, Pengbo Yu, Marina Zapater et al.

Deep neural networks (DNNs) are essential for performing advanced tasks on edge or mobile devices, yet their deployment is often hindered by severe resource constraints, including limited memory, energy, and computational power. While uniform quantization provides a straightforward approach to compress model and reduce hardware requirement, it fails to fully leverage the varying robustness across layers, and often lead to accuracy degradation or suboptimal resource usage, particularly at low bitwidths. In contrast, heterogeneous quantization, which allocates different bitwidths to individual layers, can mitigate these drawbacks. Nonetheless, current heterogeneous quantization methods either needs huge brute-force design space search or lacks the adaptability to meet different hardware conditions, such as memory size, energy budget, and latency requirement. Filling these gaps, this work introduces \textbf{\textit{SigmaQuant}}, an adaptive layer-wise heterogeneous quantization framework designed to efficiently balance accuracy and resource usage for varied edge environments without exhaustive search.

Qunyou Liu, Marina Zapater, David Atienza

Transformers have revolutionized AI in natural language processing and computer vision, but their large computation and memory demands pose major challenges for hardware acceleration. In practice, end-to-end throughput is often limited by paged data movement and interconnect bandwidth rather than raw MAC count. This work proposes a unified system-accelerator co-design approach for transformer inference that jointly optimizes a matrix accelerator and its system integration through paged streaming dataflows and explicit overlap of compute and transfer. On the hardware side, we introduce MatrixFlow, a loosely coupled 16x16 systolic-array accelerator with a page-aligned block matrix multiplication method using 4 KB tiles, a small on-chip buffer of about 20 KB, and a pipelined schedule of DMA, compute, and DMA-out to utilize interconnect bandwidth efficiently. On the system side, we develop Gem5-AcceSys, an extension of the gem5 full-system simulator that explores standard interconnects such as PCIe and configurable memory hierarchies including Direct Memory, Direct Cache, and Device Memory modes with SMMU/TLB effects. We evaluate the co-design using gem5 simulations on representative transformer models including BERT and ViT across multiple data types and system setups. Results show up to 22x end-to-end speedup over a CPU-only baseline and 5x to 8x gains over state-of-the-art loosely and tightly coupled accelerators. We further show that a standard PCIe-based host-memory design can achieve about 80 percent of the performance of on-device HBM. Overall, paged streaming and pipeline overlap, rather than large local SRAMs, are the most effective levers for efficient transformer inference under realistic system constraints.

Karan Pathak, David Atienza, Marina Zapater

The growing demands in the training and inference of Large Language Models (LLMs) are accelerating the adoption of scale-up systems that extend server shared memory through the use of Compute Express Link (CXL)-based load/store interconnects. Accurate full-system simulation of such architectures remains challenging, as existing tools (all very recent) rely on simplified or non-compliant architectural models, impacting accuracy and usability. We present CXLRAMSim, the first gem5-integrated, full-system simulator that models CXL devices at their correct position on the I/O bus, enabling the use of unmodified Linux kernels and software stack, realistic latency-bandwidth behavior and true interleaving with system DRAM. Our approach provides high-fidelity CXL.mem characterization and captures key challenges such as cache pollution when accessing CXL memory.

Valentin Gabeff, Tomas Teijeiro, Marina Zapater et al.

While Deep Learning (DL) is often considered the state-of-the art for Artificial Intelligence-based medical decision support, it remains sparsely implemented in clinical practice and poorly trusted by clinicians due to insufficient interpretability of neural network models. We have tackled this issue by developing interpretable DL models in the context of online detection of epileptic seizure, based on EEG signal. This has conditioned the preparation of the input signals, the network architecture, and the post-processing of the output in line with the domain knowledge. Specifically, we focused the discussion on three main aspects: 1) how to aggregate the classification results on signal segments provided by the DL model into a larger time scale, at the seizure-level; 2) what are the relevant frequency patterns learned in the first convolutional layer of different models, and their relation with the delta, theta, alpha, beta and gamma frequency bands on which the visual interpretation of EEG is based; and 3) the identification of the signal waveforms with larger contribution towards the ictal class, according to the activation differences highlighted using the DeepLIFT method. Results show that the kernel size in the first layer determines the interpretability of the extracted features and the sensitivity of the trained models, even though the final performance is very similar after post-processing. Also, we found that amplitude is the main feature leading to an ictal prediction, suggesting that a larger patient population would be required to learn more complex frequency patterns. Still, our methodology was successfully able to generalize patient inter-variability for the majority of the studied population with a classification F1-score of 0.873 and detecting 90% of the seizures.