Rachid Karami

Saptarshi Mitra, Yifan Zhang, Rachid Karami et al.

Mixture-of-Agents (MoA) systems improve reasoning accuracy by routing each query to multiple expert LLMs and aggregating their outputs. Efficiently executing this workload on limited GPU resources has bottlenecks. Skill-based routing creates skewed expert demand, and combining instruction-tuned LLMs with long-reasoning models results in extreme variability in generation lengths. Consequently, traditional scheduling strategies suffer from significant GPU idling and throughput collapse due to load imbalances. We present MOSAIC, a scheduling framework to accelerate MoA workloads. First, we formulate an Integer Linear Program (ILP) based scheduler that jointly optimizes expert placement and per-worker prompt assignment from offline-profiled costs, replicating reasoning experts across workers while pinning lightweight ones. Second, MOSAIC uses confidence-aware adaptive aggregation, leveraging inter-expert agreement to bypass the heavy final aggregator LLM for consensus queries. In our 4-GPU system, MOSAIC achieves up to 2.5x expert-stage, 4.23x aggregator-stage and 1.7~2.3x end-to-end speedups over the baseline scheduler, while matching accuracy within 0.1pp.

Saptarshi Mitra, Rachid Karami, Haocheng Xu et al.

The demand for machine intelligence capable of processing continuous, long-context inputs on local devices is growing rapidly. However, the quadratic complexity and memory requirements of traditional Transformer architectures make them inefficient and often unusable for these tasks. This has spurred a paradigm shift towards new architectures like State Space Models (SSMs) and hybrids, which promise near-linear scaling. While most current research focuses on the accuracy and theoretical throughput of these models, a systematic performance characterization on practical consumer hardware is critically needed to guide system-level optimization and unlock new applications. To address this gap, we present a comprehensive, comparative benchmarking of carefully selected Transformer, SSM, and hybrid models specifically for long-context inference on consumer and embedded GPUs. Our analysis reveals that SSMs are not only viable but superior for this domain, capable of processing sequences up to 220K tokens on a 24GB consumer GPU-approximately 4x longer than comparable Transformers. While Transformers may be up to 1.8x faster at short sequences, SSMs demonstrate a dramatic performance inversion, becoming up to 4x faster at very long contexts (~57K tokens). Our operator-level analysis reveals that custom, hardware-aware SSM kernels dominate the inference runtime, accounting for over 55% of latency on edge platforms, identifying them as a primary target for future hardware acceleration. We also provide detailed, device-specific characterization results to guide system co-design for the edge. To foster further research, we will open-source our characterization framework.

Mariam Rakka, Rachid Karami, Ahmed M. Eltawil et al.

Mixed-precision quantization works Neural Networks (NNs) are gaining traction for their efficient realization on the hardware leading to higher throughput and lower energy. In-Memory Computing (IMC) accelerator architectures are offered as alternatives to traditional architectures relying on a data-centric computational paradigm, diminishing the memory wall problem, and scoring high throughput and energy efficiency. These accelerators can support static fixed-precision but are not flexible to support mixed-precision NNs. In this paper, we present BF-IMNA, a bit fluid IMC accelerator for end-to-end Convolutional NN (CNN) inference that is capable of static and dynamic mixed-precision without any hardware reconfiguration overhead at run-time. At the heart of BF-IMNA are Associative Processors (APs), which are bit-serial word-parallel Single Instruction, Multiple Data (SIMD)-like engines. We report the performance of end-to-end inference of ImageNet on AlexNet, VGG16, and ResNet50 on BF-IMNA for different technologies (eNVM and NVM), mixed-precision configurations, and supply voltages. To demonstrate bit fluidity, we implement HAWQ-V3's per-layer mixed-precision configurations for ResNet18 on BF-IMNA using different latency budgets, and results reveal a trade-off between accuracy and Energy-Delay Product (EDP): On one hand, mixed-precision with a high latency constraint achieves the closest accuracy to fixed-precision INT8 and reports a high (worse) EDP compared to fixed-precision INT4. On the other hand, with a low latency constraint, BF-IMNA reports the closest EDP to fixed-precision INT4, with a higher degradation in accuracy compared to fixed-precision INT8. We also show that BF-IMNA with fixed-precision configuration still delivers performance that is comparable to current state-of-the-art accelerators: BF-IMNA achieves $20\%$ higher energy efficiency and $2\%$ higher throughput.

Rachid Karami, Sheng-Chun Kao, Hyoukjun Kwon

Among ML operators today, GEneralMatrix Multiplication (GEMM)-based operators are known to be key operators that build the main backbone of ML models. As their computational overhead dominates the overall execution time (e.g., 42.8% - 96.6% in our results), GEMM operators have been the prime optimization targets for fast ML inference. This led to advanced GPUs and accelerators available today, which provided significant boost in the GEMM performance compared to CPUs, aligned with the lesson from Amdahl's law. However, accelerating GEMM has significantly shifted the Amdahl's law's landscape for ML inference; due to the decreased GEMM execution time, the relative execution time of non-GEMM operators is now significant. Although the importance of non-GEMM performance is increasing, we have little knowledge about the non-GEMM performance horizon in the latest hardware platforms and models. Therefore, to guide non-GEMM-oriented optimizations, we conduct a thorough performance analysis of 17 widely adopted ML models in Hugging Face and Torchvision on workstation and data center platforms with/without GPUs. We discover that non-GEMM performance bottleneck is a considerable issue across all the platforms and models, accounting for 11.3% to 73.6% of total latency, on average. The challenge significantly aggravates when we apply quantization, which is a common model compression technique, due to the boosted GEMM performance and extra non-GEMM operators for dequantization and requantization. To provide insights into non-GEMM optimization targets, we demystify the most dominant non-GEMM operators for each model and deployment software. We also show that widely adopted optimizations such as operator fusion do not completely address the non-GEMM performance bottleneck, where non-GEMM operators still account for 15% to 48% of total latency.

Rachid Karami, Rajeev Patwari, Hyoukjun Kwon et al.

The integration of generative AI models, particularly large language models (LLMs), into real-time multi-model AI applications such as video conferencing and gaming is giving rise to a new class of workloads: real-time generative AI (RTGen). These workloads combine the compute intensity and dynamic execution patterns of generative models with the stringent latency and concurrency constraints of real-time inference. To meet the diverse demands of RTGen workloads, modern edge platforms increasingly adopt heterogeneous system-on-chip (SoC) architectures that integrate CPUs, GPUs, and NPUs. Despite the potential of heterogeneous SoC, the scheduling space complexity and performance implications of RTGen workloads on such platforms remain underexplored. In this work, we perform a comprehensive characterization of RTGen workloads on AMD's latest heterogeneous SoC, Ryzen AI. We construct realistic multi-model scenarios inspired by industry use cases and profile model performance across all available backends. Using this data, we evaluate five scheduling policies and their impact on both real-time metrics (e.g., deadline violation rate) and LLM performance (e.g., time-to-first-token and tokens-per-second). Our results show that scheduling decisions significantly affect workload performance (e.g., leading to a 41.7% difference in deadline violation rates on average), and highlight the need for scheduling strategies that are aware of workload dynamics and hardware heterogeneity. Our findings underscore the importance of workload-aware, dynamic heterogeneous scheduling in enabling high-performance, on-device RTGen applications.