Uploaded January 2026 | Updated November 2026, 55 minutes ago
Abstract:
Matrix multiplication sits at the core of today’s machine learning workloads, especially in large language models (LLMs) and convolutional neural networks (CNNs). As quantized neural networks become increasingly important for efficient inference, current vector architectures and SIMD units struggle to handle low-precision formats effectively.
In this talk, I will present CAMP (Cartesian Accumulative Matrix Pipeline) — a simple, scalable micro-architecture that brings substantial improvements to matrix multiplication on both ARM SVE and RISC-V SIMD-based systems. CAMP introduces a hybrid multiplier and a hierarchical design that natively supports multi-precision arithmetic (4-bit, 8-bit, and beyond), enabling higher throughput while maintaining energy efficiency.
I will discuss how CAMP accelerates quantized operations, its integration into existing vector datapaths, and the insights gained from synthesis and place-and-route on 7 nm and 22 nm technologies. Finally, I will show evaluation results on a range of LLM and CNN workloads, where CAMP delivers up to 17× and 23× performance gains over ARM A64FX and a RISC-V edge SoC, respectively — all with modest hardware overhead.
This session will provide an accessible overview of the architecture, key design decisions, and the practical implications of CAMP for future low-power AI systems.
Abstract:
Matrix multiplication sits at the core of today’s machine learning workloads, especially in large language models (LLMs) and convolutional neural networks (CNNs). As quantized neural networks become increasingly important for efficient inference, current vector architectures and SIMD units struggle to handle low-precision formats effectively.
In this talk, I will present CAMP (Cartesian Accumulative Matrix Pipeline) — a simple, scalable micro-architecture that brings substantial improvements to matrix multiplication on both ARM SVE and RISC-V SIMD-based systems. CAMP introduces a hybrid multiplier and a hierarchical design that natively supports multi-precision arithmetic (4-bit, 8-bit, and beyond), enabling higher throughput while maintaining energy efficiency.
I will discuss how CAMP accelerates quantized operations, its integration into existing vector datapaths, and the insights gained from synthesis and place-and-route on 7 nm and 22 nm technologies. Finally, I will show evaluation results on a range of LLM and CNN workloads, where CAMP delivers up to 17× and 23× performance gains over ARM A64FX and a RISC-V edge SoC, respectively — all with modest hardware overhead.
This session will provide an accessible overview of the architecture, key design decisions, and the practical implications of CAMP for future low-power AI systems.










