Table-Lookup MAC: Scalable Processing of Quantised Neural Networks in FPGA Soft Logic

Conceitos Básicos

TLMAC framework optimizes quantised neural networks for scalable lookup-based processing on FPGAs.

Resumo

The content discusses the TLMAC framework for processing quantised neural networks on FPGAs. It introduces the concept of Table Lookup Multiply-Accumulate (TLMAC) to compile and optimize quantised neural networks for scalable lookup-based processing. The framework clusters unique groups of weights to enable highly parallel computation, reducing LUT utilization and routing congestion. TLMAC significantly improves scalability compared to previous methods, allowing implementation of ImageNet-scale models on commercially available FPGAs.

• Introduction to recent advancements in neural network quantization.
• Challenges faced by previous methods in scaling lookup-based computing.
• Introduction of TLMAC as a solution for scalable processing.
• Explanation of the TLMAC process and its benefits.
• Comparison with prior works and demonstration of improved scalability.
• Details on hardware architecture, system-level implementation, place & route algorithms, and experimental results.

Estatísticas

FPGA ’24, March 3–5, 2024, Monterey, CA, USA

Citações

Principais Insights Extraídos De

Table-Lookup MAC

by Daniel Gerli... às arxiv.org 03-19-2024

https://arxiv.org/pdf/2403.11414.pdf

Perguntas Mais Profundas

重みの冗長性が量子化されたニューラルネットワークにおけるルックアップベースの計算効率にどのような影響を与えるか？

重みの冗長性は、TLMAC（Table-Lookup MAC）で実現されるルックアップベースの計算効率に大きな影響を与えます。量子化されたニューラルネットワークでは、同じ重みグループが複数回使用されることが多いため、その一部だけをLUT（Look-Up Table）配列内に保存することでリソース使用効率を向上させることが可能です。これにより、必要な情報へのアクセス時間が短縮され、高速かつ効率的な演算処理が実現されます。