20.500.12556/DiRROS-18802 Towards deploying highly quantized neural networks on FPGA using chisel We present chisel4ml, a Chisel-based tool that generates hardware for highly quantized neural networks described in QKeras. Such networks typically use parameters with bitwidths less than 8 bits and may have pruned connections. Chisel4ml can generate the highly quantized neural network as a single combinational circuit with pipeline registers in between the different layers. It supports heterogeneous quantization where each layer can have a different precision. The full parallelization enables very low-latency and high throughput inference, that are required for certain tasks. We illustrate this on the triggering system for the CERN Large Hadron Collider, which filters out events of interest and sends them on for further processing. We compare our tool against hls4ml, a high-level synthesis based approach for deploying similar neural networks. Chisel4ml is still under development. However, it already achieves comparable results to hls4ml for some neural network architectures. Chisel4ml is available on https://github.com/cs-jsi/chisel4ml. neural networks QKeras Chisel4ml true false true IEEE ZDA Angleški jezik Ni določen ©2023 IEEE Neznano 2024-04-23 09:07:58 2024-04-23 09:07:58 2024-04-24 03:39:27 0000-00-00 00:00:00 2023 0 0 Nasl. z nasl. zaslona; Opis vira z dne 25. 3. 2024; Str. 161-167 0000-00-00 PostprintKoncna Objavljeno NiDoloceno 0000-00-00 0000-00-00 2023-03-19 004 10.1109/DSD60849.2023.00032 190218499 190215427 https://ieeexplore.ieee.org/document/10456782/authors#authors 1 7438caa3-012f-11ef-b1b4-001a4af901a5 https://dirros.openscience.si/Dokument.php?lang=slv&id=25723 RAZ_Vreca_Jure_2023.pdf RAZ_Vreca_Jure_2023.pdf 1 C6531548B85A232A34C012D82828C084 9ce02cf99a866174ed5560575f589346f77d3b9beb120389a7b435a397c14d7b d0052652-0133-11ef-b1b4-001a4af901a5 https://dirros.openscience.si/Dokument.php?lang=slv&id=25735 Institut Jožef Stefan 0 0 0