Post by AimValley

5,590 followers

๐—”๐—ฐ๐—ฐ๐—ฒ๐—น๐—ฒ๐—ฟ๐—ฎ๐˜๐—ถ๐—ป๐—ด ๐— ๐—ฅ๐—œ-๐—š๐˜‚๐—ถ๐—ฑ๐—ฒ๐—ฑ ๐—ฅ๐—ฎ๐—ฑ๐—ถ๐—ผ๐˜๐—ต๐—ฒ๐—ฟ๐—ฎ๐—ฝ๐˜† ๐˜„๐—ถ๐˜๐—ต ๐—™๐—ฃ๐—š๐—” ๐—ฃ๐—ผ๐˜„๐—ฒ๐—ฟ ๐˜›๐˜ถ๐˜ณ๐˜ฏ๐˜ช๐˜ฏ๐˜จ ๐˜ค๐˜ฐ๐˜ฎ๐˜ฑ๐˜ถ๐˜ต๐˜ข๐˜ต๐˜ช๐˜ฐ๐˜ฏ๐˜ข๐˜ญ ๐˜ฃ๐˜ฐ๐˜ต๐˜ต๐˜ญ๐˜ฆ๐˜ฏ๐˜ฆ๐˜ค๐˜ฌ๐˜ด ๐˜ช๐˜ฏ๐˜ต๐˜ฐ ๐˜ฃ๐˜ณ๐˜ฆ๐˜ข๐˜ฌ๐˜ต๐˜ฉ๐˜ณ๐˜ฐ๐˜ถ๐˜จ๐˜ฉ๐˜ด ๐˜ช๐˜ฏ ๐˜ต๐˜ณ๐˜ฆ๐˜ข๐˜ต๐˜ฎ๐˜ฆ๐˜ฏ๐˜ต ๐˜ด๐˜ฑ๐˜ฆ๐˜ฆ๐˜ฅ. Proud to share the results of our work on the ART-PIVOT project, in collaboration with UMC Utrecht and NWO (Dutch Research Council), focused on accelerating adaptive MRI-guided radiotherapy planning and reducing overall treatment planning time. ๐—ง๐—ต๐—ฒ ๐—ฐ๐—ต๐—ฎ๐—น๐—น๐—ฒ๐—ป๐—ด๐—ฒ UMC Utrechtโ€™s MR-linac research pipeline runs entirely on CPU (MATLAB, FP64), where matrixโ€“vector multiplication dominates compute time. With matrices reaching sizes of up to 3M ร— 25K, the goal was to offload the most demanding kernels to an FPGA acceleratorโ€”without compromising output quality. ๐—ช๐—ต๐—ฎ๐˜ ๐˜„๐—ฒ ๐—ฏ๐˜‚๐—ถ๐—น๐˜ On the AMD Versal VCK190 (VC1902, 400 AIE tiles), we developed two accelerators targeting the primary bottlenecks: โ€ข A dense matrixโ€“vector (GEMM) accelerator using an 8ร—23 AIE tile layout, combined with a custom sort engine and add engine in the programmable logic โ€ข A sparse matrixโ€“vector (SpMM) accelerator featuring a novel compression scheme that eliminates zero-padding overhead and leverages the row-sparse structure of UMC Utrechtโ€™s matrices ๐—ง๐—ต๐—ฒ ๐—ฟ๐—ฒ๐˜€๐˜‚๐—น๐˜ Full-system VSS co-simulation shows a 4.16ร— speed-up over the MATLAB baseline for GEMM on a 3M-row workload. Interestingly, the main bottleneck was not compute, but DDR read bandwidthโ€”highlighting that the AIE array still has headroom for further performance gains. ๐—ž๐—ฒ๐˜† ๐˜๐—ฎ๐—ธ๐—ฒ๐—ฎ๐˜„๐—ฎ๐˜† AMD Versal with AIE tiles proves to be a strong fit for floating-point, compute-intensive medical imaging workloads. While DDR bandwidth remains the primary architectural constraint, there is significant opportunity to exploit parallelism within those limits. Curious how advanced compute architectures like FPGA and AI Engines can accelerate your most demanding workloads? At AimValley, we turn complex algorithms into efficient, high-performance implementations. Letโ€™s connect and explore whatโ€™s possible. #umcu #nwo #mri #fpgaacceleration #fpgapower MathWorks #versal #AIengines AimValley is a Technolution Company!

Post content