SU-E-T-473: Improvements of GPU Based Calculations Over CPU Based Calculations and the Effects of Variable Computer Hardware

Authors


Abstract

Purpose:

To compare the calculation times in Raystation treatment planning system (TPS) for different graphical processor units (GPU). Also, to determine the amount of time saved by using a GPU based calculation.

Methods:

Four different hardware configurations were utilized. (DC1) – Intel Xeon E5-2650 dual CPU with 64 GB RAM and nVidia Quadro K4000 GPU. (DC1 Titan Black) – Same configuration as (DC1) but with nVidia Titan Black GPU. (RVW1) – Intel Xeon E5-2650 CPU with 32 GB RAM and nVidia Quadro K4000 GPU. (RVW2) – Intel Xeon E5-2650 CPU with 16 GB of RAM and Quadro K2000 GPU. Dose calculation was also performed without the GPU on DC1 and RVW2. Seven clinically delivered treatment plans were elected. The plans were picked with varying IMRT complexities. Each plan was calculated consecutively eleven times, and each plan had a uniform dose grid voxel size of 2 mm3.

Results:

When compared to non-GPU calculation, the average time on DC1 with GPU improved by 21.8% (range: 9.4%–38.9%). Similarly, the average time on RVW2 with GPU improved by 42.3% (range: 35.3%–50.5%). Also, the DC1 Titan Black average calculation time was 48.6% faster (range: 13.1%–73.4%) than the nVidia Quadro GPU models. Overall, the DC1 Titan Black calculation times outperformed DC1 non-GPU by an average of 58.3% (range: 32.9%–77.5%) For the prostate and nodes study, the DC1 Titan Black reduced times from 724±2.5 seconds 205±1.0 seconds.

Conclusion:

Dose calculation time is reduced with the addition of a GPU. Also, the nVidia Titan Black GPU provided better time savings than the nVidia Quadro GPU models. Available RAM also plays a role in calculation time due to system lag when the memory reaches capacity. The clinical user must determine the best hardware specifications for his or her institution based on speed but also on monetary cost.

Ancillary