GPU computing in medical physics: A review
Number of publications relating to the use of GPUs in medical physics, per year. Data were obtained by searching PubMed using the terms “GPU,” “graphics processing unit,” and “graphics hardware” and excluding irrelevant citations.
Number of computing cores (▽) and memory bandwidth (Δ) for high-end NVIDIA GPUs as a function of year (data from vendor specifications).
Computing performance, measured in billion single-precision floating-point operation per second (GFLOPS), for CPUs (Δ) and GPUs (▽). GPUs: (A) NVIDIA GeForce FX 5800, (B) FX 5950 Ultra, (C) 6800 Ultra, (D) 7800 GTX, (E) Quadro FX 4500, (F) GeForce 7900 GTX, (G) 8800 GTX, (H) Tesla C1060, and (I) AMD Radeon HD 5870. CPUs: (A) Athlon 64 3200+, (B) Pentium IV 560, (C) Pentium D 960, (D) 950, (E) Athlon 64 X2 5000+, (F) Core 2 Duo E6700, (G) Core 2 Quad Q6600, (H) Athlon 64 FX-74, (I) Core 2 Quad QX6700, (J) Intel Core i7 965 XE, and (K) Core i7-980X Extreme (data from vendors).
The graphics pipeline. The boxes shaded in light red correspond to stages of the pipeline that can be programmed by the user.
GPU thread and memory hierarchy. Threads are organized as a grid of thread blocks. Threads within a block are executed on the same MP and have access to on-chip private registers (R) and shared memory. Additional global and local memories (LM) are available off-chip to supplement limited on-chip resources.
Scatter and gather operations in iterative reconstruction for computed tomography.
Article metrics loading...
Full text loading...