| Title |
An Entropy Model for GPU Register Compression |
| Authors |
(Minsik Kim) ; (Yunho Oh) ; (Won Woo Ro) |
| DOI |
https://doi.org/10.5573/JSTS.2025.25.3.228 |
| Keywords |
GPU; register file; register compression; entropy. |
| Abstract |
There have been continuous efforts to improve GPU energy efficiency. Register compression is a recent idea to tackle the efficiency problem, as the GPU register file is nontrivial in size and power-hungry to support massively concurrent thread execution. This paper proposes a mathematical model to estimate the ideal compression efficiency for GPU registers based on entropy theory. An ideal compression algorithm is expected to have a high compression ratio, however, the ratio varies significantly with application characteristics and also with the choice of algorithm. The proposed model can provide a theoretical bound of the ratio which is highly useful for evaluating register compression algorithms. Experimental results show an ideal case upper bound according to our model is 3.99×, while the existing algorithm based on BDI compression can achieve only 2.33×. This provides a strong case for exploring better register compression algorithms. |