A novel GPU register file organization capable achieving an improvement in IPC up to 18.6% on average.

[erek]

A GPU Register File using Static Data Compression

"GPUs rely on large register files to unlock thread-level parallelism for high throughput. Unfortunately, large register files are power hungry, making it important to seek for new approaches to improve their utilization. is paper introduces a new register file organization for efficient register-packing of narrow integer and floating-point operands designed to leverage on advances in static analysis. We show that the hardware/soware co-designed register file organization yields a performance improvement of up to 79%, and 18.6%, on average, at a modest output-quality degradation. "

https://arxiv.org/pdf/2006.05693.pdf

 

[defaultluser]

Looks interesting. Quality reduction is not going to be acceptable for rendering and anything that requires 64-bit accuracy, but could still be useful for games and deep learning.