March 2019 Summaries

The article benchmarks the NVIDIA Titan V's deep learning and machine learning performance using TensorFlow, comparing it to other commonly used GPUs such as RTX 2080 Ti, RTX 2080, Titan RTX, Tesla V100, and GTX 1080 Ti. The results show that for FP32 training, the Titan V is 42% faster than RTX 2080, 41% faster than GTX 1080 Ti, and 4% faster than RTX 2080 Ti, but only 90% as fast as Titan RTX. In contrast, for FP16 training, the Titan V is 111% faster than GTX 1080 Ti, 94% faster than Titan XP, and 70% faster than RTX 2080, with an average speed-up of +71.6% compared to FP32. The article also explores multi-GPU scaling performance, where using eight Titan Vs will be 5.18x faster than using a single Titan V, while using eight Tesla V100s is 9.68x faster than using a single Titan V. Additionally, the results show that FP16 can reduce training times and enable larger batch sizes/models without significantly impacting model accuracy.

The RTX 2080 Ti GPU performs significantly better than other GPUs in deep learning benchmarks, with some models achieving speeds up to 96% of the Titan V's performance. The GPU's use of Tensor Cores and its FP16 capabilities also contribute to its improved performance. In multi-GPU training, using more GPUs can lead to significant speedups, with 8x RTX 2080 Ti GPUs performing around 5 times faster than a single RTX 2080 Ti GPU. Additionally, the RTX 2080 Ti is found to be faster than other GPUs when training models in FP16 mode, with some models achieving speeds up to 81% faster than the Titan V. The results also show that lower precision modes like FP16 can reduce training times and enable larger batch sizes/models without significantly impacting accuracy.