January 2021 Summaries

The Tesla A100 server has been benchmarked against the Tesla V100 server, with mixed results. The A100 is found to be significantly faster than the V100 in terms of training speed for convolutional neural networks (CNNs), with 2.2x and 3.4x speedups using 32-bit precision and 1.6x and 3.4x speedups using mixed precision, respectively. In contrast, the A100 is not found to be significantly faster than the V100 in terms of training speed for language models, with only a 3.39x speedup on average. The benchmarking was conducted using PyTorch 20.10 and the same software versions were used for both servers, although results are expected soon for the A100.

The RTX A6000 is a new GPU designed for deep learning workstations and servers, shipping now from Lambda. It offers improved performance compared to other GPUs in its class, including the Tesla A100 and V100, with PyTorch convnet training speeds up to 1.5x faster than the RTX 2080 Ti and TensorFlow convnet training speeds up to 1.8x faster. The GPU features 48 GB of GDDR6 memory and is priced at $4,650. Benchmarks using Lambda's benchmark software show varying performance results among different GPUs, with some outperforming others in certain workloads.