NVIDIA H100 Surpasses Generative Artificial Intelligence, Breaks MLPerf Record

2023-06-29

NVIDIA's flagship H100 chip once again demonstrates its outstanding competitiveness, achieving the best performance in a set of MLPerf training benchmarks. The GPU, in collaboration with CoreWeave and Inflection AI, underwent new tests and set new records in various metrics. These tests were conducted on a cluster of 3,584 H100 GPUs on the CoreWeave platform. The GPUs are interconnected through InfiniBand technology, which NVIDIA claims enables exceptional performance at both individual and scalable levels. MLPerf benchmarks aim to measure the capabilities of specific hardware by calculating the time required to complete specific workloads. The tests include various LLM (large language models) and computer vision models, as well as a small number of CNN (convolutional neural networks) and RNN (recurrent neural networks). NVIDIA claims that the H100 delivers the highest performance in each benchmark test, with some notable examples including training GPT-3 in just 11 minutes and completing the ResNet benchmark test in just 0.18 minutes. It is reported that the H100 is the only chip capable of completing all benchmark tests. While NVIDIA dominates the market, its competitors have lost their competitiveness. AMD's AI accelerator products are far behind the H100. The company recently launched the M1300X AI accelerator chip, which has 192GB of VRAM and aims to compete with the H100. Although the benchmark test results for this new chip have not yet been released, the lukewarm market response is enough to indicate that this product will not pose a threat to NVIDIA in the short term. Other notable competitors include Cerebras, which is comparable to NVIDIA in training large language models in terms of cost, and Google's TPU, which excels in energy efficiency and performance but is limited by specific workloads. In the foreseeable future, it appears that NVIDIA will continue to thrive in the field of AI, benefiting from powerful computing and a well-integrated software ecosystem. Source: https://analyticsindiamag.com/nvidia-aces-generative-ai-breaks-mlperf-records/

RECENT AI NEWS

RECENT AI TOOLS