Leifeng.com news, on June 30, Beijing time, the MLCommons community released the latest MLPerf2.0 benchmark evaluation results. In a new round of testing, MLPerf has added a new object detection benchmark for training the new RetinaNet on the larger OpenImages dataset. MLperf said this new object detection benchmark can more accurately reflect applications for autonomous driving, Advanced machine learning training results for applications such as robotic obstacle avoidance and retail analytics.
The results of MLPerf2.0 are about the same as the results of v1.1 released in December last year, and the overall performance of AI is improved by about 1.8 times compared with the previous round of releases.
Twenty-one companies and institutions submitted scores for the MLPerf benchmark in the latest round, bringing the total number of submissions to more than 260.
Nvidia is still “playing the full game”
In this test, Nvidia remained the only participant to complete all eight benchmarks in version 2.0. These tests cover current popular AI use cases, including speech recognition, natural language processing, recommender systems, object detection, image classification, and more.
No accelerators other than Nvidia have run all the base tests. Nvidia, on the other hand, has completed all basic tests since it first submitted test results to MLPerf in December 2018.
A total of 16 partners used the NVIDIA platform to submit the results of this round of testing, including ASUS, Baidu, Institute of Automation, Chinese Academy of Sciences, Dell Technologies, Fujitsu, Gigabyte, H3C, HPE, Inspur, Lenovo, Ningchang and Supermicro . In this round of MLPerf benchmark results, NVIDIA and its partners accounted for 90% of all participating ecosystem partners.
This shows the good generality of the Nvidia model.
Universality provides a basis for models to work together in actual production. AI applications need to understand the user’s request, and as required, classify images, make suggestions, and respond in the form of voice messages.
To accomplish these tasks, multiple types of AI models are required to work together. Even a simple use case requires nearly 10 models, which puts a requirement on AI model versatility.
Good versatility means that users can work with the same facilities as much as possible throughout the AI process, and can also be compatible with new requirements that may arise in the future, thereby extending the service life of the infrastructure.
AI processing performance has increased by 23 times in three and a half years
In this benchmark result, the NVIDIA A100 remains the leader in single-chip performance, achieving the fastest results in four of the eight tests.
Two years ago, Nvidia used the A100 GPU for the first time in the MLPerf 0.7 benchmark, and this is the fourth time Nvidia has used the GPU to submit benchmark results.
In the three-and-a-half years since MLPerf came out, the Nvidia AI platform has achieved a 23x performance improvement in benchmarks. In the two years since the MLPerf benchmark was first submitted on the A100, the performance of the Nvidia platform has also improved by a factor of 6.
The continuous improvement in performance is due to NVIDIA’s innovation in software. Continuously unlocking more performance of the Ampere architecture, such as the heavy use of CUDA Graphs in submitting results, minimizes the startup overhead of running across multiple accelerators.
It is worth noting that NVIDIA did not choose to use its recently released Hopper GPU in this round of testing, but chose the NVIDIA A100 Tensor Core GPU based on the NVIDIA Ampere architecture.
Nvidia Narasimhan said that Nvidia prefers to focus on commercially available products, which is why Nvidia chose to submit results based on the A100 in this round.
Given that the new Hopper Tensor Cores can apply mixed FP8 and FP16 precision data, and NVIDIA is likely to use Hopper GPUs in the next round of MLPerf testing, it is foreseeable that NVIDIA’s performance in the next round of benchmarks is expected to be even greater. leap.
This article is reprinted from: https://www.leiphone.com/category/chips/JEWsTg8zc9J4gCSf.html
This site is for inclusion only, and the copyright belongs to the original author.