Nvidia Leads Again in MLPerf Benchmarks: Generative AI at the Top

Nvidia's graphics processors once again delivered the best results in one of the most prestigious benchmarks for measuring AI chip performance – MLPerf. This time, the main focus was on generative AI applications, such as large language models (LLMs), particularly Llama by Meta. In the tests, systems built by companies like SuperMicro, Hewlett Packard Enterprise, and Lenovo almost entirely dominated, taking first place in most categories.

New Test Scenarios for Generative AI

MLPerf updated its tests by adding several new scenarios, particularly for working with large language models. One new test measures the speed of token generation based on Llama 3.1, while another evaluates the response time of chatbots based on Llama 2. Additionally, a test was added to measure the speed of graph neural networks, which are increasingly used in generative AI, such as for protein folding predictions.

While competition was limited, AMD also showed strong results in two Llama 2 tests. Overall, however, Nvidia demonstrated the greatest power in the tests, especially in the closed division, where the requirements for configuration were the strictest.