Nvidia Tops New AI Benchmark for Running LLMs

• Nvidia’s H100 chips emerged as the top performer in the MLPerf benchmark for AI inference.
• Intel’s Gaudi2 chips were approximately 10% slower than Nvidia’s, but are more competitively priced.
• Google’s latest custom-built chip is expected to perform well in the MLPerf benchmark.
Advertisement

MLCommons, a benchmark group for artificial intelligence, has released the results of tests evaluating the speed of high-end hardware when running AI models.

In these tests, a chip developed by Nvidia Corp emerged as the top performer, closely followed by Intel Corp’s semiconductor.

The new benchmark, called MLPerf, focuses on a large language model with 6 billion parameters designed for summarizing news articles. This benchmark assesses the “inference” phase of AI processing, which is crucial for generative AI applications.

Nvidia’s leading submission for the inference benchmark utilized eight of its flagship H100 chips. While Nvidia has been dominant in the AI model training market, it has yet to capture a significant share of the inference market.

Nvidia’s accelerated computing marketing director, Dave Salvator, emphasized their leadership performance in various workloads, highlighting their overall strength.

Intel’s success in the benchmark is attributed to its Gaudi2 chips, developed by the Habana unit acquired by the company in 2019. Intel’s Gaudi2 system was approximately 10% slower than Nvidia’s.

Eitan Medina, Habana’s chief operating officer, expressed pride in the inferencing results, emphasizing the price-performance advantage of Gaudi2. Intel claimed that its system is competitively priced compared to Nvidia’s, roughly equivalent to the cost of Nvidia’s previous generation 100 systems, although specific chip costs were not disclosed.

Both Nvidia and Intel declined to reveal the exact pricing of their respective chips. Nvidia also announced plans to release a software upgrade aimed at doubling the performance seen in the MLPerf benchmark.

Alphabet’s Google unit offered a preview of the performance of its latest custom-built chip, which was initially announced at its cloud computing conference in August.

Also Read

AI-Powered Technology Could Make Antibody Treatments More Effective

Algorithms identify problematic regions in antibodies for non-target molecule binding. Algorithms pinpoint...

To stay informed about current events, please like our Facebook page https://www.facebook.com/BOLUrduNews/.