Benchmark test reveals how NVIDIA and Intel's AI chips perform
Artificial intelligence benchmark organization known as MLCommons has revealed the results of new tests assessing the speed at which top-of-the-line hardware can process AI models. The focus of these tests is the MLPerf benchmark, centered on a large language model with six billion parameters that summarize CNN news articles. The benchmark specifically simulates the "inference" phase of AI data processing, which plays a pivotal role in powering generative AI tools.
NVIDIA dominates with H100 chips
These benchmark tests showcase NVIDIA's chip as the top performer, closely followed by Intel's semiconductor. NVIDIA's submission for the inference benchmark employs eight of its flagship H100 chips. While NVIDIA has maintained its dominance in the market for training AI models, it has been seeking a foothold in the inference market. Dave Salvator, NVIDIA's accelerated computing marketing director, expressed their pride in delivering leadership performance across various workloads.
Intel's Gaudi2 chips close behind
Intel, on the other hand, has made significant strides with its Gaudi2 chips, which were acquired through the Habana unit in 2019. Intel's Gaudi2 system, although approximately 10% slower than NVIDIA's, has garnered attention for its price-performance advantage. Eitan Medina, Chief Operating Officer at Habana, emphasized the value proposition of Gaudi2, highlighting its competitive cost-effectiveness compared to NVIDIA's offerings. While specific chip costs remain undisclosed, Intel is positioning itself as an appealing alternative in the AI hardware landscape.
Price v/s performance
One notable aspect of the competition between NVIDIA and Intel is the balance between price and performance. While Intel asserts its system's affordability relative to NVIDIA's, the latter is gearing up for further improvements. NVIDIA recently announced plans to roll out a software upgrade that promises to double its performance based on the MLPerf benchmark.
Google's custom chip enters the fray
In addition to the NVIDIA-Intel rivalry, Alphabet's Google unit has joined the competition. Google offered a sneak peek into the performance of its latest custom-built chip, initially introduced at its cloud computing conference in August. This development adds an intriguing dimension to the AI hardware landscape, as Google aims to carve out its own niche in the rapidly evolving AI industry.
