Nvidia has accused Intel of using out-dated benchmarks in a recent comparison between its Knights Landing Xeon Phi cards and Nvidia’s own Maxwell GPU system. The comparison took place during a recent high performance computing conference; where Intel claimed that its cards provide 2.3x faster training for neural networks and 38-percent better scaling across all nodes.
The claim from Nvidia is that Intel was using an older version of the Caffe AlexNet benchmark; and that a newer version would show that Maxwell is 30-percent faster than the Xeon Phi. It also points out that Maxwell is the old architecture, and that the new Pascal GPUs would be as much as 90-percent faster.
Massaging data to present a flattering image of products is not something new to the tech industry; especially during presentations. Nvidia themselves were actually sued by Samsung for allegedly faking benchmark scores between the Exynos and Tegra processors.
Intel, on the other hand, responded to the accusation by saying that it stands by its data from the tests. Shots may have been fired as the statement says that “it is completely understandable that Nvidia is concerned about Intel in this space.”
Observers will note that comparisons done during corporate presentations don’t actually mean much. For the most part, reviewers and third parties will end up doing their own tests to determine which of the designs is better. In other words, Nvidia really has nothing to worry about over a single slide.
[Source: Ars Technica]
