Nvidia decided to dispute Intel’s benchmarks about Knights Landing Xeon Phi benchmarks and, rather controversially, decided that Chipzilla’s benchmarks are “out of date”.
The accusations from Nvidia are quite profound and can be found in detail on Nvidia’s blog. The General Manager/VP of Nvidia’s accelerated computing Ian Buck was quite clear about the thing they didn’t like in Intel’s way of benchmarking deep learning.
Perhaps they are a bit esoteric for some people, but Nvidia claims Intel benchmarks are very very ancient.
“Very very ancient,” alleges Nvidia, means that “Intel used Caffe AlexNet data that is 18 months old”.
Intel original claim is that Xeon Phi is 2.3x faster in training than GPUs, offers 38 per cent better scaling that GPUs across nodes and delivers strong scaling to 128 nodes while GPUs do not.
The way Nvidia measured things, four PCs with four Nvidia Titan X Maxwell based (last year's) cards need 8.2 hours to train the deep learning system. This figure is based on Caffe from August 2016 cuDNN5. This should be 1.3 time faster than Xeon Phi.
A PC with four Pascal based Titan X (the new ones that were just launched) needs 5.5 hours to train and is1.9 times faster than Xeon Phi.
Nvidia’s outrageously expensive $120.000 USD DGX-1 can train the same system in two hours or 5.3 times faster than Xeon Phi.
We thought that benchmark wars were a thing of the past, but of course, they are here to stay for many many years to come. As we note above, it’s all a bit confusing as it seems that the performance is something in eye of the beholder.This is why back in the days we can up with the term Benchmarketing.
This all looks similar to events of late 2008 when Intel decided not to let Nvidia make its Nehalem based chipset. Jen-Hsun decided to call a “can of whoppass” on Intel starting a public fight about chipset patents.
Nvidia won a $1.5 billion settlement over six years but it completely killed off Nvidia's chipset division. Nvidia ended the chispet production in late 2009. Nvidia tried to bully Qualcomm and Samsung into paying licencess too, but has not managed it so far.
The Nvidia’s bog entry is here.
