In this article from Tim Dettmers: Full hardware guide for Deep Learning, he states that PCI lanes speed doesn't matter a lot for Deep Learning.
But on this question, the approved answer states that a CPU with the right amount of PCI lanes is required for Deep Learning. (according to me PCI lane speed and amount of PCI lanes managed by CPU are dependent)
So, I'm looking at a benchmark for Deel Learning to compare those theories/facts.
Looking at the existing benchmarks for Deep Learning, it seems that they are all only testing GPUs capacity, comparing one Graphic card to another, focusing on matrix multiplication.
I'm wondering if there is a benchmark that scores the sum of all hardware components required for deep learning (and especially the training phase): CPU, Memory, GPU, PCIe speed, hard drive, motherboard.
To list what I have tested:
- DAWNBench/MLPerf Define the specifications of a DL benchmark, whatever the platform and the framework. Their code is partially incomplete.
- AI Benchmark primarily focusing on mobile. Simple to use with the code available and a friendly scoring system. But even with a high precision parameter, the models are tested on small batches (I guess it's to keep it usable on mobile)
- AI Matrix from Alibaba: A very heavy and complete one but testing only the GPUs.
- Lambda Labs Benchmark, I didn't get it working... non-available version of Tensorflow, non-compatible with recent Cuda versions..
(I know this is related to hardware, but as I'm looking for a software library, I post here!)