It depends on the application. To me, if you want to do some picture processing tasks, Bolstering GPU could be better. If you want to deal with time-series data or other types of signals CPU is much more beneficial.

Fazla Rabbi Mashrur ,

Depends on the meaning of "something bigger". the main issue with optimising GPU applications (e.g. CUDA) is that memory transfers are slow. If you do preprocessing on individual cores also take into account NUMA. I would worry more about the memory layout, the system bus on your board, how you upload data from disk and how you optimise the transfer to GPU memory. This will maximize throughput on whichever system you choose.

What kind of task you intended to conduct in deep learning field ?

For deep learning with image processing, you need the best memory of ram GPU to perform training works. Additionally, the more GPUs (i.e., might be, two cheaper GPUs are also okay) are attached, the faster processing speed.

Better more GPUs

It depends on what application and algorithm (how complex). All in all the bigger memory is better.

i suggest you to read this

https://medium.com/@shachishah.ce/do-we-really-need-gpu-for-deep-learning-47042c02efe2

This is only true if the software supports many cores. And even better GPU for non graphic computing

It's better to use GPU when you need to speed-up the data processing.

Moreover GPU is required when your model or system need to manage a huge amount of data.

Indeed CPU can be used to train smaller data models. If you are new and have a tight budget a 4 core CPU should be good enough. It can train slowly. GPU was actually designed for better graphical experience since they are equipped with more of RAM in the graphics adapter. Hence, obviously it is a better choice for deeplearning. You may have to however choose a proper toolset to run the code in CPU.

I work on detection and classification based applications for that GPU performs adequately better than CPU. However for series type data CPU is conventionally better than GPU to limit the extra processor.

to continue with low budget CPu configuration you must ensure The OS is linux(since most of the existing deeplearning algos give better and faster performance in linux)

Not ture always, as it depends on what kind of algorithm your are using or for which kind of application you are using.

Dear @ Fazla Rabbi Mashrur

I think you had better to prepare your model first, I mean not completely but something in very small scale and try to run it by any system you got in hand, then you will find out what is your main priority in terms of CPU or GPU.

regards

It depends on the depth of the model network which is the number of layers between input and output layers. Sometimes, to train a simple deep network we may be in need of only a single CPU with 8-core at minimum. As we train more complex models especially those models with high-resolution image datasets, we need to use A single GPU on a matrix of GPU to minimize training time. Of course cost-effective is an important factor that should be taking into account.

Yes indeed., https://aws.amazon.com/machine-learning/accelerate-amazon-sagemaker/?trk=ps_a134p000006BnGnAAK&trkCampaign=2020_HowMLisDone_campaign_landing_page&sc_channel=ps&sc_campaign=SM_AD_Google_CL_RSA_SMML_PNB&sc_outcome=AIML_Digital_Marketing&sc_geo=mult&sc_country=mult&sc_publisher=Google&s_kwcid=AL!4422!3!428016791724!b!!g!!ai%20and%20machine%20learning&sc_detail=ai%20and%20machine%20learning&ef_id=Cj0KCQjwudb3BRC9ARIsAEa-vUsU9PDQ25yQsguEYByUkfE0CpP3KLD1VaGstHLNm5UQyX-D7LEKBUYaAhV9EALw_wcB:G:s&s_kwcid=AL!4422!3!428016791724!b!!

In deep learning number of CPU cores don't matter that much unlike the GPU cores.

GPU have many weak cores and that is what accelerates the training time.

Deep learning requires more number of core not powerful cores.

And once you manually configured the Tensorflow for GPU, then CPU cores and not used for training.

So you can go for 4 CPU cores if you have a tight budget but I will prefer to go for i7 with 6 cores for a long use, as long as the GPU are from Nvidia.

And you can use Google cloud service Google Colab. This is what I use for deeper model training as Apple MacBooks don't provide the CUDA supported Nvidia GPUs.

yes,may be,depend on cpu cores.

If it is a complex model of deep learning, it mainly depends on GPU.

For deep learning, the number of cuda dores of your dedicated GPU is as decisive as the size of your video ram, which is only 6 GB for RTX 2060. The number of cuda cores is about 1900 for desktop version, against 6 cores of the I7 10th. This shows why GPUs speed up training. Actually, any GPU improves performance according to cores, but you must calculate beforehand what size your model may have later (maybe use mixed precision models, this helps reducing overall model size), since it must fit into your VRAM. This is most important, otherwise you get a bottleneck. Data is copied from RAM to VRAM, which is why AMD should be ahead at the moment due to PCIE 4.0 which is faster than Intel with PCIE 3.0, but the driver support is inadequate for deep learning, so GPU should be from NVIDIA. The best budget option for deep learning at the moment is the RTX 2070 with 8 GB of RAM. These GPUs are mainly if not exceptionally combined with I7, so the processor should be I7. Don't waste money on RAM, since you only need 1.5 times VRAM, which makes 12 GB sufficient for a 2070. If your budget only allows for a 2060, you can later get a larger external GPU and connect it via Thunderbolt. This is a good solution if you prefer a thin notebook without dedicated GPU but with Thunderbolt. You see, it is quite complicated. Maybe get a Dell G3 or G5, or some Nvidia Studio laptop, which will serve your purpose according to VRAM.

deep learning is training and inference. CPU is not the best for training as it requires a lot of compute and CPUs are not the best. This why NVIDIA or AMD GPUs are the right compute solution. For Inference here we talk about CPUs or even FPGA or ASICs. Here I think we can talk about more CPU core count will do the difference in time of inference DL operations.

First of all, a better GPU is what you need if you do a deep learning task. Then you need to think about your tools to facilitate the pre-processing of the deep learning task, whether they are complex and with high time consumption. Generally, you should buy a computer with both a powerful CPU and (more than) a GPU because you never know what your task needs in the future. My recent work needs a powerful GPU but at the same time, the pre-processing needs a powerful CPU.

Last but not least, use Intel rather than AMD if you use Matlab. I don't know whether the performance behave high difference in the other platform between the two CPUs.

GPU computation will give you better accuracy to work with deep learning models

The answer to your question is it depends on what you want to do with the network. If you want to do training, GPU will be the way to go. If inference, CPU should be better. Of course irrespective of CPU or GPU, the more the number of cores the better.

budget 6 core processor is enough for data preprocessing whether AMD/Intel. GPU selection is the big task for DL algorithms CUDA enabled GPU show better performance and faster.

GPU can save you time... I don't have now powerful workstation with gpu, so almost all my models base on LGBM or XGboost - its enough for mediocre results. However, I recommend to have machine wit powerful gpu - you will make more complex models faster.

Working with Deep learning models required GPU which offers a faster time of training with better results.

In general CPU consists of few bugger cores but in GPU it has several small cores and mainly focuses on graphic based performance. This is the main reason why GPU is preferred more for deep learning algorithms. GPU is highly specialized for parallel processing, while CPU is designed to handle many different kinds of operations.

thank you for sharing the question

Focusing on a decent GPU rather than a CPU is ideal for Deep learning.