If you read closely, he actually recommends blower style GPU.
QUOTE:
Blockquote
Update : To avoid overheating, I now use these blower-style GPUs:
1 Like
Hello everyone!
I have been researching to build my Deep Learning computer and I would like to get your opinion. The configuration is here: https://uk.pcpartpicker.com/user/rpicatoste/saved/#view=7nrvWZ
The reasoning behing the choices is the following:
AMD - Threadripper 1920X 3.5 GHz 12-Core Processor: I want to do intensive deep learning, which will include image (and therefore a lot of preprocessing), and also DRL, where simulating will require a lot of CPU+RAM.
Gigabyte - X399 AORUS PRO ATX TR4 Motherboard: This is more or less what I saw recommended by the shop for this processor. I see it has pci lanes for several GPUs and slots to expand the RAM. Other than that I donāt know if itās a good choice.
Corsair - Vengeance LPX 32 GB (2 x 16 GB) DDR4-3000 Memory: As seen from several places in the forum and in different pages related to this, more than 3000 Mhz doesnāt improve a lot. I was thinking to start with 64Gb, but the computer is very expensive already and this is really easy to update in the future (just getting another 2x16Gb RAM sticks).
Samsung - 970 Evo 1 TB M.2-2280 Solid State Drive: Seems to be the clear choice for HDD and the data for training. Extremely fast.
Western Digital - Black 4 TB 3.5" 7200RPM Internal Hard Drive: A good HDD for cold storage.
MSI GeForce RTX 2070 ARMOR 8G: Apparently it has a very good chipset. I have seen it recommended in OC places, altough it donāt plan to OC. The 2070 is the choice as better power/price relation, and since I build this having in mind the possibility of a 2nd GPU, I could end up adding an equal one or who knows, a 2080 Ti.
Cooler Master - MasterCase H500: Just following another builts (from this forums 
). I donāt have a strong opinion on case, I just want it to have space for everything. Does all the cases include dust filters? Thatās something I would like to have, but didnāt find clear indications of which cases have it.
SeaSonic - FOCUS Plus Gold 1000 W 80+ Gold: With the idea of a 2nd GPU in the future.
Processor fan: I have absolutely no idea, my plan is to ask the shop. I donāt want liquid cooling to avoid adding more potential sources of problems or maintenance. Any advise is welcome.
The current price is 2610 CHF (2324 EUR) with windows (I plan to have dual boot with Ubuntu 18.04) and assembly, in the shop where I plan to buy: https://www.ldlc.com/fr-ch/configurateur-pc/bcd3a16aa572c51c2fbeb6b263839aa0/
Thank you for any advise!
Hello Ricardo Picatoste, @rpicatoste
I just bought a configuration a lot better than this and for 2000 euros. if you are interested i can guide you to better setup. this is a lot wrong here .
Sure! Please feel free to share your choice of components and opinion
1 Like
I did the research for about a year before buying the components(you can trust me i need best pc possible ) . this is the components I would buy for about 2000 euros.
GPU : 1080 ti
CPU : intel i7 gen 7 ou 8 (donāt buy K they donāt add value here)
SSD : The sheapest you can find
RAM : the cheapest DDR4 also (2133MHz more dosenāt add value , buy more ram if you have money, it better)
Case: the cheapest full tower
powe suply : 750w is more than you need so 750 is safe (if you buy a new card you can replace it then , use the money to buy more RAM)
motherboard : you need to make sure it support the CPU (look in their site). irecommand using the same as me. Asus ROG STRIX B250F Gaming
CPU cooler : cooler master 70 around euros is more than u need
Uploading: IMG_20190415_115548.jpgā¦
Uploading: IMG_20190415_115919.jpgā¦
Uploading: IMG_20190415_120009.jpgā¦
Uploading: IMG_20190415_120109.jpgā¦
sorry for the misspellings
Uploading: IMG_20190415_115548.jpgā¦ Uploading: IMG_20190415_115919.jpgā¦ Uploading: IMG_20190415_120109.jpgā¦ Uploading: IMG_20190419_123002.jpgā¦
Iām still not convinced about the AMD cpus. Intel notoriously castrates MKL when it doesnāt detect an Intel processor.
1 Like
Thatās too bad to hear! I am hearing consistently about how AMD is doing better for this kind of use.
I will try to find some benchmark to seeā¦
Thanks for the warning!
1 Like
Consider used Xeons. E5-v3/v4 are becoming cheap on ebay, and the same stands for x99 boards with PXE/PLX chips (that is, four slots at full 16X gen3 speed).
If you use it on windows though, I found a repo where the author optimized the low level code on his own for his 1950x, itās a little faster than i7 8700k.
1 Like
Hey,guys, though I guess not many people are using threadripper nowadays, but for those that do, you can set the environment variable: export MKL_DEBUG_CPU_TYPE=5 to get a much improved performance(rival that of 8700k) with NumPy and stuff.
3 Likes
Sir can we use AMD processors instead of Intel processors? Amd Ryzen 5 + nvidia GTX 1650. Will there be any issues??
1 Like
Why nobody is going for AMD processors for Machine learning??
Only Nvidia cards(as far as I know) support cuda which is essential for deep learning.
1 Like
Ok my projected configuration is AMD RYZEN 5 with Nvidia GTX 1650. Will there be any issues?
There will not be any issues. Although, intel processors have few optimzations specifically for pytorch (Intel MKL-DNN has been integrated into official release of PyTorch by default, thus users can get performance benefit on Intel platform without additional installation steps.) You might miss out of those (https://software.intel.com/en-us/articles/getting-started-with-intel-optimization-of-pytorch)
3 Likes
Thank you very very much for your time and support. Now i got the exact picture. But is there any way to bring out optimisation in amd for Pytorch?
Keep in mind that these are Xeon CPUs and not i7s.
Iām thinking about building a box with AMD tbh but I need your opinion on that.
Performance gain
1 Like