Deep Learning PC build - Compatiblity Check

[minhnhat93]

Hello everyone, I'm looking to build a PC for Deep Learning. Specially, I want to buy a PC that has 2 GPUs with the option to add an additional GPU in the future. The CPU should have 40 PCIe lanes so that it won't bottle neck the GPUs. So far, after looking on the Internet, I have compiled this list:
https://ca.pcpartpicker.com/list/hmwzLD
PCPartPicker part list / Price breakdown by merchant

CPU: Intel - Xeon E5-2620 V4 2.1GHz 8-Core Processor ($572.32 @ Amazon Canada)
CPU Cooler: Cooler Master - Hyper 212 EVO 82.9 CFM Sleeve Bearing CPU Cooler ($36.73 @ Amazon Canada)
Motherboard: ASRock - X99 Taichi ATX LGA2011-3 Motherboard ($325.50 @ Vuugo)
Memory: G.Skill - Aegis 32GB (2 x 16GB) DDR4-2133 Memory ($288.73 @ Newegg Canada)
Storage: Samsung - 960 EVO 500GB M.2-2280 Solid State Drive ($335.98 @ Memory Express)
Storage: Toshiba - 3TB 3.5" 7200RPM Internal Hard Drive ($110.23 @ Newegg Canada)
Video Card: PNY - GeForce GTX 1080 Ti 11GB Founders Edition Video Card (2-Way SLI) ($976.48 @ Amazon Canada)
Video Card: PNY - GeForce GTX 1080 Ti 11GB Founders Edition Video Card (2-Way SLI) ($976.48 @ Amazon Canada)
Case: Corsair - Air 540 ATX Mid Tower Case ($171.60 @ DirectCanada)
Power Supply: EVGA - 1000W 80+ Gold Certified Fully-Modular ATX Power Supply ($215.23 @ Amazon Canada)
Total: $4009.28
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2017-06-12 18:07 EDT-0400

The thing that troubles me the most is the compatibility of the motherboard with the CPU as said on the pcpartpicker site: "Some Intel X99 chipset motherboards may need a BIOS update prior to using Broadwell-EP CPUs. Upgrading the BIOS may require a different CPU that is supported by older BIOS revisions."
My question is: Will I have to do a BIOS upgrade with this motherboard/CPU combination? I prefer not to because I have no experience doing this before. Also, any comment on compatibility issues are appreciated. Thank you in advance! 😀

 

[imsqam]

The PCPartPicker is a generic note and you have to go the CPU compatibility list from the manufacturer to see what BIOS version is needed to support a CPU. Here is the link
http://www.asrock.com/mb/intel/X99%20Taichi/#CPU

We can see from this list that all BIOS versions support the 2620 and therefore there is not a BIOS compatibility issue.

Curious as to why you chose an E5-2620 over an i7-6850K, better single and multicore performance while only being about $120 more. This will also allow for faster RAM.
http://cpu.userbenchmark.com/Compare/Intel-Xeon-E5-2620-v2-vs-Intel-Core-i7-6850K/m4718vs3606
https://www.cpubenchmark.net/cpu.php?cpu=Intel+Xeon+E5-2620+v4+%40+2.10GHz&id=2766
https://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i7-6850K+%40+3.60GHz&id=2800

Also curious as to why so much graphical processing power, some of the links I reviewed even talked about running Deep Learning systems without a discrete GPU.

 

[Lehan123456789]

Hi,
If you are willing to wait a bit longer, thread ripper is slated to come out in a month or so, and may well be more powerful than its Intel counterparts, with the highest end having 16 cores, 4 channel memory and 64 pcie lanes!

 

[minhnhat93]

The PCPartPicker is a generic note and you have to go the CPU compatibility list from the manufacturer to see what BIOS version is needed to support a CPU. Here is the link
http://www.asrock.com/mb/intel/X99%20Taichi/#CPU

We can see from this list that all BIOS versions support the 2620 and therefore there is not a BIOS compatibility issue.

Curious as to why you chose an E5-2620 over an i7-6850K, better single and multicore performance while only being about $120 more. This will also allow for faster RAM.
http://cpu.userbenchmark.com/Compare/Intel-Xeon-E5-2620-v2-vs-Intel-Core-i7-6850K/m4718vs3606
https://www.cpubenchmark.net/cpu.php?cpu=Intel+Xeon+E5-2620+v4+%40+2.10GHz&id=2766
https://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i7-6850K+%40+3.60GHz&id=2800

Also curious as to why so much graphical processing power, some of the links I reviewed even talked about running Deep Learning systems without a discrete GPU.

Hi, I just saw that website last night too. Thanks a lot. Cheers! 😀
The thing about deep learning is CPU speed and RAM speed are irrelevant. Our grant only allow about 4000 CAD. That's why we opted for the E5-2620 v4 which is the cheapest CPU/Motherboard combination that has 40 pcie lanes. All computation will be done on GPU. We want as many PCIe lanes for CPU and as many GPUs as possible.

Hi,
If you are willing to wait a bit longer, thread ripper is slated to come out in a month or so, and may well be more powerful than its Intel counterparts, with the highest end having 16 cores, 4 channel memory and 64 pcie lanes!

64 pcie lanes? That's so awesome. A 4 GPUs setup run at full bandwidth! It's a pity we need to buy a PC as soon as possible for our project has already started. I'm gonna check it in the future for my own Deep Learning cube 😀

 

[imsqam]

Okay, I see where this is going. There are actually cheaper Xeons with 40 PCI-e lanes, here they are, one 6-core and one 8-core.

http://ark.intel.com/compare/92993,92990,92986
https://ca.pcpartpicker.com/products/compare/gk98TW,pfvZxr,XjZ2FT/

 

[tbaierl]

I think up to four gpus are possible in a system. They use one (Xeon E5 1600) and two (Xeon E5-2600) processor boards to connect four gtx 1080 ti: https://www.cadnetwork.de/de/produkte/deep-learning
Using more than 4 gpus in one system could be a problem with the pcie bandwidth. But whats the difference between gtx 1080 ti and tesla p100, looks that they have nearly the same performance for tensorflow and other deep learning frameworks?