Deep learning setup

[__Peter__]

Hello everyone,

I want to build a PC for deep learning.

One essential is a Nvidia Titan V (https://www.nvidia.de/titan/titan-v/).

For the other parts I choose the following:

 

• Corsair RM850x as PSU (PSU)
• X99-E WS/USB 3.1 (Mainboard)
• Thermaltake Core V71 (Case)
• Crucial 8GB DDR4-2400 (RAM)
• Intel Xeon e5 1650 v4 (CPU)

What is your opinion and do you think that all parts match with each other?

 

Thanks in advance!

[Crazycatman]

6 minutes ago, __Peter__ said:

Hello everyone,

I want to build a PC for deep learning.

One essential is a Nvidia Titan V (https://www.nvidia.de/titan/titan-v/).

For the other parts I choose the following:

 

• Corsair RM850x as PS (PS)
• X99-E WS/USB 3.1 (Mainboard)
• Thermaltake Core V71 (Case)
• Crucial 8GB DDR4-2400 (RAM)
• Intel Xeon e5 1650 v4 (CPU)

What is your opinion and do you think that all parts match with each other?

 

Thanks in advance!

Just make RAM 16GB. Everything else is good. But if you're on a budget 8GB is fine? I know nothing about Deep stuff, but 8GB is a budget thing these days. Also it's "PSU" But PS (Power Supply) makes just enough sense.

[__Peter__]

4 hours ago, Crazycatman said:

Just make RAM 16GB. Everything else is good. But if you're on a budget 8GB is fine? I know nothing about Deep stuff, but 8GB is a budget thing these days. Also it's "PSU" But PS (Power Supply) makes just enough sense.

Thank you. 16 GB is also in the budget. In the long term I will install about 100 GB RAM anyway.

[lacion]

i would say go with an i9 like the 7900x that has a decent price now, its way better than a xeon, you can get up to 128gb of ram with it. 

 

Xeon cpus are tuned for servers and to keep the TDP in check they boost a lot less and are clocked lower. you will be doing better with the HEDT lineup

[Crazycatman]

1 minute ago, lacion said:

i would say go with an i9 like the 7900x that has a decent price now, its way better than a xeon, you can get up to 128gb of ram with it. 

 

Xeon cpus are tuned for servers and to keep the TDP in check they boost a lot less and are clocked lower. you will be doing better with the HEDT lineup

Can't that one over clocked reach 5GHz?

[lacion]

2 minutes ago, Crazycatman said:

Can't that one over clocked reach 5GHz?

it depends on the needs, I would not overclock beyond boost on anything that is used for critical work, especially in things like deep learning where the resources are going to be used to their limit.

 

debugging and optimizing machine learning is already hard enough to have to take overclocking into account when things start failing.

[__Peter__]

1 hour ago, lacion said:

i would say go with an i9 like the 7900x that has a decent price now, its way better than a xeon, you can get up to 128gb of ram with it. 

 

Xeon cpus are tuned for servers and to keep the TDP in check they boost a lot less and are clocked lower. you will be doing better with the HEDT lineup

I already bought the mainboard, because it was super cheap deal. So now I'm stucked to a LGA 2011-3 Sockel... So only Core-i 7 or Xeon is suitable

[lacion]

5 minutes ago, __Peter__ said:

I read that the xeon chips are better for long time processing (several days) and the Core i-Series may lead to errors at such long runtimes. Is that right?

I never experienced that before. core i9 is designed for workstation use, and most time workstations do heavy workloads for extended times.

 

I have a 7900x that I use as a development workstation and a 7980XE that doe several workloads including  ML|CV training sessions and other workloads that take a few days sometimes.

 

xeons are really just designed to keep power usage and heat on check so you can stack shit loads of them in a rack and not require a nuclear reactor to cool your datacenter.

[brob]

Is deep learning the only thing for which the machine will be used?

 

What sort of budget are you working with? (Country and currency helps.)

 

Are there plans to add more gpu?

 

There is no mention of storage in the OP. What is being contemplated?

 

What is the timeframe for acquisition?

 

In general,:

• X99 is a dated platform. It has been largely replaced by X299.
• Memory size is inadequate. Faster memory would offer better performance.
• i7-8700/i7-8700K would likely have similar or better performance than the Xeon e5-1650 v4. But it has memory limitations which may make X299, (or anticipated soon to be released sucessor), a better platform.
• If very long runs are contemplated, ECC memory should be considered. In which case motherboard and cpu choices will change.

7 minutes ago, __Peter__ said:

I read that the xeon chips are better for long time processing (several days) and the Core i-Series may lead to errors at such long runtimes. Is that right?

Not the cpu but memory. See above note on ecc (error correcting code) memory.

[__Peter__]

4 minutes ago, brob said:

Is deep learning the only thing for which the machine will be used?

 

What sort of budget are you working with? (Country and currency helps.)

 

Are there plans to add more gpu?

 

There is no mention of storage in the OP. What is being contemplated?

 

What is the timeframe for acquisition?

 

In general,:

• X99 is a dated platform. It has been largely replaced by X299.
• Memory size is inadequate. Faster memory would offer better performance.
• i7-8700/i7-8700K would likely have similar or better performance than the Xeon e5-1650 v4. But it has memory limitations which may make X299, (or anticipated soon to be released sucessor), a better platform.
• If very long runs are contemplated, the ECC memory should probably be considered. In which case motherboard and cpu choices will change.

Not the cpu but memory. See above note on ecc (error correcting code) memory.

The machine is used for developing DL models. That means testing, debugging, etc. But for a large tested network other services such as AWS are preferred. So there are no plans to add more GPUs

 

Budget is about 1500 - 2000 € without the graphic card.

 

For storage we will use a normal SSD disk, nothing special.

 

The timeframe is about one month.

 

Do you mean i can not use ECC memory with this motherboard?

 

 

[brob]

The OP motherboard does support ECC memory. The memory linked in the OP is not ECC.

 

Also, the motherboard supports quad channel memory operation and has eight slots. When buying memory an optimal kit would consist of four or eight matched memory modules. See the motherboard's support page for memory QVL that list tested memory modules.

[Razor Blade]

13 minutes ago, __Peter__ said:

Spoiler

 

The machine is used for developing DL models. That means testing, debugging, etc. But for a large tested network other services such as AWS are preferred. So there are no plans to add more GPUs

 

Budget is about 1500 - 2000 € without the graphic card.

 

For storage we will use a normal SSD disk, nothing special.

 

The timeframe is about one month.

 

Do you mean i can not use ECC memory with this motherboard?

Asus's website says that motherboard is compatible with ECC RAM.

Quote

8 x DIMM, Max. 128GB, DDR4 3200(O.C.)/3000(O.C.)/2800(O.C.)/2666(O.C.)/2400(O.C.)/2133 MHz Non-ECC, Un-buffered Memory *¹
8 x DIMM, Max. 128GB, DDR4 1866*/2400/2133 MHz ECC, Un-buffered Memory *²
8 x DIMM, Max. 128GB, DDR4 1866*/2400/2133 MHz ECC, Register Memory *²
Quad Channel Memory Architecture 
Supports Intel® Extreme Memory Profile (XMP)
** Refer to www.asus.com for the Memory QVL (Qualified Vendors Lists).
* Hyper DIMM support is subject to the physical characteristics of individual CPUs. 

 

I believe @brob was saying that if this computer is going to run long sessions and computer downtime is unacceptable, ECC memory would be a good choice. If you choose to use ECC memory you will need both a motherboard and CPU that support it. In your case both the CPU and motherboard you chose DO support ECC memory, however the memory you chose is not ECC.

[__Peter__]

17 hours ago, lacion said:

I never experienced that before. core i9 is designed for workstation use, and most time workstations do heavy workloads for extended times.

 

I have a 7900x that I use as a development workstation and a 7980XE that doe several workloads including  ML|CV training sessions and other workloads that take a few days sometimes.

 

xeons are really just designed to keep power usage and heat on check so you can stack shit loads of them in a rack and not require a nuclear reactor to cool your datacenter.

When I don't want to change the mainboard, is a good Core-i 7 better than a xeon for my use case?

[lacion]

2 hours ago, __Peter__ said:

When I don't want to change the mainboard, is a good Core-i 7 better than a xeon for my use case?

in your case the more cores the better, machine learning scales way better with parallelism than core clock.

[__Peter__]

2 minutes ago, lacion said:

in your case the more cores the better, machine learning scales way better with parallelism than core clock.

Is the CPU so important for parallelism. I thought this happens all on the GPU?

[brob]

5 hours ago, __Peter__ said:

When I don't want to change the mainboard, is a good Core-i 7 better than a xeon for my use case?

If you want to use ECC memory then you are pretty much restricted to Xeon cpu on that motherboard.

While gpu do most of the heavy work, one still needs adequate cpu resources to keep the gpu "fed". The nature of the software that will be used really determines if more or faster cores are optimal. While one might generalize, a better result will be obtained by learning how the software utilizes the cpu.

[__Peter__]

2 hours ago, brob said:

If you want to use ECC memory then you are pretty much restricted to Xeon cpu on that motherboard.

While gpu do most of the heavy work, one still needs adequate cpu resources to keep the gpu "fed". The nature of the software that will be used really determines if more or faster cores are optimal. While one might generalize, a better result will be obtained by learning how the software utilizes the cpu.

Ok, then I will buy the Xeon e5-1650 v4. @lacion You sayed that the Xeon CPUs does no produce this mcuh energy. I found this fan. Is it already an overkill?

 

[brob]

The Xeon E5-1650 v4 is rated at 140W TDP (thermal design power). It needs a fairly robust cooler.

 

The Dark Rock Pro 3 cooler will do a good job. However I believe it will interfere with the 1st PCIe slot. This is a general problem with larger coolers on the X99-WS/USB 3.1. The NH-U14S has the same problem.

 

The Noctua NH-U12S is compatible with the motherboard and is rated up to 165W TDP.

[lacion]

5 hours ago, __Peter__ said:

Is the CPU so important for parallelism. I thought this happens all on the GPU?

you can use everything, nothing is limiting you, also there is more to machine learning that just the training data and sessions

[__Peter__]

24 minutes ago, brob said:

The Xeon E5-1650 v4 is rated at 140W TDP (thermal design power). It needs a fairly robust cooler.

 

The Dark Rock Pro 3 cooler will do a good job. However I believe it will interfere with the 1st PCIe slot. This is a general problem with larger coolers on the X99-WS/USB 3.1. The NH-U14S has the same problem.

 

The Noctua NH-U12S is compatible with the motherboard and is rated up to 165W TDP.

Ah ok, thank you very much.

[Crazycatman]

On 5/14/2018 at 12:38 PM, __Peter__ said:

Ah ok, thank you very much.

Wait, will you be updating the CPU if ever needed?

[__Peter__]

21 hours ago, Crazycatman said:

Wait, will you be updating the CPU if ever needed?

You mean because of the recent security issues? I can but I don't have to.

[Crazycatman]

16 minutes ago, __Peter__ said:

You mean because of the recent security issues? I can but I don't have to.

Wdym?

[__Peter__]

Just now, Crazycatman said:

Wdym?

Spectre and meltdown or do you mean to upgrade to CPU by replacing it.

For the Spectre and metldown patch you have to flash the CPU temp memory...

[Crazycatman]

3 minutes ago, __Peter__ said:

Spectre and meltdown or do you mean to upgrade to CPU by replacing it.

For the Spectre and metldown patch you have to flash the CPU temp memory...

Oh forgot about the security stuff, honestly didn't care about it. I mean't if you have a low tier CPU were you upgrading it ever in the future?