Jump to content

Build for AI Computing. Any input is appreciated.

Wolololo

I am planning to build an rig for not only AI computing, but also for a bit of graphic rendering, hence the Threadripper 3990x.

This isn't for me but for someone else, and from what I know he will be primarily working with neural networks using TensorFlow. 

Aesthetics aren't really an issue here, only enough cooling and performance are the real considerations. Maybe reducing a bit of noise could be an option.

The question is whether the Noctua cpu cooler will be enough to cool the CPU under full load and the airflow is good enough to cool the GPU under full load. 

 

Specifications:

Motherboard - ASRock TRX40 Creator

CPU - AMD Ryzen Threadripper 3990x

CPU cooler - Noctua NH-U14S TR4-SP3

RAM - Samsung DDR4 32GB PC4-21300 ECC/REG (x8)

SSD - Samsung 970 Pro M.2 2280 1TB

HDD - Western Digital 8TB Ultrastar DC HC510 (x2)

GPU - Zotac Gaming Geforce RTX 2080 Ti D6 11GB BLOWER (x2)

PSU - Corsair AX1600i 80PLUS Titanium

Case - Corsair Obsidian 1000D

Case fan (120mm) - Noctua NF-A12x15 PWM (x10)

Case fan (140mm) - Noctua NF-A15 PWM (x4)

ETC - Nvlink

 

I would've liked to go with Quadro RTX cards, but the price difference between Geforce and RTX cards are even worse in my country (S.Korea).

 

Thanks in advance!

CPU: Ryzen 5 3600 | Mobo: ASUS ROG Strix X570-I Gaming | RAM: 16GB 3200 | GPU: MSI RX 5700 Gaming X 8GB | Case: Fractal Design Meshify C | OS: Win 10 | Storage: 250GB WD Blue SSD & 2TB Seagate BarraCuda  | PSU: Corsair HX750i | Display: Hansung TFG32Q14V 1440p @144Hz | Cooling: Fractal Design Celcius S24 Blackout | Keyboard: ABKO Hacker K660  | Mouse: G503

Link to comment
Share on other sites

Link to post
Share on other sites

7 hours ago, polo191 said:

I am planning to build an rig for not only AI computing, but also for a bit of graphic rendering, hence the Threadripper 3990x.

This isn't for me but for someone else, and from what I know he will be primarily working with neural networks using TensorFlow. 

Aesthetics aren't really an issue here, only enough cooling and performance are the real considerations. Maybe reducing a bit of noise could be an option.

The question is whether the Noctua cpu cooler will be enough to cool the CPU under full load and the airflow is good enough to cool the GPU under full load. 

 

Specifications:

Motherboard - ASRock TRX40 Creator

CPU - AMD Ryzen Threadripper 3990x

CPU cooler - Noctua NH-U14S TR4-SP3

RAM - Samsung DDR4 32GB PC4-21300 ECC/REG (x8)

SSD - Samsung 970 Pro M.2 2280 1TB

HDD - Western Digital 8TB Ultrastar DC HC510 (x2)

GPU - Zotac Gaming Geforce RTX 2080 Ti D6 11GB BLOWER (x2)

PSU - Corsair AX1600i 80PLUS Titanium

Case - Corsair Obsidian 1000D

Case fan (120mm) - Noctua NF-A12x15 PWM (x10)

Case fan (140mm) - Noctua NF-A15 PWM (x4)

ETC - Nvlink

 

I would've liked to go with Quadro RTX cards, but the price difference between Geforce and RTX cards are even worse in my country (S.Korea).

 

Thanks in advance!

Your overall selection sounds really good. Remember to add another fan on that NH-U14S.

 

Make sure that 11GB is enough for his needs, or that their models can easily be split into 2 GPUs (giving 22GB total of vram). If not, going for a single Quadro card would be the best option since those are the only way to get over 20GB of vram in a single card, even if you lose on raw compute power.

 

I've heard that the Obsidian 1000d is perfect for water cooling and whatnot, but I'm not sure about air-cooling. Bayne give a look at some Fractal cases or a MasterBox from Cooler Master?

FX6300 @ 4.2GHz | Gigabyte GA-78LMT-USB3 R2 | Hyper 212x | 3x 8GB + 1x 4GB @ 1600MHz | Gigabyte 2060 Super | Corsair CX650M | LG 43UK6520PSA
ASUS X550LN | i5 4210u | 12GB
Lenovo N23 Yoga

Link to comment
Share on other sites

Link to post
Share on other sites

19 hours ago, igormp said:

Your overall selection sounds really good. Remember to add another fan on that NH-U14S.

 

Make sure that 11GB is enough for his needs, or that their models can easily be split into 2 GPUs (giving 22GB total of vram). If not, going for a single Quadro card would be the best option since those are the only way to get over 20GB of vram in a single card, even if you lose on raw compute power.

 

I've heard that the Obsidian 1000d is perfect for water cooling and whatnot, but I'm not sure about air-cooling. Bayne give a look at some Fractal cases or a MasterBox from Cooler Master?

Thank you so much for your input.

I definitely included that additional fan in the spec sheet. Thanks for the reminder :) 

 

I'd like nothing more than to go with a single Quadro RTX 6000 (lower end Quadros have insignificant difference in VRAM than 2080Ti), that is only if I had another couple thousand dollars. But, the fact is that a single Quadro 6000 is nearly 1.5x the price of 2 2080Tis in my country. For now, I think 2 2080Tis linked through Nvlink should suffice.

 

Regarding the Obsidian 1000d, I am planning to have 8x 120mm fans in the front for intake and 2x 120mm (back), 3x 140mm (top) for exhaust. Since the case is huge and therefore has lots of room for airflow, I was thinking that the air cooling should be good enough to cool the components once under load. I'd love your input on this. 

 

 

CPU: Ryzen 5 3600 | Mobo: ASUS ROG Strix X570-I Gaming | RAM: 16GB 3200 | GPU: MSI RX 5700 Gaming X 8GB | Case: Fractal Design Meshify C | OS: Win 10 | Storage: 250GB WD Blue SSD & 2TB Seagate BarraCuda  | PSU: Corsair HX750i | Display: Hansung TFG32Q14V 1440p @144Hz | Cooling: Fractal Design Celcius S24 Blackout | Keyboard: ABKO Hacker K660  | Mouse: G503

Link to comment
Share on other sites

Link to post
Share on other sites

3 hours ago, polo191 said:

Thank you so much for your input.

I definitely included that additional fan in the spec sheet. Thanks for the reminder :) 

 

I'd like nothing more than to go with a single Quadro RTX 6000 (lower end Quadros have insignificant difference in VRAM than 2080Ti), that is only if I had another couple thousand dollars. But, the fact is that a single Quadro 6000 is nearly 1.5x the price of 2 2080Tis in my country. For now, I think 2 2080Tis linked through Nvlink should suffice.

Ouch. In this case, I guess there's nothing much you can do anyway. The 2x 2080Tis should really be more than enough anyway :)

 

3 hours ago, polo191 said:

Regarding the Obsidian 1000d, I am planning to have 8x 120mm fans in the front for intake and 2x 120mm (back), 3x 140mm (top) for exhaust. Since the case is huge and therefore has lots of room for airflow, I was thinking that the air cooling should be good enough to cool the components once under load. I'd love your input on this. 

I'm not really an expert when it comes to cooling, but I guess this should be fine as long as air can come in and out of the case itself, instead of recirculating inside of it.

 

Also, another thing: wouldn't it be a good idea to have a regular 2.5" SSD to use as a scratch disk for faster data manipulation? I guess that could also be done on the NVMe, but who knows.

FX6300 @ 4.2GHz | Gigabyte GA-78LMT-USB3 R2 | Hyper 212x | 3x 8GB + 1x 4GB @ 1600MHz | Gigabyte 2060 Super | Corsair CX650M | LG 43UK6520PSA
ASUS X550LN | i5 4210u | 12GB
Lenovo N23 Yoga

Link to comment
Share on other sites

Link to post
Share on other sites

20 hours ago, igormp said:

Ouch. In this case, I guess there's nothing much you can do anyway. The 2x 2080Tis should really be more than enough anyway :)

Yeah, I wish I had a bigger budget... I mean I'm building a high-end system after all.. I've seen some tensorflow performance benchmarks on 2x 2080Tis linked with NVlink and it seems like it's the best for those who want to go with a consumer-grade option instead of Quadro cards. Check this out: https://www.pugetsystems.com/labs/hpc/RTX-2080Ti-with-NVLINK---TensorFlow-Performance-Includes-Comparison-with-GTX-1080Ti-RTX-2070-2080-2080Ti-and-Titan-V-1267/

 

Quote

I'm not really an expert when it comes to cooling, but I guess this should be fine as long as air can come in and out of the case itself, instead of recirculating inside of it.

 

Also, another thing: wouldn't it be a good idea to have a regular 2.5" SSD to use as a scratch disk for faster data manipulation? I guess that could also be done on the NVMe, but who knows.

Hmm, a very rudimentary calculation of the airflow generated by 8x Noctua NF-A12x15 PWM 120m fans seem to be more or less equal to a medium-sized Vornado fan set to max speed. (I have no idea whether if this is the right way and there are probably some factors missing; I did this by getting the CFM of each 120mm fan and multiplying by 8 and comparing it to a Vornado 633 Medium Air Circulator). With less CFM as exhaust, it should constitute a positive pressure environment, but a problem could be the large difference in Total_IntakeCFM - Total_ExhaustCFM, possibly causing air to recirculate? Openings in the case (especially in a case this big) and the cooling of the recirculating air might play a role in cooling, but then again, I have no background in aerodynamics and such. What's your thoughts on this?

 

Yup, that's a good idea. Preprocessing data outside the model could be better with a dedicated drive. I know for certain that it can be done on the NVMe, so I guess I should add a 500GB NVMe. Thanks for the suggestion!

CPU: Ryzen 5 3600 | Mobo: ASUS ROG Strix X570-I Gaming | RAM: 16GB 3200 | GPU: MSI RX 5700 Gaming X 8GB | Case: Fractal Design Meshify C | OS: Win 10 | Storage: 250GB WD Blue SSD & 2TB Seagate BarraCuda  | PSU: Corsair HX750i | Display: Hansung TFG32Q14V 1440p @144Hz | Cooling: Fractal Design Celcius S24 Blackout | Keyboard: ABKO Hacker K660  | Mouse: G503

Link to comment
Share on other sites

Link to post
Share on other sites

7 hours ago, polo191 said:

Yeah, I wish I had a bigger budget... I mean I'm building a high-end system after all.. I've seen some tensorflow performance benchmarks on 2x 2080Tis linked with NVlink and it seems like it's the best for those who want to go with a consumer-grade option instead of Quadro cards. Check this out: https://www.pugetsystems.com/labs/hpc/RTX-2080Ti-with-NVLINK---TensorFlow-Performance-Includes-Comparison-with-GTX-1080Ti-RTX-2070-2080-2080Ti-and-Titan-V-1267/

I've seen that post before, the nvlink on geforce cards really doesn't bring much to the table, but I guess another extra 5% of performance might be worth anyway. In the end, both 2080Tis will sure beat the crap out of any Quadro compute wise.

 

7 hours ago, polo191 said:

Hmm, a very rudimentary calculation of the airflow generated by 8x Noctua NF-A12x15 PWM 120m fans seem to be more or less equal to a medium-sized Vornado fan set to max speed. (I have no idea whether if this is the right way and there are probably some factors missing; I did this by getting the CFM of each 120mm fan and multiplying by 8 and comparing it to a Vornado 633 Medium Air Circulator). With less CFM as exhaust, it should constitute a positive pressure environment, but a problem could be the large difference in Total_IntakeCFM - Total_ExhaustCFM, possibly causing air to recirculate? Openings in the case (especially in a case this big) and the cooling of the recirculating air might play a role in cooling, but then again, I have no background in aerodynamics and such. What's your thoughts on this?

I'm really not that knowledgeable when it comes to airflow things inside of a case, so I believe you're better off with the guys at the air cooling subforum haha

FX6300 @ 4.2GHz | Gigabyte GA-78LMT-USB3 R2 | Hyper 212x | 3x 8GB + 1x 4GB @ 1600MHz | Gigabyte 2060 Super | Corsair CX650M | LG 43UK6520PSA
ASUS X550LN | i5 4210u | 12GB
Lenovo N23 Yoga

Link to comment
Share on other sites

Link to post
Share on other sites

13 minutes ago, igormp said:

I've seen that post before, the nvlink on geforce cards really doesn't bring much to the table, but I guess another extra 5% of performance might be worth anyway. In the end, both 2080Tis will sure beat the crap out of any Quadro compute wise.

You're right. Since I can't go with the quadro route, I might as well squeeze the heck out of the 2080Tis anyway I can. 

13 minutes ago, igormp said:

I'm really not that knowledgeable when it comes to airflow things inside of a case, so I believe you're better off with the guys at the air cooling subforum haha

Ahaha yeah, you seemed knowledgeable so I couldn't hesitate asking. I'd better do that.

 

Thanks I really appreciate the input! 

 

CPU: Ryzen 5 3600 | Mobo: ASUS ROG Strix X570-I Gaming | RAM: 16GB 3200 | GPU: MSI RX 5700 Gaming X 8GB | Case: Fractal Design Meshify C | OS: Win 10 | Storage: 250GB WD Blue SSD & 2TB Seagate BarraCuda  | PSU: Corsair HX750i | Display: Hansung TFG32Q14V 1440p @144Hz | Cooling: Fractal Design Celcius S24 Blackout | Keyboard: ABKO Hacker K660  | Mouse: G503

Link to comment
Share on other sites

Link to post
Share on other sites

For this kind of system maybe consider a Gen4 nvme ssd? Possibly something like a Corsair mp600 or even a Sabrent Rocket? 

 

Obsidian 1000 is massive overkill for an air-cooled system. That case is meant for dual loop custom watercooling with giant radiators. 

 

The 1600w PSU is also slightly ott. Your system should pull around 900 to to 950w under full load, so unless you are overclocking heavily, a 1200w psu should be plenty. Of course there's nothing wrong with having the extra headroom a 1600w would provide. 

Link to comment
Share on other sites

Link to post
Share on other sites

10 hours ago, ErA.l said:

For this kind of system maybe consider a Gen4 nvme ssd? Possibly something like a Corsair mp600 or even a Sabrent Rocket? 

I'm not really sure whether there will much benefit from a faster storage when using Tensorflow. I mean Samsung 970 Pros are fast as they are. I think some questions to consider would be whether they have a performance increase worth the extra dollar? And are they stable enough?

Do you think you can help me better understand your reasoning?

 

10 hours ago, ErA.l said:

Obsidian 1000 is massive overkill for an air-cooled system. That case is meant for dual loop custom watercooling with giant radiators. 

 

The 1600w PSU is also slightly ott. Your system should pull around 900 to to 950w under full load, so unless you are overclocking heavily, a 1200w psu should be plenty. Of course there's nothing wrong with having the extra headroom a 1600w would provide. 

I actually have a reason for this, I saw that the Corsair 1000D supports other larger sized motherboards as well, namely the SSI EEB and EATX classes. The person that I'm building this for also said that there might be a need to convert it into a dual core workstation, so it'll be future proof in case that happens. Plus, I know for sure that he's planning to buy Quadro RTX cards further down the line so in that case I may have to implement a water cooling solution. Also, I think it justifies the use of 1600w PSUs as well (3990x + 2x 2080Tis + 2x Quadro RTX 6 or 8000s)?

 

 

CPU: Ryzen 5 3600 | Mobo: ASUS ROG Strix X570-I Gaming | RAM: 16GB 3200 | GPU: MSI RX 5700 Gaming X 8GB | Case: Fractal Design Meshify C | OS: Win 10 | Storage: 250GB WD Blue SSD & 2TB Seagate BarraCuda  | PSU: Corsair HX750i | Display: Hansung TFG32Q14V 1440p @144Hz | Cooling: Fractal Design Celcius S24 Blackout | Keyboard: ABKO Hacker K660  | Mouse: G503

Link to comment
Share on other sites

Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

×