Jump to content

Hi all. I'm a machine learning researcher who wants to build a pc for the first time. My primary use case will be training ML models, and therefore using multiple gpu's is extremely efficient. I'd like to get two RTX 3080's; however, I'm not sure exactly how to go about it. Would two founders editions be infeasible due to airflow? What I mean is that for two cards plugged in next to each other, the hot air blowing out of one would seem to result in the second overheating. Secondly, since I would have them running for potentially weeks at a time, I worry about the longevity of such an approach. Would custom watercooling be the only solution in this case? And if not, what would people suggest I do?

 

Budget (including currency): 3500 USD

 

Country: 

USA (Massachusetts)

Games, programs or workloads that it will be used for: 

 

JAX (https://github.com/google/jax) XLA compiled ML model training. 

 

Other details (existing parts lists, whether any peripherals are needed, what you're upgrading from, when you're going to buy, what resolution and refresh rate you want to play at, etc): 

Current part list:

-Fractal Design Meshify C Tempered Glass

-MSI Meg X570 Unify Motherboard (https://www.amazon.com/gp/product/B07ZG6YG1K/ref=ox_sc_act_title_3?smid=A1FF9ODJZIUKGX&psc=1)

-AMD Ryzen 9 3900X

-Corsair AXi Series, AX1600i (https://www.amazon.com/gp/product/B0787G1Z74/ref=ox_sc_act_title_5?smid=ANNSEW49C6U7Y&th=1)

-Corsair Vengeance RGB Pro 64GB (2x32GB) (https://www.amazon.com/gp/product/B081BTGNDW/ref=ox_sc_act_title_8?smid=ATVPDKIKX0DER&psc=1)

-2x RTX 3080

Link to comment
https://linustechtips.com/topic/1243460-building-a-dual-gpu-system/
Share on other sites

Link to post
Share on other sites

2 hours ago, RenOS said:

Hi all. I'm a machine learning researcher who wants to build a pc for the first time. My primary use case will be training ML models, and therefore using multiple gpu's is extremely efficient. I'd like to get two RTX 3080's; however, I'm not sure exactly how to go about it. Would two founders editions be infeasible due to airflow? What I mean is that for two cards plugged in next to each other, the hot air blowing out of one would seem to result in the second overheating. Secondly, since I would have them running for potentially weeks at a time, I worry about the longevity of such an approach. Would custom watercooling be the only solution in this case? And if not, what would people suggest I do?

 

Budget (including currency): 3500 USD

 

Country: 

USA (Massachusetts)

Games, programs or workloads that it will be used for: 

 

JAX (https://github.com/google/jax) XLA compiled ML model training. 

 

Other details (existing parts lists, whether any peripherals are needed, what you're upgrading from, when you're going to buy, what resolution and refresh rate you want to play at, etc): 

Current part list:

-Fractal Design Meshify C Tempered Glass

-MSI Meg X570 Unify Motherboard (https://www.amazon.com/gp/product/B07ZG6YG1K/ref=ox_sc_act_title_3?smid=A1FF9ODJZIUKGX&psc=1)

-AMD Ryzen 9 3900X

-Corsair AXi Series, AX1600i (https://www.amazon.com/gp/product/B0787G1Z74/ref=ox_sc_act_title_5?smid=ANNSEW49C6U7Y&th=1)

-Corsair Vengeance RGB Pro 64GB (2x32GB) (https://www.amazon.com/gp/product/B081BTGNDW/ref=ox_sc_act_title_8?smid=ATVPDKIKX0DER&psc=1)

-2x RTX 3080

i thought sli was gone?

 

https://www.pcworld.com/article/3573384/rip-sli-nvidia-geforce-rtx-30-series-slams-the-final-nail-in-multi-gpus-coffin.html

I have dyslexia plz be kind to me. dont like my post dont read it or respond thx

also i edit post alot because you no why...

Thrasher_565 hub links build logs

 

Link to post
Share on other sites

9 hours ago, thrasher_565 said:

SLI is only needed for gaming - machine learning doesn't use SLI.

Link to post
Share on other sites

12 hours ago, RenOS said:

snip

 

sli only works with the rtx 3090. buy that instead and add a second gpu later if you need.

6 minutes ago, RenOS said:

SLI is only needed for gaming - machine learning doesn't use SLI.

what? is is the other way around. games do not use sli anymore. well some do, but usually the performance is lower with 2 gpus than with single gpu.

 

12 hours ago, RenOS said:

What I mean is that for two cards plugged in next to each other, the hot air blowing out of one would seem to result in the second overheating. Secondly, since I would have them running for potentially weeks at a time, I worry about the longevity of such an approach. Would custom watercooling be the only solution in this case? And if not, what would people suggest I do?

if the temps are under 100c longetivity is not the problem. but since rtx 3080 does not use sli you should go with single rtx 3090 instead.

 

12 hours ago, RenOS said:

Current part list:

so you already have those?

 

with that budget get  rtx 3090. and good fans in the case. like noctua redux or be quiet silent wings 3.

 

but if yo udo not have those:

much better cpu and motherboard

better case. good fans. 2x on the top and 1 to the rear for exhaust

for single gpu good 850w psu is more than plenty.

I Use my knowledge as business owner and self taught technician aswell as an AI to help people. AI might be controversial but it actually works pretty well 90% of the time.

Link to post
Share on other sites

27 minutes ago, SavageNeo said:

sli only works with the rtx 3090. buy that instead and add a second gpu later if you need.

what? is is the other way around. games do not use sli anymore. well some do, but usually the performance is lower with 2 gpus than with single gpu.

 

if the temps are under 100c longetivity is not the problem. but since rtx 3080 does not use sli you should go with single rtx 3090 instead.

 

so you already have those?

 

with that budget get  rtx 3090. and good fans in the case. like noctua redux or be quiet silent wings 3.

 

but if yo udo not have those:

much better cpu and motherboard

better case. good fans. 2x on the top and 1 to the rear for exhaust

for single gpu good 850w psu is more than plenty.

Thank you for the advice! Just for your information, the JAX / Tensorflow optimizing compiler doesn't care about having SLI. It's typical to be utilizing 4/8 gpu's in computing clusters, so having SLI be a requirement would be completely infeasible.

 

 

Also: I rejected using a CPU air cooler due style of the 3080 cooling. From what I've read, it'll be blowing hot air near the cpu, making air coolers much more inefficient. 

Link to post
Share on other sites

1 hour ago, SavageNeo said:

sli only works with the rtx 3090. buy that instead and add a second gpu later if you need.

what? is is the other way around. games do not use sli anymore. well some do, but usually the performance is lower with 2 gpus than with single gpu.

SLI has NEVER been a thing for compute-focused workloads. Those can make use of multi gpu in anyway you can imagine, even using 2 different GPUs will work. NVLink adds about 5~10% performance due to the extra bandwidth for crosstalk between cards.

 

14 hours ago, RenOS said:

I'd like to get two RTX 3080's; however, I'm not sure exactly how to go about it. Would two founders editions be infeasible due to airflow? What I mean is that for two cards plugged in next to each other, the hot air blowing out of one would seem to result in the second overheating.

Those cards are still somewhat blower-like, I guess. The fan on the back works exactly like a regular blower-style GPU, so I guess those will work fine right next to the other. If you're still worried, there'll probably be blower-style cards made by partners later on, so you could just wait.

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 post
Share on other sites

The AX1600 psu is too long for the Meshify C case.

 

You might consider a workstation motherboard. No flash, designed for heavy workloads.

 

At the very least you will want to consider cooling available on non reference design gpu. The ideal would be liquid cooling for the cpu and all gpu.

 

PCPartPicker Part List

CPU: AMD Ryzen 9 3900X 3.8 GHz 12-Core Processor  ($429.99 @ B&H) 
Motherboard: Asus Pro WS X570-ACE ATX AM4 Motherboard  ($372.33 @ Amazon) 
Memory: Crucial Ballistix 64 GB (2 x 32 GB) DDR4-3600 CL16 Memory  ($389.26 @ Amazon) 
Case: Fractal Design Define 7 ATX Mid Tower Case  ($169.99 @ Walmart) 
Power Supply: Corsair 1600 W 80+ Titanium Certified Fully Modular ATX Power Supply  ($492.00 @ Amazon) 
Total: $1853.57
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2020-09-06 14:42 EDT-0400

80+ ratings certify electrical efficiency. Not quality.

 

Link to post
Share on other sites

2 hours ago, brob said:

The AX1600 psu is too long for the Meshify C case.

Should I find a different case? I'm also considering the https://www.thermaltakeusa.com/level-20-xt.html for future watercooling support and also having enough space to hold two 3080's. I'm considering using an extender cable to bring down the second 3080 down to the bottom of the case.

Link to post
Share on other sites

11 hours ago, RenOS said:

Should I find a different case? I'm also considering the https://www.thermaltakeusa.com/level-20-xt.html for future watercooling support and also having enough space to hold two 3080's. I'm considering using an extender cable to bring down the second 3080 down to the bottom of the case.

 

The Define 7 has more than sufficient room. You might also consider some of the Phanteks models.

 

I would not use an extended cable. Both the MEG X570 and Pro WS X570 motherboards have sufficient air space between the first two gpu slots when using double width cards.

80+ ratings certify electrical efficiency. Not quality.

 

Link to post
Share on other sites

8 hours ago, brob said:

I would not use an extended cable. Both the MEG X570 and Pro WS X570 motherboards have sufficient air space between the first two gpu slots when using double width cards.

This would be true for blower cards; however, I'm worried about the new 3080 style of cooling where one card would blow into the other.

Link to post
Share on other sites

7 hours ago, RenOS said:

This would be true for blower cards; however, I'm worried about the new 3080 style of cooling where one card would blow into the other.

I think you are needlessly worrying. There is a full slot width open between the first and second dual width gpu. One just has to insure good airflow through the case. But as I said in my initial post the ideal would be to use liquid/hybrid cooled gpu. 

80+ ratings certify electrical efficiency. Not quality.

 

Link to post
Share on other sites

8 hours ago, RenOS said:

This would be true for blower cards; however, I'm worried about the new 3080 style of cooling where one card would blow into the other.

The new cards still have a kind of blower-like cooler that pushes air out of the case. The card on top will always be hotter than the 2nd one, no matter what you do, but in this case I'd say that the end result is more than acceptable.

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 post
Share on other sites

  • 1 month later...
On 9/6/2020 at 8:34 AM, RenOS said:

Hi all. I'm a machine learning researcher who wants to build a pc for the first time. My primary use case will be training ML models, and therefore using multiple gpu's is extremely efficient. I'd like to get two RTX 3080's; however, I'm not sure exactly how to go about it. Would two founders editions be infeasible due to airflow? What I mean is that for two cards plugged in next to each other, the hot air blowing out of one would seem to result in the second overheating. Secondly, since I would have them running for potentially weeks at a time, I worry about the longevity of such an approach. Would custom watercooling be the only solution in this case? And if not, what would people suggest I do?

 

Budget (including currency): 3500 USD

 

Country: 

USA (Massachusetts)

Games, programs or workloads that it will be used for: 

 

JAX (https://github.com/google/jax) XLA compiled ML model training. 

 

Other details (existing parts lists, whether any peripherals are needed, what you're upgrading from, when you're going to buy, what resolution and refresh rate you want to play at, etc): 

Current part list:

-Fractal Design Meshify C Tempered Glass

-MSI Meg X570 Unify Motherboard (https://www.amazon.com/gp/product/B07ZG6YG1K/ref=ox_sc_act_title_3?smid=A1FF9ODJZIUKGX&psc=1)

-AMD Ryzen 9 3900X

-Corsair AXi Series, AX1600i (https://www.amazon.com/gp/product/B0787G1Z74/ref=ox_sc_act_title_5?smid=ANNSEW49C6U7Y&th=1)

-Corsair Vengeance RGB Pro 64GB (2x32GB) (https://www.amazon.com/gp/product/B081BTGNDW/ref=ox_sc_act_title_8?smid=ATVPDKIKX0DER&psc=1)

-2x RTX 3080

I think you should buy a 3090 instead because it has 24 gigs of VRAM which you'll end up using for AI/ML related tasks. You can always upgrade later :)

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

×