RTX 3090 SLI - We Tried so Hard to Love It

[AlexTheGreatish]

 

Officially Nvidia has killed SLI but the 3090 has SLI fingers... lets give it a go.

 

Buy ASUS ROG Strix GeForce RTX 3090

On Amazon (PAID LINK): https://geni.us/XDIki

On Newegg (PAID LINK): https://geni.us/TNeCJ

On B&H (PAID LINK): https://geni.us/wEVs

 

Buy MSI MEG Z490 GODLIKE

On Amazon (PAID LINK): https://geni.us/jDxY1VT

On Newegg (PAID LINK): https://geni.us/hpgEWa

On B&H (PAID LINK): https://geni.us/hry1

 

Buy Intel Core i9-10900K Processor

On Amazon (PAID LINK): https://geni.us/rJ3e

On Newegg (PAID LINK): https://geni.us/OqUA

On B&H (PAID LINK): https://geni.us/wtzSjn

 

Buy Seasonic PRIME TX-1000 PSU

On Amazon (PAID LINK): https://geni.us/7SfX

On Newegg (PAID LINK): https://geni.us/W6wkoAm

On B&H (PAID LINK): https://geni.us/5cxl

 

Buy Corsair AX1600i PSU

On Amazon (PAID LINK): https://geni.us/vWYWVT

On Newegg (PAID LINK): https://geni.us/wMGJ

On B&H (PAID LINK): https://geni.us/xUWb

 

Purchases made through some store links may provide some compensation to Linus Media Group.

 

 

 

[riba2233]

3:28

 

"I love Ryzen as much as anyone does at this point, but if you are running 2 RTX3090's, you are going to need the fastest possible gaming cpu, which for now anyway is 10900K".

 

Ok, maybe, but apparently you are fine with 1/4 of max bandwidth for each card?

 

BTW I died when I saw Alex laughing, so honest, his face got really red

 

[jones177]

I retired my SLI setup but I have a 1300 watt PSU just in case I want to at least test it in the future.

[HenrySalayne]

4 minutes ago, riba2233 said:

Ok, maybe, but apparently you are fine with 1/4 of max bandwidth for each card?

Fair point. So PCIe 3.0 x16 is still enough, but will 3.0 x8 hurt the performance? Could be tested without SLI TBH.

[Drama Lama]

SLI has been slowly dying for years

 

SLI is dead Long live SLI

[Metallic-Lighght]

When even the manufacturer needs you to send a card back to them because of the shortage of the cards, we have an issue. 

 

Spoiler

Nvidia, get yourself together.

 

[Nystemy]

Have been using a dual GPU setup for years. But have never actually configured them to run in SLI...

But to a degree, I have to agree with Nvidia, since they have made the statement that it should be up to the game if it runs multiple GPUs or not, since having explicit control over it is much better than just brute forcing it like SLI traditionally does.

Though, that doesn't change the fact that having a bridge between the cards to toss resources back and forth is a good idea. After all, more bandwidth is usually not a bad thing when it comes to linking together processing units, be it GPUs or CPUs. _{Kinda wish that motherboard manufacturers could repurpose some of the PCIe lanes for inter CPU connectivity. (Ie, use the lanes for QPI/UPI or Infinity Fabric.) Since that could make some multi socketed systems a lot more interesting for some applications.}

[bungbloopity]

Anyone know what case he was being used in the video? Thanks!

[riba2233]

2 minutes ago, bungbloopity said:

Anyone know what case he was being used in the video? Thanks!

O11 dynamic razer edition?

[bungbloopity]

2 minutes ago, riba2233 said:

O11 dynamic razer edition?

Yep thats the one! Thanks!

[Luscious]

I had a similar power problem with a CM V1200. My 4-way 980Ti system pulls exactly 1140W from the wall under load or just a little over 1000W actual PSU load. For a 1200W PSU to crap it's pants at 1000W is a massive fail - shame on you Cooler Master.

 

After figuring out the cause of my system problem I unceremoniously dumped that POS and installed an EVGA 1600W T2. Runs like a champ without a hiccup and have not had a single power problem ever since. Two thumbs up for EVGA.

 

Regarding SLI - having multiple graphics cards in a system has more benefits than just gaming. Being able to pool the VRAM together can have benefits when running simulations. If your software allows for configuring additional graphics cards that can also take advantage. Maybe you are in a situation where you need to transcode a massive library of video footage but don't want to tie up your main GPU for multitasking, or maybe you want to do work and game at the same time.

 

In my case F@H is the main reason I have 4 cards, since even when I am folding on three I can still game. Then when I am not gaming I can push all four and have it run overnight as well. It's all water cooled with massive radiators with lots of fans so temperatures aren't an issue. The case I am using (CaseLabs Mercury S8 + pedestal) is also perfectly suited for exactly such a multi-card setup.

 

I would be hesitant about dropping in four water cooled 3090's if I was building today though, just because you will definitely need a second PSU to even get it to work, and finding a case that can house a 2nd full size PSU and accommodate radiators is tough (RIP CaseLabs). Four 3080's though would be doable with a single 1600W unit if the cards have just 2x8pin.

[The1Dickens]

@AlexTheGreatish Any plans to retest this setup when more games come out that utilize that GPU streaming feature? Or are you folks over it?

[Helpful Tech Witch]

8 hours ago, riba2233 said:

Ok, maybe, but apparently you are fine with 1/4 of max bandwidth for each card?

 

the bottleneck on ryzen is worse that the 1-2% lower scores/fps for the pcie 3rd gen.

[dogwitch]

4 hours ago, Luscious said:

I had a similar power problem with a CM V1200. My 4-way 980Ti system pulls exactly 1140W from the wall under load or just a little over 1000W actual PSU load. For a 1200W PSU to crap it's pants at 1000W is a massive fail - shame on you Cooler Master.

 

After figuring out the cause of my system problem I unceremoniously dumped that POS and installed an EVGA 1600W T2. Runs like a champ without a hiccup and have not had a single power problem ever since. Two thumbs up for EVGA.

 

Regarding SLI - having multiple graphics cards in a system has more benefits than just gaming. Being able to pool the VRAM together can have benefits when running simulations. If your software allows for configuring additional graphics cards that can also take advantage. Maybe you are in a situation where you need to transcode a massive library of video footage but don't want to tie up your main GPU for multitasking, or maybe you want to do work and game at the same time.

 

In my case F@H is the main reason I have 4 cards, since even when I am folding on three I can still game. Then when I am not gaming I can push all four and have it run overnight as well. It's all water cooled with massive radiators with lots of fans so temperatures aren't an issue. The case I am using (CaseLabs Mercury S8 + pedestal) is also perfectly suited for exactly such a multi-card setup.

 

I would be hesitant about dropping in four water cooled 3090's if I was building today though, just because you will definitely need a second PSU to even get it to work, and finding a case that can house a 2nd full size PSU and accommodate radiators is tough (RIP CaseLabs). Four 3080's though would be doable with a single 1600W unit if the cards have just 2x8pin.

i had a spare 1050 its the pysix card now for 2070 installed

[BabaGanuche]

8 hours ago, Nystemy said:

_{Kinda wish that motherboard manufacturers could repurpose some of the PCIe lanes for inter CPU connectivity. (Ie, use the lanes for QPI/UPI or Infinity Fabric.) Since that could make some multi socketed systems a lot more interesting for some applications.}

AMD already does this on their server platform.

 

https://m.youtube.com/watch?v=Oe_hwJ7hl7A

 

[HenrySalayne]

5 hours ago, Luscious said:

I had a similar power problem with a CM V1200. My 4-way 980Ti system pulls exactly 1140W from the wall under load or just a little over 1000W actual PSU load. For a 1200W PSU to crap it's pants at 1000W is a massive fail - shame on you Cooler Master.

Looking at the total power draw from the wall is pretty much pointless. You may still trip the over-current protection on certain voltage rails. And "exactly" is another word I certainly wouldn't use with affordable power meters. Switching power supplies (so basically every modern power supply) especially with quickly changing loads are beyond the capabilities of many "wall wart" power meters when it comes to accuracy.

[PeachGr]

I don't know many things about SLI, but isn't that for blender, editing and things like that? Obviously it's not for gaming

[Nystemy]

3 hours ago, BabaGanuche said:

AMD already does this on their server platform.

 

https://m.youtube.com/watch?v=Oe_hwJ7hl7A

 

Then I'll hope Intel aslo starts doing it. Even if it does add some complexity to the IO system.
But the benefits can outweigh the costs for a lot of applications that are less NUMA aware, or can't easily be optimized for the low bandwidth.

Though, would also be nice if AMD supported more than 2 sockets..... Intel after all goes all the way to 8, even if that is niche and weird, and the fact that a multi node system becomes more logical at that point regardless... But it's nice to not need to drag out a tone of network cards just to have a bunch of CPUs picking away at some latency sensitive, yet parallel task.

[The Sloth]

14 hours ago, AlexTheGreatish said:

Buy ASUS ROG Strix GeForce RTX 3090

 

Your giving me false hope. 

[GoldenLag]

@GabenJr any news on a possible followup in regards to the tripping issues? testing other units, or some sort of followup. as there are known 550w/650w units that will happily work with the 3080/3090. 

 

tbf, i was just waiting for the tx1000 to trip. as it is one of the known models that are sensetive. And to be fair to the PSU makers, this is really not their fault that Ampere spikes high/ is heavy transiense wise. Vega was also known to trip certain PSU models, and its really is just protections working as intended. 

[Rux]

What I really liked about this video was also seeing what happens when the cards draw too much power and the PSU shuts down, not sure why I found this so intriguing 

[riba2233]

15 hours ago, HelpfulTechWizard said:

the bottleneck on ryzen is worse that the 1-2% lower scores/fps for the pcie 3rd gen.

Yeah try with 2nd gen, because that's equal to bandwidth they run them at (gen3 x8 = gen2 x16).

 

 

[BuckGup]

On 10/15/2020 at 12:10 PM, yaboistar said:

this comment was brought to you by the LTT folding team. if you, too, have two 3090's sitting around, consider donating them to science.

I think people would rather mine and make $230/month

[GDRRiley]

can this be re tested with zen3. given they will be the fastest CPUs in a few weeks and have PCIE gen4

[GodAtum]

the testing is all wrong. people who will buy his will have 3 4K monitors. Be nice to see benchmarks for that.