Friday, December 8th 2017

NVIDIA TITAN V Lacks SLI or NVLink Support

Earlier today, we brought you a story about NVIDIA TITAN V setting you back by up to $7,196 for two cards and two $600 NVLink cables. We got word from NVIDIA that the card neither features NVLink, nor supports SLI, and have since edited it. The NVLink fingers on the TITAN V card are rudiments of the functional NVLink interface found on the Tesla V100 PCIe, being developed by NVIDIA, as the TITAN V, Tesla V100, and a future Quadro GV100 share a common PCB. The NVLink fingers on the TITAN V are concealed by the base-plate of the cooler on one side, and the card's back-plate on the other; so the female connectors of NVLink bridge cables can't be plugged in.

With the lack of SLI support on what is possibly it's fastest graphics card based on the "Volta" architecture, NVIDIA seems to have responded to market trends that multi-GPU is dying or dead. That said, it would be interesting to see if professional overclockers chasing benchmark leaderboard glory pick up the TITAN V, as opposed to two TITAN Xp in SLI or four Radeon RX Vega 64 in 4-way CrossFireX.
Add your own comment

49 Comments on NVIDIA TITAN V Lacks SLI or NVLink Support

#26
Captain_Tom
Th3pwn3rI love when I tell someone wanting to go SLI not to and they still do. Then they experience the bugs and lack of support.
The sad thing is I remember doing 2 x HD 6950's and getting good performance in every game I played. I said to myself "Some day I will have the money to get an ultra 3-way sli/cf build.".

Now here I am and it's completely awful. Instead I built an excellent ITX build (Which is cool because I travel a lot), and in fact I don't see mega PC's being viable again for at least 2-3 years....
Posted on Reply
#27
R0H1T
There's no way NVlink would work, even if it were enabled. It's for IBM's POWER only AFAIK.
Posted on Reply
#28
EarthDog
Captain_TomI don't see mega PC's being viable again for at least 2-3 years....
I'll bite... what in 2-3 years do you see making "mega" PCs (whatever that is) viable?
Posted on Reply
#29
mouacyk
hyp36rmaxTHIS THIS THIS. A lot of people don't understand how to configure a setup like this. I generally only recommend SLI/CF if you play with multi-monitor, high refresh rates and / or 4K+ resolutions. Many people discounting SLI/CF hear stories or only have experience with low-midrange GPU's, which isn't even worth it in the first place. This setup requires you to get the best single GPU available and double it for maximized performance.
It didn't use to be this way. In the Fermi and Kepler days, you could SLI two mid-range (460's, 560's and 660's) for about half the price of a high end and achieve near the same average frame rate. SLI can still work this way, but requires way more experimentation with NVidia Inspector and setting custom flags.
EarthDogI'll bite... what in 2-3 years do you see making "mega" PCs (whatever that is) viable?
Wild guess... the crash of crypto's everywhere making people hold their fiats closer to their chest.
Posted on Reply
#30
londiste
R0H1TThere's no way NVlink would work, even if it were enabled. It's for IBM's POWER only AFAIK.
Why? IBM is using it for Power9 but this is primarily developed and used by NVidia. NVLink can do both GPU-CPU and GPU-GPU connections, supposedly in a mesh configuration. The closest competitor to the same functionality today is AMD's Infinity Fabric.
Posted on Reply
#31
R0H1T
londisteWhy? IBM is using it for Power9 but this is primarily developed and used by NVidia. NVLink can do both GPU-CPU and GPU-GPU connections, supposedly in a mesh configuration. The closest competitor to the same functionality today is AMD's Infinity Fabric.
Well that's what I'm saying, it works only with POWER CPU's, it was after all developed in collaboration with IBM, for SUMMIT.

It's proprietary, not unlike IF from AMD or Intel's ominpath (competitor to Infiniband). This doesn't mean that it can't work with x86 or ARM CPU's, in the future, just that it cannot right now.
Posted on Reply
#32
hyp36rmax
mouacykIt didn't use to be this way. In the Fermi and Kepler days, you could SLI two mid-range (460's, 560's and 660's) for about half the price of a high end and achieve near the same average frame rate. SLI can still work this way, but requires way more experimentation with NVidia Inspector and setting custom flags.


Wild guess... the crash of crypto's everywhere making people hold their fiats closer to their chest.
Yea I do remember those days, it was a great value at the time.
Posted on Reply
#33
Th3pwn3r
hyp36rmaxWhat GPU's and monitors were you using?
260 GTXs. The same monitors I have now, Asus VH236Hs.
Posted on Reply
#34
Captain_Tom
EarthDogI'll bite... what in 2-3 years do you see making "mega" PCs (whatever that is) viable?
It's pretty obvious Crossfire and SLI worked incredibly well 8 years ago due to the fact that it was new, but matured - and AMD and Nvidia were busy trying to beat each other at a major feature (Like G-Sync vs Freesync now).

But eventually it just wasn't profitable to keep applying so much manpower to an awesome, but WAY underutilized feature. AMD got higher performance, but Nvidia's more consistently worked (some years). Nvidia phased it out slowly - First Nvidia limited the first Titan to 3-way, then Pascal went with 2-way, and now the new Titan has ZERO sli support. Oh, and AMD only supported it through Polaris because they had to (But now Vega is out).

For now all of the competing API's and engines are making mGPU pointless to support. AMD was right when they insisted that DX12 move mGPU to the driver level, but DX12/Vulkan won't be the standard for at least 2-3 years. Once they ARE the standard, I expect mGPU to maybe become bigger than it ever has been.... But it is a BIG "maybe."
Posted on Reply
#35
Prima.Vera
LycanwolfenI have run SLI for years. Almost every game that said it did not support it I got it to run in SLI. Currently running two 1070ti's now in SLI and all my games are running 4k with no issues at all. Average fps is 100+ in 4k. Single card ran around 30 fps which for me is not enough even 60fps I personally do not like. I talked to so many new gamers that said SLi sucked and didn't work. 80% of the time lol Sli wasn't even enabled on there drivers. 10% of the time it wasn't configured correctly and maybe 10% of the time the frame rendering was not configured.
....
So you're basically saying that adding another card on SLI, is increasing the performance for you ...333% ?
Cool story bro.
Posted on Reply
#36
Jism
Prima.VeraSo you're basically saying that adding another card on SLI, is increasing the performance for you ...333% ?
Cool story bro.
I stacked a Geforce 930 with hypermemory and a tweaked NV/AMD crossfire bridge right next to my RX480 8GB and got 320 FPS in Doom on 4K ultra.
Posted on Reply
#37
newtekie1
Semi-Retired Folder
mouacykThere's no competition above the 1080, so they're selling the crappiest SKU first for the most amount of money. Economics of opportunity.
This is GV100 fully maxed out, it doesn't get better than this.
Posted on Reply
#38
Patriot
newtekie1This is GV100 fully maxed out, it doesn't get better than this.
It is not fully maxed out...the caches have been trimmed and the reduction in memory directory correlates to lower memory bandwidth.
But it is still good. Priced like a Quadro V6000 tho.
Posted on Reply
#39
GoldenX
The only sad thing about losing SLI, is that it means Voodoo is truly dead.
Posted on Reply
#40
Patriot
GoldenXThe only sad thing about losing SLI, is that it means Voodoo is truly dead.
(hugs his Voodoo3) don't say that, you scare the little ones.
Posted on Reply
#41
newtekie1
Semi-Retired Folder
PatriotIt is not fully maxed out...the caches have been trimmed and the reduction in memory directory correlates to lower memory bandwidth.
But it is still good. Priced like a Quadro V6000 tho.
Well, as fully maxed out as a desktop GPU is going to get. There aren't shaders disabled like the previous early Titan releases.
Posted on Reply
#42
Deathlokke
I don't really get most of the complaints in this thread. This is a V100 with a reduced memory bandwidth and priced at a 70% discount. What is the problem here? This isn't aimed at gamers in any way, and is instead designed to be a reduced-price card for those working with deep learning and AI software, not to play Crysis with. If it was designed for gaming they'd have disabled the tensor cores and some of the FP64. Personally, I'm amazed at how good a value this card is for what it's designed for.
Posted on Reply
#43
GC_PaNzerFIN
Don't worry, NVLink support will come in Titan Vv Gold Edition available only from NVIDIA at the low low price of $4999. With partially disabled drivers of course. Wouldn't want to touch the sales of the GV100 Quadro/Tesla products would we. :)
Posted on Reply
#44
Captain_Tom
The funny thing is I WOULD buy this in a second if it turns out its tensor cores and bandwidth can be put to use for mining at say 100 ETH + 5000 MH/s...
Posted on Reply
#45
londiste
R0H1TWell that's what I'm saying, it works only with POWER CPU's, it was after all developed in collaboration with IBM, for SUMMIT.

It's proprietary, not unlike IF from AMD or Intel's ominpath (competitor to Infiniband). This doesn't mean that it can't work with x86 or ARM CPU's, in the future, just that it cannot right now.
You literally bolded the parts supporting your narrative and completely ignored the rest. NVLink also works with GPU-GPU connections.
Both Omnipath and Infiniband are in a slightly different niche, generally for connecting systems on a step higher level.
Posted on Reply
#46
Vayra86
HAHAHAHA

This is the final middle finger to all those noobs saying Titan is 'a gaming card'. "But if I want to and I can, why wouldn't I buy one for gaming?"

Suck on that
Posted on Reply
#47
iO
newtekie1Well, as fully maxed out as a desktop GPU is going to get. There aren't shaders disabled like the previous early Titan releases.
It has 4 disabled SMs as the full chip has 5376 shaders.
Posted on Reply
#48
R0H1T
londisteYou literally bolded the parts supporting your narrative and completely ignored the rest. NVLink also works with GPU-GPU connections.
Both Omnipath and Infiniband are in a slightly different niche, generally for connecting systems on a step higher level.
Yes & it's an enterprise feature geared towards compute heavy systems, that's why it'll not be enabled on desktop systems, probably with the exception of pro cards like Quadro.
Posted on Reply
#49
Th3pwn3r
Vayra86HAHAHAHA

This is the final middle finger to all those noobs saying Titan is 'a gaming card'. "But if I want to and I can, why wouldn't I buy one for gaming?"

Suck on that
If they can afford it I don't think they care what it's for. If it's expensive people will want it just for that reason alone so they can brag.
Posted on Reply
Add your own comment
Nov 21st, 2024 14:12 EST change timezone

New Forum Posts

Popular Reviews

Controversial News Posts