New Titan ? Nandato?

Noticeably shorter than the existing Titan/1080ti.

 

That would make sense if it was using HBM.

 

Interesting... it's hard to think it'll dissipate heat anywhere near my preferences with a cooler like that.

I've always ripped the stock cooler off for a water block anyways but I wish I didn't have to do that.

 

Titan GOLD!?! Likely uses GP100 if its genuine, like the Quadro GP100. Doesn't seem shorter to me though, and the Quadro card is not shorter even with HBM.

 

yea looks normal length. HBM cards are much shorter than that from what we have seen.

Also NVlink means nothing for intel/AMD systems. Only IBM supports it last i heard.

 

Agreed!
http://www.nvidia.com/object/nvlink.html

 

One major consideration if this is not fake image; GP100 has FP64/FP32/FP16 as primary focus compared to current Titans/GP102.
That said they could be doing the GP100 as a Titan for those scientific smaller projects so they also have access to FP16 inferencing and bringing back FP64 to a cheaper product as well for same reasons.

You would not be able to mix the Titan cards though in same PC/node as they have different CUDA compute versions and different instructions/operations.
So not sure what to think as the scientists/engineers have Quadro GP100 although this is not really priced in an accessible way, personally weighting this more to fake for now but I can see uses and market for it *shrug*.
Cheers

 

DGX and HGX is fitted with Intels CPUs and still uses NVlink.

 

interesting but still no nvlink to the CPU....so this is being used like an SLI bridge. Just more elaborate.

IIRC NVlink was also supposed to go to the CPU too but I think only IBM said it was going to implement it. At least I thought that was the point was to get my BW to the CPU.

Whats DGX and HGX?

 

It goes to the Power8/9 with a different connection than PCIe and uses such as Blue Link 25G that supports NVLink; basically used for NVlink and historically CAPI.

But Shintai was just pointing out that NVLink is usable by cards that still connect traditionally using PCIe with the CPU.
TBH it is much more than SLI bridge because this is specifically for systems-apps that are professional visuation or more importantly scientific compute/modelling/etc and is a true mesh connection while SLI is crude in comparison.

Cheers

 

like i said more elaborate

 

You cannot use SLI in same way as the NVLink
When I say SLI is crude, I really mean NVLink is not even an evolution of SLI.
Cheers

 

elaborate
involving many carefully arranged parts or details; detailed and complicated in design and planning.

 

POWER is dead anyway, so it was more a novelty thing.

DGX=8 P100+2 Xeons (BRW-EP) in a box.
HGX=8 V100+2 Xeons (SKL-SP) in a box.

 

Still PCIe to CPU :/ PCIe 6 cant come soon enough :/

 

Ask yourself what is the primary purpose of SLI, then ask yourself what is the primary purpose of the cohesive mesh interconnect of NVLink?
One comes back to scaling geometry and fill rate, the other goes waaay beyond this in how the two GPUs are connected in terms of cache/memory and cores used/compute and scaling across multiple operations/frameworks/applications.

CHeers

 

https://arstechnica.com/gadgets/2017/02/nvidia-quadro-gp100-price-details/

yeah interesting stuff man

 

do they both link GPUs?
Yes

Do they both transfer data?
Yes

Is NVLink complicated in design and planning?
Yes

Is NVLink more complicated in design and planning than SLI bridge?
Yes

....I stand by my statement since it is 100% accurate....you unnecessarily technical dip.

 

So is nvlink more like having a homogeneous GPU rather than the bullshit known as SLI/crossfire?

 

yep

 

After Nvidia is removing SLI from lower mainstream cards like GP106. Perhaps NVlink will replace the SLI connector in the future. Not just for gaming, but also for lower Tesla models and Quadro cards.

 

That's really exciting then.

 

Yes and no about the excitement part. While it IS technically better in every way then SLI, due to it's nature it will be far better suited for splitting compute loads then graphics loads. Gaming won't benefit much over SLI in the same ways compute does. NVlink is great for HPC, but is still meaningless to gaming for the time being and short term future. I think it will be a good while before anything like this makes it to the consumer level if ever other then limited, niche scenarios like multi-GPU, single card uses, if even that.

Long term might play out differently and who knows, eventually we may see NVlink replace SLI, but not today or tomorrow or even any time soon I think. Just my opinion though.

 

Yeah.
But may be sooner than much later (being years) but comes down to API integration for consumer and key will be for such to mature on both DX12 and Vulkan.
We know DX12 with EMA support is the new alternative, especially as Nvidia has sort of given up on SLI and AMD I think has put Crossfire more into the hands of the devs/studios rather than themselves.

This is a good example of it working on DX12 and EMA, and this is early days for the tech:



But really needs both Vulkan and DX12 to go heavily into maturing this, along with Nvidia and AMD also pushing it with their engineers and support, and for it to gain greater support to have more relevance and to get peformance to acceptable levels.
Yeah not soon, but it does not necessarily mean too long away although for it to really and be adopted by studios it requires several other parties that I mention.
AMD at least seems to be pushing this with DX12, so at some point Nvidia will need to join in.

Of course that is me totally ignoring the elephant in the room being DX11
So will not be all studios or games.
But NVlink could logically work well with these APIs (Nvidia would create the libraries required to integrate).
A big consideration is whether trickle down this tech for other Tesla/Quadros; if so it is likely to be left on with the consumer card - does not mean it can be used but more incentive for Nvidia to make it work for consumers in this situation.
If they do not trickle down on Tesla/Quadro other SKUs, well unlikely to see it on consumer.
Trickle down I could see it on GV102 quite quickly though, makes sense but that would lose another differentiator between GV100 and GV102.
But this is digressing from the OP regarding just how different and advanced NVLink is relative to SLI.
Cheers

 

I think someone is trolling hard lol. I would expect leaks of gtx 2080 before I see leaks of titan volta. lol.

 

http://www.overclock.net/t/1630852/fb-videocardz-picture-of-volta-titan-leaked-by-nvidia-intern

 

What an idiot

 

What does a Intern care he probably would leave in a few weeks anyway.

 

His internship could have turned into a lucrative career. Now no technology company in the world will touch him. He committed career suicide by identifying himself. I hope he likes manning cash registers at 7-11.

 

He's a Hero in my Book
All he needs to do is NOT put Nvidia down as his previous employer on his next application.
I like how they have the name of the Card on the reference design.
Instead of Geforce GTX they must of changed to molds at Fox Conn.
I'm going to use that Picture he snapped as my Wallpaper for my Lock screen.

 

he signs an NDA too...he/she is in serious shit.

Could have also been set up....so innocent until proven guilty

Not hard to set someone up like that

 

Well they could be doing something like the g80, the g80 was ready 1 year before launch, the chip dates for the g80 was 1 year prior to launch.

 

the car???

 

I was thinking that, too. Maybe it was just a test to out the weak ones.

 

LOL no not a volvo was it a volvo, can't remember has been a long time since I looked at them, 8800 gtx and gts

 

ah...was so confused hahaha

googled g80 and got a car and was like was he referring to a self driving car or something that this was going in? hahha

 

lol

 

G80 is the designation of the Tesla core found in the 8800GTX, Ultra, and (original) GTS. So razor1 was correct .

 

I had both the 8800 GTS 320mb and 640mb. I remember G80 fondly.