Nvidia Volta Speculation Thread

[Deleted member 13524]

of course, they are stock pilling HPC Volta GV100 with HBM2, not the consumer version.

From this 5-day-old article from Anandtech, TSMC only has 12nm FFC which is a pseudo-half node from 16FFC.
High-volume production is expected for 2018, yet you think nvidia has been stocking 12FFC Volta chips one full year before that, to the point of increasing a company-wide inventory growth of 3%?



BTW, 16FFC is an area-optimized version of 16FF+, but this seems to come at the cost of performance and/or power efficiency (if Kirin 960's results are anything to go by).
If Volta really uses 12FFC, maybe we should expect similarly or lower-clocked but wider GPUs in the Volta generation.

 

[Clukos]

I expect an increase in die size just like Kepler -> Maxwell. 314 mm2 GP104 equivalent should be somewhere in the ~400mm2 range, GP102 replacement should be 500-550mm2.

 

[Blazkowicz]

I suppose it's possible GV100 uses 16FF+, and that other Volta products might use something else (as evidenced with GP107 doing its own thing, and perhaps GP108 using the same)
I also think the larger consumer Volta (GP102 and GP104) might be on 16FF+ still.

GP108 cards (Pascal GT 1030) haven't been announced yet. It could also be an Android console or set-top box using Parker. Or even just the FCC-approved Shield handheld with a TX1).

Perfect graphics card (for some users), except for the DVI-D, bit of a let down except if you need cost effective dual link DVI or plain DVI.
I can't believe it does 10 gigapixels/s, used to be 1 gigapixel/s was impressive.
Won't be dirt cheap as full GK208 cards are over 60€ here.

R.I.P. GM108.

 

[Deleted member 13524]

I expect an increase in die size just like Kepler -> Maxwell. 314 mm2 GP104 equivalent should be somewhere in the ~400mm2 range, GP102 replacement should be 500-550mm2.

Weren't Kepler and Maxwell using the same 28HP process?
In this case we'll have Pascal at 16FF+ and then Volta at 12FFC (C stands for Compact). Volta built on the half-node evolution of the "compact" version of 16FF+.
Transistor density should be at least some 30% higher, so nvidia could be using similarly-sized GPUs as successors this time.

Perfect graphics card (for some users), except for the DVI-D, bit of a let down except if you need cost effective dual link DVI or plain DVI.

There are some other leaked cards. Which is why I think this is the GPU responsible for that 3% inventory increase.
This one fits all the "needs" that HTPC users with tiny cases may have: Low-Profile, Passive, HDMI 2.0 and DisplayPort.

Most of these chips are probably going into mid-range laptops, though.

I suppose it's possible GV100 uses 16FF+, and that other Volta products might use something else (as evidenced with GP107 doing its own thing, and perhaps GP108 using the same)

It's true that all rumors and info that's been coming up about the 12FFC Volta seems to refer to a consumer replacement of GP102, mainly because of SK Hynix coming on record for that 384-bit card with GDDR6.
I guess a 16FF+ compute-oriented Volta with HBM2 could have been in mass production for a while.
But isn't it too early to announce a successor to GP100 (i.e. wouldn't it cause an osborne effect)?

 

[Blazkowicz]

Wasn't GV100 announced before GP100?
We've known of the POWER9 + Volta supercomputer for a long time.
Sure, has nothing to do with an upcoming launch. I also don't know what's 3% inventory but building up GV100 may be far lower than that.

 

[CarstenS]

Regarding the inventory: Maybe you have to pay for early risk production runs up front so that wannabee techeleads do not clog the pipelines? Just a thought.

 

[xpea]

Tesla V100
815mm2, TSMC 12nm FFN, 21B transistors

my sources were right !

 

[xpea]

wow wow wow

biggest update since Fermi !

 

[Deleted member 13524]

What's TSMC 's 12 FFN?

Dedicated Tensor cores apparently do 2*FP16 MUL + FP32 ADD at a very high rate (exclusively for 4x4 matrix processing?), hence the 120 mixed TFLOPs.
900GB/s HBM2 means it's using chips running at 1.8Gbps, up from the ~1.4Gbps in P100.


P.S.: Obligatory bullshit moment: showing a trailer for the 1-year-old Kingslaive CG movie as "what Takeshi Nozue thinks games will look like in the future". Yeah ok.



P.S.2: The 5120 ALUs are apparently running at ~1.47GHz to give those 15 FP32 TFLOPs.
Quite the achievement for a >800mm^2 chip.

 

[Razor1]

TSMC 12nm is another half node

with that much processing power, yeah games will look like that, the character I showed in the pc gaming forum, guess what, is a cinematic model, just structured to run in pc game engine. 300k untessellated polys, each piece of the armor moves separately, just like in real life. It will run on today's best hardware, so in another 2 years or even next gen graphics cards, its going to run really well.

I just learned how to use substance painter in the past 2 or three weeks, and now learning how to rig (well I know how, just getting better at it since this rig is quite complex). Just imagine people that have years of experience in texturing and animation, what they can do?

 

[Clukos]

Tesla V100
815mm2, TSMC 12nm FFN, 21B transistors

my sources were right !

My brain was right also, bigger dies -> more cuda cores for Pascal -> Volta. Expecting ~420 mm2 for the GP104 replacement and around ~600 mm2 for GP102 replacement

 

[Picao84]

While quite the achievement for such a big chip like Totten said, it is not that much of an increase in FP32, only 3 TFlops, so I don't believe that "games will look like that", Razor. The added logic for the "Tensor Cores" might have not allowed a larger increase I guess. On the other hand, I doubt we will see GV100 sold as a graphics chip, just like GP100 was not. The consumer versions will probably have none or very few "Tensor Cores", so maybe they can have more Cuda Cores and higher FP32 throughoutput?

 

[Razor1]

While quite the achievement for such a big chip like Totten said, it is not that much of an increase in FP32, only 3 TFlops, so I don't believe that "games will look like that", Razor. The added logic for the "Tensor Cores" might have not allowed a larger increase I guess. On the other hand, I doubt we will see GV100 sold as a graphics chip, just like GP100 was not. The consumer versions will probably have none or very few "Tensor Cores", so maybe they can have more Cuda Cores and higher FP32 throughoutput?

Oh it will, I'm already working on a game that will look like that lol

 

[Infinisearch]

it is not that much of an increase in FP32, only 3 TFlops,

Isn't it 5Tflops more? GP100 is 10.x tflops IIRC. Thats a 50 percent improvement.

 

[McHuj]

Wow. I didn't know a 800+ mm chip was possible. When is thing shipping?

 

[iMacmatician]

NVIDIA also announced the DGX-1 with Volta, the DGX Station (4x V100), and the HGX-1 (8x V100 for cloud computing).

(From AnandTech.)

The DGX Station is 1500 W with 4 GPUs so does that imply 375 W per GPU?

Wow. I didn't know a 800+ mm chip was possible. When is thing shipping?

2017 Q3 for the DGX-1 Volta. Also,

All DGX_1 orders are now going to be for DGX-1V, it seems

 

[Deleted member 13524]

Wow. I didn't know a 800+ mm chip was possible. When is thing shipping?

Personal DGX PCs will be available in Q3. I guess they're really using GPUs made at 12FF risk production like @CarstenS suggested.
One Personal DGX costs $69000, so that may be why they can resort to that.

Maybe 12FFN = 12FFC in risk production?

TSMC 12nm is another half node

The only 12nm reported from TSMC so far is 12FFC. I was wondering what 12FFN is.

with that much processing power, yeah games will look like that

A 15 TFLOP GPU (25% more than a Titan Xp) can make games look like Kingslaive?

Haha... maybe at 480p30 or less.

My brain was right also, bigger dies -> more cuda cores for Pascal -> Volta. Expecting ~420 mm2 for the GP104 replacement and around ~600 mm2 for GP102 replacement

15B transistors in GP100 occupied 610mm^2.
815mm^2 with 21B transistors actually seems like it has a very similar transistor density, so you might be right.
I don't know what this 12FFN is, so.. ¯\_(ツ)_/¯

 

[Picao84]

Isn't it 5Tflops more? GP100 is 10.x tflops IIRC. Thats a 50 percent improvement.

You are right but I was comparing with GP102, which has 12.1 TFlops. But the point was that 3TFlops are probably not enough to render Kingslaive in real time

 

[DavidGraham]

VideoCardz says the chip has disabled cores in the Tesla V100, full chip is 84 SMs (5376 cores). I think consumer versions should run at higher clocks as well.
https://videocardz.com/69378/nvidia-announces-tesla-v100-with-5120-cuda-cores

 

[Razor1]

A 15 TFLOP GPU (25% more than a Titan Xp) can make games look like Kingslaive?

Haha... maybe at 480p30 or less.

I just stated a 10 TF card can do it, game developers are held back by consoles, the assets that I'm making will not run on today's consoles, just too many polys, too many bones, and too much on the lighting. And it will run on 2k reses. Yeah at 30 FPS on a gtx 1080 but hell that's ok. Right now, I can animated 10 of these characters running around on a gtx 970 30 fps, no advanced lighting (using UE4's default GI), on a Titan P, its running well above 100 fps. So I see no problem games to come out and look like that and run on Volta. As long as they aren't limited by consoles, its not a problem.