• Hey Guest. Check out your NeoGAF Wrapped 2025 results here!

NVIDIA Volta Unveiled (GV100 new GPU Architecture)

Thanks

Hmm... I highly doubt they will announce anything gaming related pertaining to Volta at this event. They usually hold their own events for things like that.

But it seems like this new architecture is their big leap. I always felt like Pascal was thrown in as a little stop gap.

But the upside to this is that this conference will just cause NVIDIA stock to soar even higher.. I'm thinking NVIDIA will be over $200 before the end of the year at this rate.
 
I don't think I could cope with anyone other than Jen Hsun presenting NVIDIAs plans and products. Sometimes cuts off other presenters / tales over presentations and delivers awkward jokes sometimes, but you can tell the man is passionate about his company.
 
So if this chip cost 3 billion to make. Anyone here have an idea what say the Tegra x1 chip cost to make and the previous nvidia chip. To give an idea if 3 billion is lot.
 
AmyS said:
I'll bet Nvidia's next architecture (Einstein, perhaps finally?) might cost $4B+

Especially moving to a new node, 7nm or 7nm+

2019/2020 ?
Click to expand...

If the data Center, machine learning, AI and other forms of GPU non-gaming used continue to grow (which is looking like a sure thing) then yeah, we will most likely continue to see R&D investment rise.
 
So will Volta be their next release or will they launch an in-between series of cards? I seem to remember reading that they were planning on releasing a rebranding of the 10 series cards with higher clocks.
 
It's hard to take these R&D numbers seriously, to be honest.
3 billion, with, say $100k per engineer on average (aren't they mostly based in Asia with lower salaries?) is 30'000 MAN-YEARS.

As of 2014 nVidia employed 8.8k.
Entire stuff would need to work on Volta for 3 years non-stop exclusively.

There sure are some "material" costs, but I doubt they are significant..
 
llien said:
It's hard to take these R&D numbers seriously, to be honest.
3 billion, with, say $100k per engineer on average (aren't they mostly based in Asia with lower salaries?) is 30'000 MAN-YEARS.

As of 2014 nVidia employed 8.8k.
Entire stuff would need to work on Volta for 3 years non-stop exclusively.

There sure are some "material" costs, but I doubt they are significant..
Click to expand...
It certainly seems 'slightly' exaggerated.
Although I'm pretty sure the design process takes quite bit longer than 3 years.

We saw Volta in roadmaps from GTC2013.
 
pottuvoi said:
It certainly seems 'slightly' exaggerated.
Although I'm pretty sure the design process takes quite bit longer than 3 years.

We saw Volta in roadmaps from GTC2013.
Click to expand...

It was supposed to come way earlier according to 2013 roadmap. And I doubt entire 8.8k team was working on it. (there are always multiple projects in parallel)
 
Omeganex9999 said:
I think they said something like "this year we are launching several products, including new graphic cards..." and so on.

I might have got that wrong though, since I was watching it at work.
Click to expand...

Well, they've already launched a bunch of new graphics cards this year so it doesn't tell us much. But with GV100 availability Q3/Q4 this year there really is no reason why gaming Volta shouldn't be coming around the same time.
 
To that end, I am excited to announce that our Graphics design team in Munich has achieved 16Gbps data rates in our high speed test environment—another first for memory industry. The left picture shows the data eye opening at 16Gbps based on a critical PRBS pattern sequence, with great timing and voltage margin. The right image below shows stable data timing margin (horizontally) versus data rate (vertically), from our base sort speed of 10Gbps up to an unprecedented 16Gbps. This result is based on measurements on a meaningful sampling size of our mass production G5X silicon – not theoretical simulation data.
Click to expand...
https://www.micron.com/about/blogs/2017/june/what-drives-our-commitment-to-16gbps-graphics-memory

It's entirely possible that first consumer Volta cards will use GDDR5X at 14 or even 16 Gbps which is basically the same speeds GDDR6 is expected to launch on in 2018.
 
dr_rus said:
https://www.micron.com/about/blogs/2017/june/what-drives-our-commitment-to-16gbps-graphics-memory

It's entirely possible that first consumer Volta cards will use GDDR5X at 14 or even 16 Gbps which is basically the same speeds GDDR6 is expected to launch on in 2018.
Click to expand...
Nice.

That was exactly what I said on other thread... GDDR5x 12Gbps was not the limit and it can reach 13-14Gbps soon.

Said that GDDR5x at 14Gbps is more than enough for any new gaming graphic card to launch in the next year... the issue of low bandwidth happens more with HPC cards where HBM2 is a better option.
 
ethomaz said:
Nice.

That was exactly what I said on other thread... GDDR5x 12Gbps was not the limit and it can reach 13-14Gbps soon.

Said that GDDR5x at 14Gbps is more than enough for any new gaming graphic card to launch in the next year... the issue of low bandwidth happens more with HPC cards where HBM2 is a better option.
Click to expand...

Yeah, even 14 Gbps (which seems more possible than 16 for a mid range launch) should provide a nice 40% bandwidth increase on the same 256 bit bus compared to launch version of GTX1080 (320->448 GB/s). Add whatever memory saving improvements Volta may have here over Pascal and you're looking at the same +50% bandwidth increase as what is expected from flops too in Volta.
 
Are there any upcoming games that will really take advantage of this power or is this more along the lines of play your games at 100+ fps?
 
Newboi said:
Unless I'm reading the slides wrong, it looks like Nvidia's unified memory claims were specific to using GV100 over Nvlink.

Anyway, I wonder what if any of what's been shown can be an indicator for architecture improvements useful for games. I'm guessing the independent thread scheduling might prove somewhat useful?
Click to expand...

The new cache is more likely to transfer over to consumer gpus than the thread scheduling imo. That will be helpful.
 
kevm3 said:
Are there any upcoming games that will really take advantage of this power or is this more along the lines of play your games at 100+ fps?
Click to expand...

Well, we are still unable to play the highest end games completely maxed out at 4K with consistently greater than 60fps on a single card.

Also, the advent of VR will make the need for more powerful GPUs more relevant than ever. It's been said that 16K per eye will be the necessary resolution needed to achieve true presence. That's a long way away. I assume wide field of view AR will be even more demanding.
 
Newboi said:
Unless I'm reading the slides wrong, it looks like Nvidia's unified memory claims were specific to using GV100 over Nvlink.
Click to expand...
There are some additional features when used over NVLink with P9 but unified memory is there in each Pascal and will obviously be there in each Volta GPU as well. Volta PCIE will also have some improvements over Pascal PCIE.

Newboi said:
Anyway, I wonder what if any of what's been shown can be an indicator for architecture improvements useful for games. I'm guessing the independent thread scheduling might prove somewhat useful?
Click to expand...
Most of Volta's improvements disclosed so far are (understandably) targeting improved compute efficiency and if I'd had to guess I'd say that all of them will be useful in games - but not all games, of course.

Thread scheduling will of course be similar between all Volta GPUs as this is a base level of architecture which you certainly don't want to be different between your GPUs as you'd need two different drivers otherwise. Whatever can be saved here by not using the new scheduling in gaming chips will be more than covered by using these chips in Quadros and Teslas (some people seem to forget that so called "gaming GPUs" are used for HPC/DL as well as the top dog with HBM2).

Volta's updated memory architecture with L1 being very close in its efficiency to a dedicated LDS will certainly affect games in a major way. Unified memory is unlikely to be used in gaming API drivers, in the same way as it is now with Pascal.

The interesting part is the "gaming Volta" SM structure as this time it would be harder to just double the SIMD units while cutting out the HPC stuff from them - GP100 had two SIMD units per SM, GV100 has four of them per SM already. Going with eight per SM seems a bit high from an optimal load point of view - although this may be mitigated by the new scheduling capabilities.

In any case, I expect GV104 with 3584 FP32 SPs and peak throughput around 15 TFlops and GV102 with 5376 FP32 SPs and peak throughout around 20 TFlops. Minimal rate FP64 will be kept for compatibility, highly likely that double rate FP16 will be absent again and it's possible that they'll keep at least one tensor core per SM for the same reasons they keep a couple of FP64 and FP16 units in gaming Pascal chips. 15 TFlops with additional capabilities of the new architecture should be enough to be able to beat 1080Ti by some 20-30% which is what people have grown to expect from a Gx104 part these days.
 
Renekton said:
Likely GV104 first.

1080Ti owners are not going to be happy.
Click to expand...

It's not like the approximate dates were not known when 1080TI was released. Same thing happened with 980TI and 1080. You only got 6 months or so usually after TI release till the new cards now days.
 
StereoVsn said:
Hmm... I wonder if they are going to run into bandwidth problems. HBM2 is again not ready (hence most likely AMD delays) and GDDR6 is 2018. Guess we shall see Volta sooner then later at this pace.
Click to expand...

GDDR6 will have more or less the same bandwidth as GDDR5X.
 
NVIDIA Formally Announces PCIe Tesla V100: Available Later This Year
Interestingly, unlike the Tesla P100 family, NVIDIA isn't offering a second-tier PCIe card based on salvaged chips; so this generation doesn't have an equivalent to the 12GB PCIe Tesla P100. NVIDIA's experience with GP100/interposer/HBM2 assembly as well as continuing production of HBM2 has likely reduced the need for memory-salvaged parts.
Click to expand...
Higher clocks than the P100 PCIE which is impressive for the increase in die size and transistors number.
 
Will consumer Volta chips use the 12nm FFN process? Also is this 12nm process actually smaller than Intel's 14nm process? I've read that GF and TSMC have been claiming smaller processes without shrinkage in all areas of the chip.
 
Newboi said:
Will consumer Volta chips use the 12nm FFN process? Also is this 12nm process actually smaller than Intel's 14nm process? I've read that GF and TSMC have been claiming smaller processes without shrinkage in all areas of the chip.
Click to expand...

12FFN is the same as 16FF+ but with further improvements made specifically for NV.
 
RedlineRonin said:
Hmmmm

Almost exactly a 50% increase in single precision output over pascal PCIe.

Wonder if that's the increase we'd see on consumer side?
Click to expand...
Needed a significant increase in transistor count and die size to get there though.

Still, amazing power efficiency improvements.
 
You must log in or register to reply here.

Similar threads

NVIDIA secures 94% of AIB GPU market
2 3
108
1K
LavitzSlambert replied
KeplerL2: PS6 Won’t Use Full RDNA 5 GPU Architecture, It’s Claimed
Opinion Rumor Hardware  6 7 8 9 10
497
4K
Bojji replied
Full path tracing might not be such a GPU hog in the future, as Nvidia claims over 2x performance improvements in new research
Hardware Platform 
41
307
S0ULZB0URNE replied
Nvidia Warns of 'Very Tight' GPU Supplies for at Least Next 2 Quarters
News Hardware 
5
261
Cumulus replied
NVIDIA and Samsung are making new GeForce RTX 3060 GPUs
News Hardware 
20
557
asdasdasdbb replied
Top Bottom