Nvidia Volta discussion thread

[Life-Binder]

https://www.pcgamesn.com/nvidia/nvid...ass-production

Next-gen Nvidia Volta GPUs are getting crafted at the end of the year


Nvidia’s silicon partner, TSMC, have just announced they will be entering volume production of 12nm FinFET chips in the fourth quarter of 2017. That’s the lithography they’re using for the gaming-focused Nvidia Volta GPUs, making an early 2018 launch of their next-gen GPUs a real possibility.

The first Nvidia Volta GPUs have already rolled off the production line and into the hands of the professionals. Those are the hyper-expensive Tesla V100 chips though, not the sorts of graphics silicon likely to find their way into our gaming rigs.

The actual Volta GPUs us normies are going to be able to buy won’t quite be the parallel processor monster the 5,376 core Tesla is, but they’re still liable to be pretty chunky chips nonetheless. TSMC have produced the 12nm FinFET chips for the pro-level cards, but those will have only been in very limited volumes.

The high-volume 12nm FinFET parts are due to start going into mass production at the end of this year. That indicates a 2018 launch for the new Nvidia Volta graphics cards.


If I were a gambling man (I would go by the name of Gamblor) I’d be tempted to put a couple of pennies on Nvidia releasing their new cards in March next year.

https://www.overclock3d.net/news/gpu...n_march_2018/1

Nvidia's next generation of Volta GPUs are expected to release in March 2018


Nvidia has already confirmed that their upcoming Volta series of gaming GPUs will not be released anytime soon, with many expecting Volta gaming series GPUs to release in the first half of 2018.

Today some rumours have a emerged which states that Nvidia's gaming-oriented Volta series GPUs will release in March 2018, which makes a lot of sense given the current production schedule for GDDR6 memory from companies like SK Hynix. TSMC's new 12nm process, which Volta is based on, is also expected to be ready for large scale production by the end of the year, giving Nvidia another reason not to rush production, as higher yields and production rates are promised later.

Right now the only Volta-based GPU that Nvidia produces is their colossal Tesla V100, which is suffering from low-yields due to the maturity of the 12nm production node and the sheer size and scale of the Tesla V100. To put it into perspective the Tesla V100 has almost 2x the transistor count of the GTX 1080 Ti, with Nvidia stating that the GPU sits at the limits of manufacturability with today's process technology.

In short, Nvidia's Volta plans will ensure that their chosen memory type will be in greater supply and that TSMC's 12nm production node is more than ready for the mass production of Volta gaming GPUs. It is also worth remembering that Nvidia does not feel in any way rushed to release Volta, given the performance of their Pascal architecture when compared to competing hardware.

[Life-Binder]

https://www.nvidia.com/en-us/titan/titan-v/

well here it is

the first Volta Titan came out


its a 3000$ Titan, since it still has active Tensor cores and is aimed at AI researchers first and foremost .. but it is based on Volta arch and has 5120 CUDA cores and is already supported in the latest 388.59 GF drivers, so you can game perfectly fine with it and I assume it will beat 1080Ti easily (any game benches yet ?)

- - - Updated - - -

for gamers 3000$ is too much, we gotta wait for that ~$800-900+ 2080Ti


but as a "cheap" V100 its a bargain

[Cyrops]

https://www.nvidia.com/en-us/titan/titan-v/

well here it is

the first Volta Titan came out


its a 3000$ Titan, since it still has active Tensor cores and is aimed at AI researchers first and foremost .. but it is based on Volta arch and has 5120 CUDA cores and is already supported in the latest 388.59 GF drivers, so you can game perfectly fine with it and I assume it will beat 1080Ti easily (any game benches yet ?)

- - - Updated - - -

for gamers 3000$ is too much, we gotta wait for that ~$800-900+ 2080Ti


but as a "cheap" V100 its a bargain

I'm sure some whale will buy 4 of those in no time.

[Strawberry]

Isn't Volta scraped for us gamers? Titan V is not a gaming card even though it will be about 30-50% faster than the latest Titan.
Didn't they say they are skipping Volta and going straight for gaming Ampere?

EDIT:
Found a discussion here:
https://www.reddit.com/r/nvidia/comm...ta_for_ampere/

The news have been circulating on tech websites for a few weeks now however I am not sure if that's a rumor or not or how true it is.

[Sorshen]

$3k... lmfao, AMD get your shit straight!

[Life-Binder]

well this Titan V has video outputs and is supported now in regular GF drivers, so you cant say its not a gaming card ... so, technically its is Volta + gaming

but perhaps 2018 actual gaming-oriented GF cards will be Ampere rather than Volta

[KingSapmi]

$3k... lmfao, AMD get your shit straight!

Might be a joke going *swoosh* above my head here, but what does this have to do with AMD? Their lack of high-end GPU or?

[Life-Binder]

the 3000$ is chiefly because its a 815 mm2 (!!!) monster GPU with all the tensor cores .. AMD is secondary concern here, though of course them being so far behind at the highest-end matters too

[Sorshen]

Might be a joke going *swoosh* above my head here, but what does this have to do with AMD? Their lack of high-end GPU or?

Yeah, people were applauding AMD for tackling Intel and promoting healthy competition and actual progress this year.
But in the mean time, their failure in the GPU market has put nvidia in a commanding position - they can basically charge w/e cause they have nothing to worry about, can even skip a 'tick' in their roadmap. Still, screw nvidia for being greedy fuckers.

[Evildeffy]

Considering how much of the chip's build is wasted on Double Precision and Tensor hardware, something games will never utilize, I'm going to guess that this year's GV102 variant will probably outperform this thing in games.

This is a full GV100 die, same type of die as the V100 used in supercomputers ... just with graphical outputs in card form.

So this year may actually mark the start of a divergence of the Titan line of dies to the consumer line of dies.
GV102 may actually result in the 1180Ti/2080Ti (naming scheme is unclear as of yet, do not assume either please) being the strongest gaming card out there.

That said ... nVidia being nVidia (and honestly probably 99% of us would do as well) will price the living crap out of because of current lack of high-end competition.

Unless Adrenalin holds some form of "Miracuru" we'll be "meh"-ing our way along with AMD and unless their 14nm GPU process is absolutely terrible and TSMC's 12nm GPU process doubles everything from power efficiency to clock rates that will remain the same when the refresh happens in Q1 -> Q2.

[KingSapmi]

Yeah, people were applauding AMD for tackling Intel and promoting healthy competition and actual progress this year.
But in the mean time, their failure in the GPU market has put nvidia in a commanding position - they can basically charge w/e they like cause they have nothing to worry about, they can even skip a 'tick' in their roadmap. Still, screw nvidia for being greedy fuckers.

Thanks for clearing that up. But yeah, I agree. I was holding off for the hopes of RX Vega to hit the shelf and compete at a higher level. However, should the prices drop, the FreeSync/Vega system might be more viable for me, personally. Right now I've got a 1ms 144hz FreeSync monitor paired with a MSI GTX 1070 Gaming X. Not ideal, but workable until a better solution comes my way.

[Vash The Stampede]

Might be a joke going *swoosh* above my head here, but what does this have to do with AMD? Their lack of high-end GPU or?

Supply and demand. Without a proper competitor you get to pay $3k for a graphics card. Just like you got to pay over $250 for a overclockable Intel i5 without hyperthreading, and now Intel might even make a 6 or 8 core CPU in that price range thanks to AMD's Ryzen.

[DeltrusDisc]

https://www.nvidia.com/en-us/titan/titan-v/

well here it is

the first Volta Titan came out


its a 3000$ Titan, since it still has active Tensor cores and is aimed at AI researchers first and foremost .. but it is based on Volta arch and has 5120 CUDA cores and is already supported in the latest 388.59 GF drivers, so you can game perfectly fine with it and I assume it will beat 1080Ti easily (any game benches yet ?)

- - - Updated - - -

for gamers 3000$ is too much, we gotta wait for that ~$800-900+ 2080Ti


but as a "cheap" V100 its a bargain

HOLY GOD LOOK AT THE SIZE O' THAT THING!!!

That chip is a fucking monster, holy shit.

Seriously, WTBBBBBBBBBB benchmarks, holy fuck.

I'm sure some whale will buy 4 of those in no time.

Limit TWO per customer. Rofl, as if. I'd like just one, thanks!

- - - Updated - - -

Considering how much of the chip's build is wasted on Double Precision and Tensor hardware, something games will never utilize, I'm going to guess that this year's GV102 variant will probably outperform this thing in games.

This is a full GV100 die, same type of die as the V100 used in supercomputers ... just with graphical outputs in card form.

So this year may actually mark the start of a divergence of the Titan line of dies to the consumer line of dies.
GV102 may actually result in the 1180Ti/2080Ti (naming scheme is unclear as of yet, do not assume either please) being the strongest gaming card out there.

That said ... nVidia being nVidia (and honestly probably 99% of us would do as well) will price the living crap out of because of current lack of high-end competition.

Unless Adrenalin holds some form of "Miracuru" we'll be "meh"-ing our way along with AMD and unless their 14nm GPU process is absolutely terrible and TSMC's 12nm GPU process doubles everything from power efficiency to clock rates that will remain the same when the refresh happens in Q1 -> Q2.

I'm tempted to change my subscription to just "GTX 20xx series Club" just because of your comment. <3

Seriously, I don't get why so many people are so bothered about which way NVidia goes, ultimately it won't matter one bit, but I am still betting on 20 as opposed to 11.

[Gray_Matter]

That said ... nVidia being nVidia (and honestly probably 99% of us would do as well) will price the living crap out of because of current lack of high-end competition.

Here is the thing. I don't think this beast matters for most of us. I just want a good graphics card for $200. I don't care whether it's NVidia, AMD or even Intel.

For me, the top of the range fight keeps getting overblown. It's like buying a bicycle because Rolls Royce makes it when there are better ones out there on price/performance.

[ZenX]

If the tensor FPs is as advertised, it would definitely be a revolutionary co-processor considering SLId GTX 580s max rate of ~3 TFPs for AlexNet.

First of all, gaming performance of GPUs since CUDAs introduction have never been a priority for me when purchasing one. Heck, I am using a gt 1030 for playing WoW at a 1080p display at mid-high settings while using my 1080 for compute-tasks which I have been intending to pair with an AIB Vega (RX 460 and 1050 Ti being the ideal pair are really anemic for compute tasks considering their slot occupancy) that I gave up on (Thumbs-up to AMD for the worst launch ever).
However much I had the blower type coolers while not being a fan of liquid cooling, I was thinking of going for a Titan Xp finally justifying the price for it.

Yet, nVidia had again reminded me why I dislike them regardless of them being the better card manufacturer. Better tech at a harsh premium within such a short lifecycle making you regret your purchase.

Since hearing about the Tesla V100 and its tensor cores a day ago, I have been eager to get my hands on such a card. Upon seeing the announcement for Titan V, I initially had my hopes up for getting one and then I saw the price which basically decimated them and then devastated me to the core. Basically, this is what Vega FE with its $1000 should have been but thanks to the turn-coat who left RTG probably aware of the crap he pulled, it fell on its head and now got itself kicked in the guts (Basically, 4xVega FE @FP16 ~= 1xV100 tensor FP).

Guess when they state "scientists" and "researchers", they indicate those who don't spend money out of their pocket for buying these. Nonetheless, I guess now I gotta wait for Intel to release the Lake Crest coprocessor at competitive pricing as they are hoping to break nVidia's hold on the market. While RTG is "trying to do its best" sniffing glue (Not saying AMD as I am still thankful to them for Ryzen)

[Kagthul]

Here is the thing. I don't think this beast matters for most of us. I just want a good graphics card for $200. I don't care whether it's NVidia, AMD or even Intel.

For me, the top of the range fight keeps getting overblown. It's like buying a bicycle because Rolls Royce makes it when there are better ones out there on price/performance.

This also isnt a consumer card. Its meant for professional uses, which is why it is full-fat GV100. This is for people who need compute muscle but dont need as much as a Tesla card provides (or costs).

Given its low clocks, it probably actually WONT be a terribly great gaming card (i mean, im sure it will outperform even a Titan X-p, judt because it has so many more cores) compared to its consumer Volta cousins when they launch. Id expect the Titan X-v or 1180Ti to soundly outperform it for gaming at a fraction of the cost.

[Evildeffy]

Here is the thing. I don't think this beast matters for most of us. I just want a good graphics card for $200. I don't care whether it's NVidia, AMD or even Intel.

For me, the top of the range fight keeps getting overblown. It's like buying a bicycle because Rolls Royce makes it when there are better ones out there on price/performance.

Whether it matters to you or not is irrelevant really..

nVidia will gouge even their low to mid-range cards if they so see the opportunity.

So even if you go for that nVidia will still do the same.

And like I said prior .. I am going to assume the GV102 variant WILL outperform this cardbin terms of gaming.

- - - Updated - - -

If the tensor FPs is as advertised, it would definitely be a revolutionary co-processor considering SLId GTX 580s max rate of ~3 TFPs for AlexNet.

Be aware that the TFPs mentioned is not TeraFlops but Tensor Flops ... nVidia's measuring standard.

Just making sure you're not confusing it.

[ZenX]

Whether it matters to you or not is irrelevant really..

nVidia will gouge even their low to mid-range cards if they so see the opportunity.

So even if you go for that nVidia will still do the same.

And like I said prior .. I am going to assume the GV102 variant WILL outperform this cardbin terms of gaming.

- - - Updated - - -



Be aware that the TFPs mentioned is not TeraFlops but Tensor Flops ... nVidia's measuring standard.

Just making sure you're not confusing it.

I am aware and as I work on Neural Networks utilizing (basically) tensor calculus, it would be of great use to me. Now the question is whether if the 100 TFPs is for tensor cores only or did nvidia go ahead and basically derived results as an aggregate of their FP16, FP32 and FP64 compute units while all them operating at FP16 etc (Not sensible but marketing people love to pump the exaggeration) Although, considering the nvidia-docker it might actually be a possibility as well.

[Deleted]

3k... ooooooooooook lol.

[Life-Binder]

Since hearing about the Tesla V100 and its tensor cores a day ago, I have been eager to get my hands on such a card. Upon seeing the announcement for Titan V, I initially had my hopes up for getting one and then I saw the price which basically decimated them and then devastated me to the core.

this is a full tensor enabled 815 mm2 GV100 chip that was essentially slashed down from 10,000$-15,000$ to 3000$ at the cost of 4GB VRAM, some bandwidth and certified drivers (and NVLink ?)

I dont think in that sense you can complain about the price at all, its arguably even "cheap" for this kind of chip


just stop thinking about it as any kind of gaming card or even as you thought of the old Titans

- - - Updated - - -

I mean if you specifically need the "Tensor-flops" for your AI research then 110+ Tensor-flops for $3000 is better price/performance than any other kind of compute card in existance, inc the Vegas

while also using less power and taking up less space and producing less heat

- - - Updated - - -