Fudzilla - "Nvidia to showcase Volta in early May"

'showcase' probably means a bunch of slides and Powerpoint demos...mainly meant to rain on Vega's upcoming release
 
If it is true I would expect only the P100 successor to be revealed. Seems way too early for any Volta-based consumer products - I could only see it happening if Vega forces their hand.
 
Most are expecting it to showcase in May because that is the GTC May date.
They showcased P100 in April 2017 at the GTC event.

And the comments about wood screws...
Ah how many went on about the P100 being a vapour product, and yes full stack GDX-1 was available outside of specialist supercomputer/science lab contracts in Q3 with 1 week turnaround, the core clients with certain project contracts or certain high profile research were receiving the P100 before Q3.
Cheers
 
It was a joke. I'm sometimes like TaintedSquirrel. I will praise and pillory both brands as the mood takes me. :ROFLMAO:
 
Algrim said:
It was a joke. I'm sometimes like TaintedSquirrel. I will praise and pillory both brands as the mood takes me. :ROFLMAO:
Click to expand...
More directed at what was said in the past (with the P100) rather than now, but yeah sorry does come across as directed at you specifically.
I can see many sites/forums and BitsandChips/Semi-accurate/etc all again saying it is vapor ware and wooden screws :)
And 1000000s of posts on WCCFT all agreeing.

Groundhog day again after the GTC.
Cheers
 
realworld said:
NVIDIA doesn't even look at AMD anymore. They look right through them as if they don't exist.
Click to expand...

Sadly, I think this statement is more and more accurate after every release cycle. I would love Vega to smack NV right in the face (because I always root for the underdog) but I don't think it will. *Before you flame me, I just want to point out I own a 1070, and I just bought a 1080ti*.
 
This would be awfully fast turnaround time, just a year after P100.

But hey, the HBM2 and process tech are now solved problems, so who knows?

I was thinking more realistically September.
 
I hope someone gets really excited for Volta and sells me a 1080 or 1080ti FE super cheap :)
 
I just hope optimization keeps up with Pascal since we are talking about Volta here. Remember how well Kepler dropped off once Maxwell came into the fold.
 
Okay having a 1080ti. And looking at Volta specs. 15tflops. More cores and same clock.

I honestly think Volta is not going to be what we thought it would be for gaming. And look at that die size. Pretty disappointed in those specs, I think it's designed purely for deep learning with not too much increase in gaming.
 
NKD said:
Okay having a 1080ti. And looking at Volta specs. 15tflops. More cores and same clock.

I honestly think Volta is not going to be what we thought it would be for gaming. And look at that die size. Pretty disappointed in those specs, I think it's designed purely for deep learning with not too much increase in gaming.
Click to expand...
It's pretty clear that Volta was not developed for games, it was developed for GPU accelerated computing. That is the market driver right now, hopefully AMD can keep their gaming hardware moving. I realize that yes, I'm sure Volta will perform wonderfully in games with it's beefy engine, it just appears that at least V100 is meant to crunch numbers more than make pretty moving pictures
 
Napoleon said:
It's pretty clear that Volta was not developed for games, it was developed for GPU accelerated computing. That is the market driver right now, hopefully AMD can keep their gaming hardware moving. I realize that yes, I'm sure Volta will perform wonderfully in games with it's beefy engine, it just appears that at least V100 is meant to crunch numbers more than make pretty moving pictures
Click to expand...

Vega is no different than GV100 in terms of market focus.
 
We get this every year.
Last year P100 " that is HPC so consumer will not gain".... not quite.
"Pascal is just Maxwell with higher clocks".... not quite.
Now we are getting again "V100 is HPC so consumer will not see gains", yes they will have similar level of gains.
"Volta is just more cores same clocks".. Err very much no it is not.
It is a fundamental shift in the GPC-SM-Cuda core design also changing behaviour of Caches and also Thread Scheduling, while again making another jump in efficiency just like Kepler-->Maxwell.

Here are the stats for that point:
It is still same 300W as the P100, ergo any improvements comes from further efficiency/IPC.
They increase the die by 33.6% while impressively keeping with same 300W and yet they go further:
FP32 compute increases by 41.5% or 2x (yeah depends upon function with Tensor).
FP64 compute increases by 41.5%
FP16 compute increased by 41.5% or 4x (yeah depends upon function with Tensor).
NVLink 2 increased BW and links supported (big improvement).
Memory BW increased by 25%.

So that is another jump in efficiency, performance, and also arch changes relating to mix precision/Thread Scheduling/Cache behaviour/accelerated matrix functions/etc.

Cheers
 
Last edited:
bexxxxxx said:
.. and same power. Common, GP100 to GV100: 1 year, 50% more perf, and 0% more power. Not impressive at all?
Click to expand...

if you change the sentence to ..and same power. Common, Fiji to Vega: 50% more performance and 0% more power. not impressive at all?. you would have EVERY AMD fanboy doing party at home saying every kind of non-sense things as how AMD is far superior to Nvidia, that's how things work in this world.
 
Rvenger said:
I just hope optimization keeps up with Pascal since we are talking about Volta here. Remember how well Kepler dropped off once Maxwell came into the fold.
Click to expand...

can you prove kepler dropped performance?. can you prove maxwell dropped performance?... look, what Nvidia do, is basically what AMD is starting to do right now, "new arch, forget the old, focus on the new one" that's it, since Polaris ALL of the optimizations and main focus is for Polaris, in fact all of their performance gains are described with polaris cards as those are the newer ones even having Fury Series which are stronger, people have all kind of mental diarrhea saying that Nvidia purposefully gimp the performance of older cards, and nope, that isn't right, they just make all the optimizations for newer architectures and they change the architectures way more frequent than AMD which hey, are still using the same GCN since 6 years ago, they are just optimizing and refining it, that's why people have that fake sensation that AMD cards age better and all kind of non-sense shit.

IF you compare TODAY games from the same timeline from kepler versus GCN 1.1 you will find than the 780Ti still perform better than the 290X because the games are optimizations were made to "actual" and "modern" hardware for their time, same as games from the HD7970 versus GTX 680... same from 980 to 390X.. and same shit will happens with AMD once they release VEGA, ALL of their game and drivers optimizations will be based for and to VEGA cards, people will be lucky if any of those optimizations benefit older cards due architecture similarities. can we say then AMD dropped Hawaii and Fiji because polaris?. ;)
 
Look yea Volta will be faster now doubt. But people here were making it seem like it's gonna be second coming of God! Like it was goons be 100-200% faster. We are looking at 40% performance increase it seems. Big ass die I was expecting more. I am sure it will do good. Looks like pascal with more cores. I don't see much ipc increase other than more cores. Not saying it's bad but hype was unreal for Volta.
 
NKD said:
Look yea Volta will be faster now doubt. But people here were making it seem like it's gonna be second coming of God! Like it was goons be 100-200% faster. We are looking at 40% performance increase it seems. Big ass die I was expecting more. I am sure it will do good. Looks like pascal with more cores. I don't see much ipc increase other than more cores. Not saying it's bad but hype was unreal for Volta.
Click to expand...

big role of that die size are due the new tensor Core those are huge, if you delete de tensor core from the SM you will have way smaller SM so way smaller Die size.
 
NKD said:
Look yea Volta will be faster now doubt. But people here were making it seem like it's gonna be second coming of God! Like it was goons be 100-200% faster. We are looking at 40% performance increase it seems. Big ass die I was expecting more. I am sure it will do good. Looks like pascal with more cores. I don't see much ipc increase other than more cores. Not saying it's bad but hype was unreal for Volta.
Click to expand...

NKD you miss some big points.
The die increased by 33.6% but compute rating has increased by 41.5%, or 2x (fp32) or 4x (fp16) if able to make use of the Tensor cores/matrix function, and all in the same TDP.
Explain how that is not an improvement?
In general Nvidia has a chart showing 50% performance over P100 across a range of real world HPC operations, but hang on this is same TDP and only 33.6% increased die :)
Those latter HPC application results are not using Tensor matrix functions.
Oh and within that same TDP we have 50% more NVLink connectors each with 50GB/s bandwidth and also 25% more HBM BW.
And you really do not think these changes are impressive when it is same node albeit latest iteration?
And that is ignoring aspects such as Thread Scheduling or Cache performance behaviour changes.
Cheers
 
Last edited:
NKD said:
We will see how it all translates to gaming.
Click to expand...
What just like the P100 and Pascal GTC 2016 presentations that had all the talk about arch improvements and was purely HPC focused including heavy focus on NVLink 1/HBM2/FP64/mixed precision/etc?

How did it pan out for consumer version with GTX1070/1080/1080ti/Titan?
They did rather well wouldn't you say?
But this point is different to your post before and that was you being critique of just how well (or not in your opinion) Volta improves over Pascal.
Anyway pretty clear there is efficiency/IPC improvements over Pascal.
This change is in many ways comparable to Kepler-->Maxwell arch jump.
Cheers
 
Last edited:
CSI_PC said:
What just like the P100 and Pascal GTC 2016 presentations that had all the talk about arch improvements and was purely HPC focused including heavy focus on NVLink 1/HBM2/FP64/mixed precision/etc?

How did it pan out for consumer version with GTX1070/1080/1080ti/Titan?
They did rather well wouldn't you say?
But this point is different to your post before and that was you being critique of just how well (or not in your opinion) Volta improves over Pascal.
Anyway pretty clear there is efficiency/IPC improvements over Pascal.
Cheers
Click to expand...

We will see. If we go by pure tflops. It won't be as impressive as going from maxwell to pascal. I mean 1080 was literally 70-80% faster than 980. I doubt we will see those gains. But I am sure 40-50% will be there to have. It may not be worth the upgrade over my 1080ti.
 
NKD said:
We will see. If we go by pure tflops. It won't be as impressive as going from maxwell to pascal. I mean 1080 was literally 70-80% faster than 980. I doubt we will see those gains. But I am sure 40-50% will be there to have. It may not be worth the upgrade over my 1080ti.
Click to expand...


That was a node shrink and so not really comparable in same way.
This is an arch improvement on same node and that is what you need to consider, to do 45%-55% with same TDP on same node is massive especially when that is without the optimisation possible for Volta!
Even if this the latest iteration of 16nm.
Imagine if this is then shrunk we would get those said additional gains on top as well and it looks like this latest arch is even more suited to such a node shrink cycle.
Nvidia is really pushing the tech with Volta in terms of efficiency and performance/IPC-functions.
Cheers
 
Last edited:
well theoretical flop numbers and frequency aside, Volta is probably faster than 40%-50% faster than Pascal. nV is expecting to have what x2 the perf/watt with Volta, if I remember correctly? If they get anything close to that even 33% of that (which is just the improvements of the 16nm node to the "new" 12nm node) its going to end up like the same going from Maxwell to Pascal.

We still don't know what the new Tensor units are like and how they are structured and how the gaming GPU's will use it but just a hint of the variable scheduling, we can assume the efficiency of Volta is going to be quite a bit higher.

Just think about rapid pack math for Vega, but Volta is capable of doing that and full precision on the same units at any time without any sort of performance hit that we see with Pascal.

The amount of unit increase with a frequency increase which, nV did state they will be increasing frequencies (in V100 they are not since its such a large chip), its pretty safe to assume they will hit higher than 40% performance, and likely hood of hitting more than 50% looks to be good too. Just the unit increase amount looks to be around 50%.
 
Last edited:
razor1 said:
well theoretical flop numbers and frequency aside, Volta is probably faster than 40%-50% faster than Pascal. nV is expecting to have what x2 the perf/watt with Volta, if I remember correctly? If they get anything close to that even 33% of that (which is just the improvements of the 16nm node to the "new" 12nm node, its going to end up like the same going from Maxwell to Pascal.

We still don't know what the new Tensor units are like and how they are structured and how the gaming GPU's will use it but just a hint of the variable scheduling, we can assume the efficiency of the Volta is going to be quite a bit higher.
Click to expand...

Tensor for matrix function is meant to be around 2x performance gain for FP32 and 4x for FP16 with real world frameworks/apps.
It may have very select uses, but that is great improvement.
CHeers
 
CSI_PC said:
Tensor for matrix function is meant to be around 2x performance gain for FP32 and 4x for FP16 with real world frameworks/apps.
It may have very select uses, but that is great improvement.
CHeers
Click to expand...


I think it can be used in a lot of apps :) mainly because I think its the new scheduler that is capable of enabling of Tensor cores, gotta remember, a core is a core, how they are utilized and the way they are fed is much more important in graphics than just their capabilities as stand alone units, hence the increased register space too.

This is what nV has been doing with their GPU's since Tesla, they aren't just adding units, but the do that and how those units interact with each other. This is very noticeable with Maxwell on the same node as Keplar. Fermi was a bust because there were some fundamental changes in the architecture that increased power consumption to the point it was a over board, but as we can see the performance was still there. And if we start looking at Fermi and Tesla, the performance improvements would have been pretty good if that power consumption wasn't so much out of reach of the competition. When looking at CUDA based apps I remember seeing 300% to 400% increase in performance of Fermi from Tesla. It was just a misstep nV had. And this had nothing to do with the hardware scheduler something more fundamental than that caused the power consumption issue.
 
Last edited:
The "I'm going to wait until X releases" crowd is going to eat this up. Pushing back their upgrade like they were going to do it but this came up.
Still trying to find a replacement for that Kepler card.
 
:D Hey look, you can get a founders edition Volta card for the low low price of 149,995. Sounds about right ;)
 
You must log in or register to reply here.
Back
Top