Nvidia volta gpu tested

[vanished]

3 hours ago, pyrojoe34 said:

Your math is backwards btw. It’s a 42.8% increase in CUDA cores and a 50% increase in FP32/64. What you calculated was the decrease from Volta to Pascal, not the increase from Pascal to Volta. 

Do we know if clock speed has changed?  Because with those numbers, it means they're doing about 17% more with each core.  Now, how that breaks down into a clock speed change and an IPC change I don't yet know.

[Dackzy]

9 minutes ago, CostcoSamples said:

I disagree.  AI for games is just around the corner and it will require hardware acceleration.  The need for more GPU power will not diminish, and die size will forever be dictated by manufacturing cost vs consumer demand.  My point is simply that manufacturing cost will diminish.

We already use AI in games....

[vanished]

5 minutes ago, Dackzy said:

We already use AI in games....

Well, that's a pretty broad topic.  There's been very basic AI in games for a long time, but I think he's talking about more sophisticated stuff like machine learning and neural nets, etc.

[Dackzy]

3 minutes ago, Ryan_Vickers said:

Well, that's a pretty broad topic.  There's been very basic AI in games for a long time, but I think he's talking about more sophisticated stuff like machine learning and neural nets, etc.

Yep the AI in games is very basic. Just wanted to make it clear that we already use AI and we actually have used AI in games for a long ass time. 

[AresKrieger]

 

Seriously we don't need another BS guess what leaks are real run like last time, odds are all of this info is either fake, irrelevant or both so can we at least wait till some card actually is shown before guessing its performance

[leadeater]

6 hours ago, Misanthrope said:

Sorry but this is likely the big chip and a synthetic bench: Real world it will be a bit of a jump, maybe 10 to 20% jump in performance vs current Pascal cards but it won't "revolutionize" shit.

 

Because even if it did, Nvidia would just keep that big Volta chip in their back pockets and release a cut version. In Ken's own word they don't have to do much better when AMD Radeon is just so behind.

Those big jumps in performance shown are all from the Tensor cores that have nothing to do with gaming and won't even be on the gaming dies. It's no more than 50% faster since the the theoretical compute performance figures shown are just that, it won't scale perfectly of course.

[CostcoSamples]

34 minutes ago, Ryan_Vickers said:

Well, that's a pretty broad topic.  There's been very basic AI in games for a long time, but I think he's talking about more sophisticated stuff like machine learning and neural nets, etc.

Just because game devs call it AI doesn't make it AI...   Also the so called game "AI" (which is basically just a collection of if-then statements) does not require GPU acceleration.  

[vanished]

1 minute ago, CostcoSamples said:

Just because game devs call it AI doesn't make it AI...   Also the so called game "AI" (which is basically just a collection of if-then statements) does not require GPU acceleration.  

I mean, all intelligence of every kind is just a series of binary decisions like that  It's just a matter of how many there are and how complex they are, etc.

[leadeater]

13 hours ago, CostcoSamples said:

This massive V100 die proves a point I tried to make a while back:  Large dies are the future.  We are maybe 5 years away from the limits of silicon, which most agree is around 5nm.  Eventually, yields on 5nm will be so good that it will be easy to make larger and larger dies.  They might even increase wafer size up to 600 mm or larger.  Today, 815mm² is considered massive and only for super computers, but I think we'll see consumer GPUs that size or larger, eventually.

I don't think you watched the Nvidia Volta announcement then where they stated they only get one working V100 die from an entire wafer, that's not the future of anything. You can go back through the history of both companies and look at the die sizes, neither of them have wanted to go over the 600mm2 die size and for good reason, fab yields are terrible.

 

Node shrinking only makes yields go up be because there are more dies per wafer. Reducing defects per mm2 is not effected by node size, not directly anyway. Those improvements come from improvements in fabrication processes and equipment.

 

Either way you look at it the larger the die the lower the yields and that has held true right through to today and there is no sign of that ever changing. All Volta dies in GeForce cards will be below 600mm2, that I am very confident on.

 

 

[MeDownYou]

14 hours ago, leadeater said:

I don't think you watched the Nvidia Volta announcement then where they stated they only get one working V100 die from an entire wafer, that's not the future of anything. You can go back through the history of both companies and look at the die sizes, neither of them have wanted to go over the 600mm2 die size and for good reason, fab yields are terrible.

 

Node shrinking only makes yields go up be because there are more dies per wafer. Reducing defects per mm2 is not effected by node size, not directly anyway. Those improvements come from improvements in fabrication processes and equipment.

 

Either way you look at it the larger the die the lower the yields and that has held true right through to today and there is no sign of that ever changing. All Volta dies in GeForce cards will be below 600mm2, that I am very confident on.

 

 

Let's not forget the inherent benefit of a smaller die on top of that which is more die per wafer, so closer to a 2x the good small dies out of a single wafer

 

Also I appreciate you posting this as it breaks it down quite nicely with no guess work

[CostcoSamples]

17 hours ago, leadeater said:

I don't think you watched the Nvidia Volta announcement then where they stated they only get one working V100 die from an entire wafer, that's not the future of anything. You can go back through the history of both companies and look at the die sizes, neither of them have wanted to go over the 600mm2 die size and for good reason, fab yields are terrible.

 

Node shrinking only makes yields go up be because there are more dies per wafer. Reducing defects per mm2 is not effected by node size, not directly anyway. Those improvements come from improvements in fabrication processes and equipment.

 

Either way you look at it the larger the die the lower the yields and that has held true right through to today and there is no sign of that ever changing. All Volta dies in GeForce cards will be below 600mm2, that I am very confident on.

 

 

Thanks for the formula.  My point is that the number of defects per unit area tend to decrease the longer a node is in production. If for example they stayed at 16nm for the next 20 years, we can safely assume defects per unit area will be a very small number.  If somehow they could get that number to zero (probably impossible), you could make a die that uses the entire wafer!