Nvidia volta gpu tested

[Giantpie12]

https://www.game-debate.com/news/23739/first-nvidia-tesla-volta-v100-gpu-benchmarks-reveal-more-than-double-pascal-performance#discussForm

 

according to this article the next gen Nvidia gpus completely blow the pascal line up out of the water, if it is true we would be seeing a leap in gpu performance not seen in quite a while imo.

Quote

Nvidia Tensor Cores are new to Volta and are specially designed for AI training and inference applications. They’re mixed-precision FP16/FP32 cores specifically designed for deep learning, offering up to 6x the performance of previous generation Pascal P100’s FP16 operations, and 12x the performance in terms of FP32 operations. This is a specialised machine learning device that is several times faster than the current-gen, although this AI optimisation won’t feed through directly to gaming performance. The end result is a GPU that’s incredibly fast for Deep Learning; an AI dream for many research institutions.

 

Quote

As for the Nvidia DGX-1, preliminary benchmarks give it a score of 743,537 in single-core Geekbench 4 compute tests. The second fastest system ever recorded in the benchmark database is the current P100 system, which scores ‘just’ 320,031. A workstation PC with nine Quadro GP100 graphics cards can score 278,706. That’s an astounding improvement, more than doubling the peak benchmark score within a single generation.

 

-Thanks @TVwazhere

Edited September 22, 2017 by TheRandomness
Fixed.

[PCGuy_5960]

Geekbench? Really? Please ban this benchmark, it sucks!

[Ryujin2003]

I just realized why he never quoted. That article is complete garbage. That's not tested. It's speculation with nothing else. No deep dive analysis, legit benchmarks, or actual performance outputs... Click bait title.

And Geek Bench?

 

I guess there last question is can it run Crisis or  Mine?

[NeoAthlon]

None of this is new news, these are tesla cards, not video cards. The die size on these makes them extremely cost prohibitive for gpu use.

[Alariel]

Yeah I'm gonna wait for some more sources and a different bench on this one.

[TheRandomness]

But then you realise the V100 has 5120 CUDA cores, compared to the 3584 of the P100. That's a 30% increase in CUDA cores. Now, comparing FP64/FP32, you'll see that those have increases of 33%. So basically, according to some really dodgy 'benchmark' Volta has minimal IPC improvements.

 

EDIT: Lol I can't do maths.

13 minutes 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. 

[goodtofufriday]

 

Is this really the same class of gpu die they are comparing? it seems like their super high end used in machine learning applications. Pretty damn sure its not gonna be a 2080

[NeoAthlon]

https://www.anandtech.com/show/11367/nvidia-volta-unveiled-gv100-gpu-and-tesla-v100-accelerator-announced

 

pretty much all the same info, from 4 months ago

 

 

[Princess Luna]

It'd be nice if the Volta TITAN X was the full chip since begin like in Maxwell family... might be a worth purchase on day one release... I walked away from buying a 1080ti for myself to replace the TITAN XM in this expectation... if by some chance it is a crippled chip again like happened in Pascal then I'll go GTX 2080...

[Guest]

24 minutes ago, PCGuy_5960 said:

Geekbench? Really? Please ban this benchmark, it sucks!

Geekbench is pretty accurate, within 100-200% margin of error... 

[Guest]

1 minute ago, Tedny said:

people, that won't be consumer Card, at least not the first generation.... like it was with 400-700 series 

don't tell me that I can't buy a thousand dollar gpu to play minecraft 

[manikyath]

1 minute ago, deXxterlab97 said:

Geekbench is pretty accurate, within 100-200% margin of error... 

seeing how some houses are built.. thats acceptable margin in some sectors

[System Error Message]

volta will have higher clocks too. Hence the 50% increase.

 

But the point is that these arent game benchmarks, and its a question if the consumer cards will get the same thing and if we can use them in our codes and games. It would be a big improvement for game AI.

[TidaLWaveZ]

24 minutes ago, goodtofufriday said:

Is this really the same class of gpu die they are comparing? it seems like their super high end used in machine learning applications. Pretty damn sure its not gonna be a 2080

I'm not too sure about the validity of the data but if accurate the p100 and v100 would be in the same class (professional/tesla/datacenter type).  You're correct, super high end - not Geforce.

[Dylanc1500]

I hate to burst bubbles but this is information that has been publicly available since May. @leadeater and I had a coversation about GV100 in a previous thread from a couple months ago. NVidia has it all published on this page and has since May : https://devblogs.nvidia.com/parallelforall/inside-volta/

[Giantpie12]

4 minutes ago, Dan Castellaneta said:

No sources within the articles, no meaningful benchmarks, all speculation. I thought Volta wasn't dropping until later next year anyways.

Seriously, tech news is really as bad, if not worse than cable news when it comes to this garbage.

there are multiple sources citing the same tests that were run

[Misanthrope]

 

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.

[CostcoSamples]

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.

[DildorTheDecent]

>Game debate

 

kek. 

[CostcoSamples]

1 hour ago, VegetableStu said:

the future looks to be extremely expensive ._.

Good yields and larger wafers will bring prices down a lot...  It will happen, mark my words!  

 

In 1999 the GeForce GPU was around 80 mm².  Today, the GTX 1080 ti is 471mm², and even that is not considered to be really "large".  Chips have been getting bigger for a long time.  That trend will accelerate once we hit the 5nm wall.

[pyrojoe34]

2 hours ago, TheRandomness said:

But then you realise the V100 has 5120 CUDA cores, compared to the 3584 of the P100. That's a 30% increase in CUDA cores. Now, comparing FP64/FP32, you'll see that those have increases of 33%. So basically, according to some really dodgy 'benchmark' Volta has minimal IPC improvements.

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. 

[TheRandomness]

12 minutes 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. 

Whoops

Yeah it's been a while since I've messed with percentages and such. So I'll just quote that in my op.

[Dylanc1500]

Here are some better charts directly from NVidia that directly compare it with Pascal.

[MeDownYou]

3 hours ago, CostcoSamples said:

Good yields and larger wafers will bring prices down a lot...  It will happen, mark my words!  

 

In 1999 the GeForce GPU was around 80 mm².  Today, the GTX 1080 ti is 471mm², and even that is not considered to be really "large".  Chips have been getting bigger for a long time.  That trend will accelerate once we hit the 5nm wall.

Eh smaller chips mean more chips per wafer we moved away from large sizes and are moving back into them, possibly for the last time or two. We will then get smaller and more efficient again until we hit a point we're node size doesn't shrink for a while and we make bugger chips to compensate, however I'm guessing in 15 years we will be at the point were moderate cards are fine for 4k like cards are for 1080p today. This continues until we have cards like that for the resolution that we can't tell the difference between

[CostcoSamples]

57 minutes ago, MeDownYou said:

Eh smaller chips mean more chips per wafer we moved away from large sizes and are moving back into them, possibly for the last time or two. We will then get smaller and more efficient again until we hit a point we're node size doesn't shrink for a while and we make bugger chips to compensate, however I'm guessing in 15 years we will be at the point were moderate cards are fine for 4k like cards are for 1080p today. This continues until we have cards like that for the resolution that we can't tell the difference between

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.