Discussion: Nvidia Pascal

[renz496]

Heard that nvidia double the VRAM without actually naking it more expensive than the 12GB version. Plus just because nvidia refeshing their fastest quadro it doesn't mean it has something to do with pascal. Remember with pascal nvidia bring in the double presicion stuff back. And unlike AMD nvidia pro line up is consist of tesla and quadro. And both serve two different market. So it is possible quadro based pascal will not coming out soon but it doesn't mean tesla based pascal also going to be delayed. It is possible nvidia pascal has been booked early by HPC customer like what happen to GK110 making most of the available chip go to HPC customer first.

 

[mr91]

You might be correct - do you think a Telsa card with HBM2 or a Titan card will be released in the next couple of months?

 

[renz496]

hard to say actually. it seems HBM2 will not be on time if we want to see the card in Q3 or even Q4 this year. seen one of the article that hynix will start volume production of 4GB HBM2 sometime in Q2 or Q3. why need to specifically to mention 4GB only? probably high capacity HBM2 got delayed further? it might be possible to see the the geforce part (including titan) getting GDDR5/X while only the tesla part get HBM (like the leaked chart suggested). one things that is very clear though: nvidia need pascal out ASAP especially the one that offering double precision because of intel threat in HPC segment with their Phi. and this is where it gets interesting actually. initially nvidia intend to use HMC for pascal. to be honest i would not be surprise if nvidia tesla based pascal end up using HMC while geforce and quadro variant end up using GDDR5/X. some said it is impossible because that's mean nvidia have to design two chip with different memory controller to make it happen and historically nvidia did not make specific design that only cater on one segment only. but recently nvidia did make GK210 that is specifically made for tesla only. so such a thing is not totally impossible with nvidia. and recently there are speculation about the existence of 'GP102'.

 

[AdmiralDonut]

hard to say actually. it seems HBM2 will not be on time if we want to see the card in Q3 or even Q4 this year. seen one of the article that hynix will start volume production of 4GB HBM2 sometime in Q2 or Q3. why need to specifically to mention 4GB only? probably high capacity HBM2 got delayed further? it might be possible to see the the geforce part (including titan) getting GDDR5/X while only the tesla part get HBM (like the leaked chart suggested). one things that is very clear though: nvidia need pascal out ASAP especially the one that offering double precision because of intel threat in HPC segment with their Phi. and this is where it gets interesting actually. initially nvidia intend to use HMC for pascal. to be honest i would not be surprise if nvidia tesla based pascal end up using HMC while geforce and quadro variant end up using GDDR5/X. some said it is impossible because that's mean nvidia have to design two chip with different memory controller to make it happen and historically nvidia did not make specific design that only cater on one segment only. but recently nvidia did make GK210 that is specifically made for tesla only. so such a thing is not totally impossible with nvidia. and recently there are speculation about the existence of 'GP102'.

SK Hynix are not the only producers of HBM2. In fact, SK Hynix are not even the main suppliers for NVIDIA, Samsung are, and their HBM2 production is already in full swing. Mass production has already begun, with 4GB chips already available, and the 8GB chips due somewhere around September time.

 

[stavrosmast]

Lets just wait 2 weeks for further Info...

 

[mr91]

hard to say actually. it seems HBM2 will not be on time if we want to see the card in Q3 or even Q4 this year. seen one of the article that hynix will start volume production of 4GB HBM2 sometime in Q2 or Q3. why need to specifically to mention 4GB only? probably high capacity HBM2 got delayed further? it might be possible to see the the geforce part (including titan) getting GDDR5/X while only the tesla part get HBM (like the leaked chart suggested). one things that is very clear though: nvidia need pascal out ASAP especially the one that offering double precision because of intel threat in HPC segment with their Phi. and this is where it gets interesting actually. initially nvidia intend to use HMC for pascal. to be honest i would not be surprise if nvidia tesla based pascal end up using HMC while geforce and quadro variant end up using GDDR5/X. some said it is impossible because that's mean nvidia have to design two chip with different memory controller to make it happen and historically nvidia did not make specific design that only cater on one segment only. but recently nvidia did make GK210 that is specifically made for tesla only. so such a thing is not totally impossible with nvidia. and recently there are speculation about the existence of 'GP102'.

Interesting - Hopefully we will see a HBM 2 Titan Flagship launched in tandem with the GP104...

 

[renz496]

If samsung can be ahead of Hynix offering high capacity HBM2 that would be interesting.

 

[Mrinal_2]

Any latest news on pascal gpus??? I am so tired of waiting....dammit.....really need a gpu upgrade

 

[stavrosmast]

Patience is the key 😀

 

[renz496]

at this point just wait for nvidia GTC event next month

 

[Non-Euclidean]

I think the funniest thing is seeing people parrot "10x faster"

It wont be 10 times quicker.

Its PR additive math

"The idea is that if we look at all the improvements coming up with Pascal compared to Maxwell, they will collectively add up to make it “roughly” 10 times more efficient at deep learning compute tasks. Pascal will feature 3x the memory bandwidth of Maxwell, 2x peak single precision compute performance and 2x the performance per watt."
http://wccftech.com/nvidia-pascal-gpu-gtc-2015/

 

[renz496]

Most people just pay attention to that "10x faster" and did not look at further details. They should watch the actual keynote when JHH actually on stage explaing about that 10 times faster stuff.

 

[Mrinal_2]

I think the funniest thing is seeing people parrot "10x faster"

It wont be 10 times quicker.

Its PR additive math

"The idea is that if we look at all the improvements coming up with Pascal compared to Maxwell, they will collectively add up to make it “roughly” 10 times more efficient at deep learning compute tasks. Pascal will feature 3x the memory bandwidth of Maxwell, 2x peak single precision compute performance and 2x the performance per watt."
http://wccftech.com/nvidia-pascal-gpu-gtc-2015/

That is still amazing i guess..didnt get most of it though..but still gives me nerdgasm 😛

 

[imrazor]

2x the performance per watt."
http://wccftech.com/nvidia-pascal-gpu-gtc-2015/

Hmm, does that translate to "2x the framerate at the same resolution"? If so, it's still exciting. Too bad 10x is just marketing BS.

 

[TechyInAZ]

2x the performance per watt."
http://wccftech.com/nvidia-pascal-gpu-gtc-2015/

Hmm, does that translate to "2x the framerate at the same resolution"? If so, it's still exciting. Too bad 10x is just marketing BS.

At this time right now, a 10x jump in performance is nearly impossible. Now, if we were back in the mid 2000s, sure that could be believeable, but with how gradual the performance boosts of all next gen products have become, 10x is only fantacy.

 

[renz496]

even seeing double the performance in one go is no longer possible now. instead gpu maker go the route that will enable them to maximize profit: releasing new parts that only increase in 10%-20% over the previous generation.

 

[turkey3_scratch]

GPU and CPU makers both know they will soon hit physical size limitations of transistors, so they are going to hold back on having large jumps in performance so that they can prolong the years to come before there is a dead end in further technological advancements.

 

[stavrosmast]

^^^^^

 

[turkey3_scratch]

And then programmers can no longer be lazy, they'll have to optimize the crap out of software like the 80s and 90s.

 

[mr91]

even seeing double the performance in one go is no longer possible now. instead gpu maker go the route that will enable them to maximize profit: releasing new parts that only increase in 10%-20% over the previous generation.

Seems like approximately every 2 years give or take the performance doubles...

680-980

780 - 980 ti

I believe we already came to a consensus in beginning of this thread that 10x is not realistic and most likely applies to the the Architectures potential in deep learning...

Competition From AMD and the fact that the PC platform is beginning to have a larger audience is the reason we're seeing such huge increase.

CPU's don't double in performance ever two years...

Finally the fact that compute cards DNA are being recycled for the gaming cards...

 

[stavrosmast]

The only thing I that I hate with what nvida does is that They hype all us with the Some more billion transistors and the HBM 2 in June and then they are like ohhps Sorry this year you will only have GDDR5/X and in 2017 we will release the really expensive hbm2 ones..

 

[stavrosmast]

Also

http://techfrag.com/2016/03/25/rumour-nvidia-pascal-facing-problems-with-asynchronous-compute/

 

[Nohuhhu]

The only thing I that I hate with what nvida does is that They hype all us with the Some more billion transistors and the HBM 2 in June and then they are like ohhps Sorry this year you will only have GDDR5/X and in 2017 we will release the really expensive hbm2 ones..

Did they actually say that EVERY GPU will have GDDR5/X? That none has HBM this year?

 

[renz496]

To me it doesn't matter if the cards having HBM or GDDR5. As long as the performance is there then it should be fine. Right now both amd and nvidia have node shrink to help reducing power consumption. And about pascal having trouble with async compute honestly i don't know what to say yet. Because sometimes nvidia did come out with the unexpected result. For example many believed back in 2012 that 680 will end up more power hungry than 7970 and they say it is impossible for nvidia to have 6Ghz GDDR5 given how bad their GDDR5 memory controller are compared to AMD back in fermi days. And with fermi many believed that nvidia will never going to beat amd in tessellation because it is the first time for nvidia while amd already have the experience ever since with their DX10 hardware.

 

[imrazor]

It seems Intel had a lot of trouble getting good yields of 14nm chips. Will NVidia (and AMD) have similar problems, or has everything been ironed out?