NVIDIA Volta Rumor Thread

[zuzu]

by swiss lab test ,gp100 vs gv100 gv100 cuda core is 1-3% slower
other stuff are same ,plus gv100 just have more cores and is faster from amd vega cca 20%
in deep learning etc stuff ,and minus is vega can use ssd like L4 memory .gv100 cant
second gddr6 will be rdy in 1q 2018 and cards at best q2 2018 (that mean gv104 +)
and last gp104 vs gv104 just moar cuda cores (same cores this is just 12nm shrink to be on same or + lvl with rx vega )
nvidia is scared of rx vega ,so they pay for custom 12nm to beat it
with 20-30% more cores and +10% just from shrink
so i hope gv102 (rumor say it's cca 600mm2 and very difficult to make on 12nm)
will be at least summer time and 50% over last titan
7nm will bring new dx12/vulcan compatible core arch
and gp100 tech is useless for gaming ,so they have 2 teams one for gx 100 pro
and one that most of time do shrink and optimisse same dx11 maxwell cuda cores )
i hope volta will bring new cuda arch but 12nm proces will be just tooo expensive
one word we got from fab is in cca 100 waffers only 2 good gv100 chips ,good is usable
like gp100 (max 70% of cores usable)
so not good ...

 

[zuzu]

oh one more ,,dont know why amd slash drivers of fe vega
but some test numbers for rx vega are scarryy
we all know that gaming bio is dx11 90% ,and most of it are tile raster cuda optimised
and now vega use same tech + some new dx12 tech that we don't have in pascal
i hope it will be flop like fiji but ,nvidia wont invest millions for short process like 12nm if rx vega
is flop ,or drop prices of quadro boards under 1k $
great for us but ...
hope navi will be just refresh and new dx12 cuda arch smash them again

 

[Puffnstuff]

After seeing Vega fall flat upon launch I'm more interested in Volta as it will probably be in my next build.

 

[zuzu]

problem is they never launch rx vega .... and by all these pro app tests titan xp is in pain )

 

[TheF34RChannel]

Glad I found this thread! So what's the latest and greatest speculation on release dates concerning Titan and '2080 (Ti)' consumer cards?

 

[zuzu]

i hope never ,it's 600+mm2 hot beast of chip .titan xp take close to 400w in oc mode and wont drop under 80c ,wonder what will double size chip do ..... they need wait 7nm ,,but it's hot now under legs ... striking from intel side and now from amd side is bad

 

[ddogg]

After seeing Vega fall flat upon launch I'm more interested in Volta as it will probably be in my next build.

That wasn't really Vega but yeah, I'm not holding my breath for when the gaming card comes out later this month.

 

[Puffnstuff]

That wasn't really Vega but yeah, I'm not holding my breath for when the gaming card comes out later this month.

When rx debuts I'm not really expecting anything to change except for the card color and the excuses used for a lack of performance. If the market forces pricing down like it did with ryzen then consumers will benefit from it. Its been a long time since I've used an AMD gpu and it will probably stay that way.

 

[zuzu]

how i read ,fe vega have only full pro drivers ,,gaming ones are demo ,with manny new stuff disabled
so rx vega cause nvidia to flush patettic 12nm with only 15% shrink ,will that be good for us who knows ..

 

[TheF34RChannel]

When rx debuts I'm not really expecting anything to change except for the card color and the excuses used for a lack of performance. If the market forces pricing down like it did with ryzen then consumers will benefit from it. Its been a long time since I've used an AMD gpu and it will probably stay that way.

Same old, same old then eh. Me too, all ready for consumer Volta whenever it may land and I cannot wait for the leaks to surface.

 

[zuzu]

may i ask why ?? it's same gpu like titan ti or dx11 ??
i mean why upgrade from dx11 to dx 11 just a bit faster ??
i will wait 7nm dx12 arch ,before upgrade ,,this volta will be 4th dx11 gpu they sell us

 

[Puffnstuff]

may i ask why ?? it's same gpu like titan ti or dx11 ??
i mean why upgrade from dx11 to dx 11 just a bit faster ??
i will wait 7nm dx12 arch ,before upgrade ,,this volta will be 4th dx11 gpu they sell us

Well in my case I'm still running a gtx 780 ti so anything newer is still an upgrade for me.

 

[Pinstripe]

may i ask why ?? it's same gpu like titan ti or dx11 ??
i mean why upgrade from dx11 to dx 11 just a bit faster ??
i will wait 7nm dx12 arch ,before upgrade ,,this volta will be 4th dx11 gpu they sell us

Where did you get this information from that the 7nm chip will be the first true DX12 arch?

7nm AFAIK isn't ready until 2019.

 

[zuzu]

tsmc say 1st half of 2018 soc ,second part of year gpu
goflo say winter 2017 mid 2018 for cpu/gpu rdy
sammy say 2018 7nm 5nm 2019

and dx12 arch ,,volta wont be one ,they tape gpu with same cuda cores ...
so if they dont make dx12 arch on 7nm or after volta , rip nvidia

 

[TheF34RChannel]

Well in my case I'm still running a gtx 780 ti so anything newer is still an upgrade for me.

Oh wow! I could never do that because newly released GPUs always make me itchy so I have to buy one every generation. Quickest path to immediate performance increase.

 

[Puffnstuff]

Oh wow! I could never do that because newly released GPUs always make me itchy so I have to buy one every generation. Quickest path to immediate performance increase.

I can't afford to do that so I hold on to what I've got. Maybe in the future I'll be able to obtain a current generation card.

 

[TheF34RChannel]

I can't afford to do that so I hold on to what I've got. Maybe in the future I'll be able to obtain a current generation card.

Completely understandable! I have to save up myself as well so don't make such purchases lightly. I skipped the 600 series though. Selling the current card on time goes a long way though, and I'll do just that to get my much awaited Volta sweetness.

Shouldn't we by now also get the codename of its successor?

 

[TheF34RChannel]

NVIDIA's Future MCM GPU Design Detailed - 256 SMs With Over 16000 Cores, Multiple TeraBytes of Bandwidth

https://www.overclock3d.net/news/gp...-chip_gpu_modules_to_scale_past_moore_s_law/1

 

[Puffnstuff]

NVIDIA's Future MCM GPU Design Detailed - 256 SMs With Over 16000 Cores, Multiple TeraBytes of Bandwidth

https://www.overclock3d.net/news/gp...-chip_gpu_modules_to_scale_past_moore_s_law/1

Thanks for sharing that information. It only makes sense that they'd also have to implement this kind of system to facilitate different performance amongst their product offerings. Next I'd like to see a socket cooler bracket akin to that of a cpu to facilitate swapping out the cooling. I'm still hoping that RX Vega will force team green to lower prices on their stuff so we can all benefit from it.

 

[Samwell]

It's just research at the moment. Who knows when it might come into chips. I wouldn't expect Nvidia to do it before the 5nm node in 2021. Next year Volta as the big architectural step. 2019 just a shrink at 7nm with small enhancements like pascal and then maybe such a concept.

 

[Puffnstuff]

It's just research at the moment. Who knows when it might come into chips. I wouldn't expect Nvidia to do it before the 5nm node in 2021. Next year Volta as the big architectural step. 2019 just a shrink at 7nm with small enhancements like pascal and then maybe such a concept.

At least they consistently move forward leading the market segment for performance. AMD does have leadership in the hype department.

 

[TheF34RChannel]

It's just research at the moment. Who knows when it might come into chips. I wouldn't expect Nvidia to do it before the 5nm node in 2021. Next year Volta as the big architectural step. 2019 just a shrink at 7nm with small enhancements like pascal and then maybe such a concept.

Most probably, yes. We'll get some nice cards in between in that case. But I don't see any other realistic way forward.

I hope they'll develop what they paper proposed.

 

[Samwell]

Most probably, yes. We'll get some nice cards in between in that case. But I don't see any other realistic way forward.

I hope they'll develop what they paper proposed.

Long term it might be the only possibility, but i'm not so enthusiastic as other people about that concept. People tend to forget all the disadvantages of it. I read the paper myself and to get within 10% of a monolythic chip they had to add links with 768Gb/s speed for each dia. This will take a bit of die space, but the bigger problem is you're loosing energy efficiency. Moving data over the pcb costs quite a bit energy compared to moving data on chip. That's why they're comparing the energy efficiency only with a multigpu system on the same board, where the multidie of course wins. But compared to a monolythic chip it would be much more interesting. I'm pretty sure that they also calculated that, but then you would see the disadvantage of the concept, so they left the information out of the paper.

 

[TheF34RChannel]

Could it be a case of needing to take a step back in order to prevail over a monolithic chip a few generations later? A strange musing really. Wouldn't make sense if the new design would lose performance over the old one, temporary or not.

 

[zuzu]

did they say how will prevent sli problem with multi gpu thing ?
sli problem 1 gpu 100% 2 gpu 70% 3 gpu 50% etc