NVIDIA Volta Rumor Thread

[Arachnotronic]

xpea ..tests say gp100 vs gv100 core to core ,,they are same
so dont know what u read but volta is pascal with moar cores
i wish it was dx12 in hw but ,,

They're not the same core-to-core. The SM is totally different as is the very ISA that those SMs run.

 

[xpea]

xpea ..tests say gp100 vs gv100 core to core ,,they are same
so dont know what u read but volta is pascal with moar cores
i wish it was dx12 in hw but ,,

Do you understand how an ALU works ? Of course Pascal and Volta offer same performance core to core (at same speed). I maybe shock you if I say that Fermi has also same performance core to core than Volta... and its normal ! However, is Fermi same than Volta and can reach same final performance ?
A good read for your education on Volta :
https://devblogs.nvidia.com/parallelforall/inside-volta/
Than tell me its same than Pascal

 

[zuzu]

i guess u dont understand that all of these use kepler sm's
but kepler is old process full of bugs ,pascal is 16nm volta 12 ,so perf came from sm number and clock fixes and not new cores
i read that swiss lab test gp100 vs vp100 cards ,on same clock i think 1ghz
same perf minus 1-2% for volta - extra cores in volta
so even slower

 

[xpea]

i guess u dont understand that all of these use kepler sm's
but kepler is old process full of bugs ,pascal is 16nm volta 12 ,so perf came from sm number and clock fixes and not new cores
i read that swiss lab test gp100 vs vp100 cards ,on same clock i think 1ghz
same perf minus 1-2% for volta - extra cores in volta
so even slower

Please don't embarrass yourself and try to learn something. You waste everybody's time
Have a nice day

 

[zuzu]

you 2 ,,, and i know what i talk ,and i guess u only dream ,same like we dream that one day nvidia will make async driver for dx11 pascal ..

 

[Samwell]

you 2 ,,, and i know what i talk ,and i guess u only dream ,same like we dream that one day nvidia will make async driver for dx11 pascal ..

No, it seems you have no idea what you're talking about. Kepler and Maxwell already have different SMs and Voltas SM is clearly different than Pascals. And i definately believe people like Ryan more, when they write that Volta is the biggest architecture jump since Fermi.

 

[TheF34RChannel]

Do you understand how an ALU works ? Of course Pascal and Volta offer same performance core to core (at same speed). I maybe shock you if I say that Fermi has also same performance core to core than Volta... and its normal ! However, is Fermi same than Volta and can reach same final performance ?
A good read for your education on Volta :
https://devblogs.nvidia.com/parallelforall/inside-volta/
Than tell me its same than Pascal

Can I ask who all your replies are to? I seem to be missing something or maybe have gone completely off the rails?

Never mind, got it!

 

[nvgpu]

https://developer.nvidia.com/cuda-toolkit/whatsnew

CUDA 9 is the fastest software platform for GPU-accelerated applications. It has been built for Volta GPUs and provides faster GPU-accelerated libraries, improvements to the programming model, compiler and developer tools. With CUDA 9 you can speed up your applications while making them more scalable and robust.

Download the CUDA 9 Release Candidate (RC) today to try the latest release.

 

[french toast]

Is clear Volta SM is very different from Pascal.
Does anyone have any information on how useful having independent FP32 and INT32 shaders are in real world apps? Is this purely compute benefit or could be used in gaming?
Also how does this compare to Vega?

 

[mohit9206]

No, it seems you have no idea what you're talking about. Kepler and Maxwell already have different SMs and Voltas SM is clearly different than Pascals. And i definately believe people like Ryan more, when they write that Volta is the biggest architecture jump since Fermi.

Yeah i don't believe that. Same thing was said about Maxwell and then Pascal.Biggest architecture jump since (insert the latest upcoming product here) .Its just hype. I'm not saying Volta is just a refresh but the media outlets job is to hype products up.

 

[Glo.]

Please stop to spread non sens and start to write better English, this forum is not a phone messenger app
Volta is even more advanced than Vega in DX12 specs. It has finer granularity in threads control (hardware based), better geometry performance and miles ahead in efficiency that its not even funny.


^^This
No more Pascal is planned. Nvidia is on full tilt with Volta from now on

Nvidia architectures differ in how they are fed by Register File size.

256 KB RFS was from Kepler architecture feeding different amounts of cores.

192 Cores/256 RFS for Kepler.
128 Cores/256 KB RFS for Maxwell, hence the throughput increase both in compute and in gaming.
128 cores/256 KB RFS for Pascal consumer GPUs(hence no increase in core throughput of the cores, compared to Maxwell).
64 cores/256 KB RFS for GP100 chip.
Volta, GV100 has the same 64 cores/256 KB RFS.

So it depends from where you are coming. Nvidia can reuse GP100, Pascal architecture and still offer per clock, per core increase in throughput compared to GP102, for example. But it does not matter what the name is for GV100 chip.

It is still Pascal architecture, with added few changes to architecture layout, and added Tensor cores, but it is still in the most meaningful way - the same Pascal architecture, but on refined 16 nm, called 12 nm.

Zuzu is correct on this. Per clock and per core you will not see a difference in FP32 applications between GP100 and GV100. But every GPU that will use GP100, or GV100 architecture will see improvement compared to GP102, per clock and per core. It is such funny situation, that Nvidia can use, for example 768 CUDA cores for GTX 2050 Ti, clocked at the same clock speeds, and still will get improvement in performance, compared to GTX 1050 Ti. Improvement to the degree that 1.6 GHz 768 GV107 Chip can be faster than GTX 1060 6 GB.


As to my question. Coffee Lake 35W CPU are going Live Q1 2018. Does anyone know when we can expect GV107 chip released by Nvidia? I would like to tie new small form factor build together with Intel release.

 

[Glo.]

Do you understand how an ALU works ? Of course Pascal and Volta offer same performance core to core (at same speed). I maybe shock you if I say that Fermi has also same performance core to core than Volta... and its normal ! However, is Fermi same than Volta and can reach same final performance ?
A good read for your education on Volta :
https://devblogs.nvidia.com/parallelforall/inside-volta/
Than tell me its same than Pascal

It does not offer the same performance core to core as Volta .

If that would be correct, we would not see improvement core to core in Maxwell vs Kepler, when they have changed the amount of cores fed by 256 KB Register File size, from 192 to 128. Those cores in Maxwell have become less starved for resources, and this little change increased the throughput of the cores.

That is why 2048 CUDA core Maxwell GPU was faster than 2880 CUDA core Kepler GPU, both in Gaming and Compute applications.

GV100 and GP100 have the same 64 core/256 KB Register File Size. New ISA can only help with several instructions, but always most important part that has biggest impact on performance will be core throughput.

 

[imported_bman]

From Nvidia's earning call:

Hans Mosesmann - Rosenblatt Securities, Inc.
As a follow-on, when can we expect Volta in the consumer gaming market? Thanks.

Jen-Hsun Huang - NVIDIA Corp.
Volta for gaming, we haven't announced anything. And all I can say is that our pipeline is filled with some exciting new toys for the gamers, and we have some really exciting new technology to offer them in the pipeline. But for the holiday season for the foreseeable future, I think Pascal is just unbeatable. It's just the best thing out there. And everybody who's looking forward to playing Call of Duty or Destiny 2, if they don't already have one, should run out and get themselves a Pascal.

Looks like a Q1 or Q2 2018 launch.

 

[SteveGrabowski]

From Nvidia's earning call:



Looks like a Q1 or Q2 2018 launch.

Crap, looks like I'll stick with my 970 for another year then.

 

[Puffnstuff]

Crap, looks like I'll stick with my 970 for another year then.

I feel ya as I have to stick with my 780 ti a while longer too.

 

[PeterScott]

I feel ya as I have to stick with my 780 ti a while longer too.

I wish. I have been waiting for GTX 1060 prices to drop to replace my 8800GT.

 

[TheF34RChannel]

Crap, looks like I'll stick with my 970 for another year then.

I feel ya as I have to stick with my 780 ti a while longer too.

I wish. I have been waiting for GTX 1060 prices to drop to replace my 8800GT.

Gosh now I feel like a spoiled brat for wishing to exchange my GTX 1080 for a GTX 2080 ASAP (alongside a new 6C/12T CPU) for games like Watch_Dogs 2 etc.! Different usages, different needs I suppose

All three of you will be in for a treat when you pop in whatever Pascal or Volta card; there will be teary eyes with disbelief and simultaneous joy Reminds me of my first proper GPU upgrade (FX5200 -> 8800GTS = shock and awe).

 

[SteveGrabowski]

All three of you will be in for a treat when you pop in whatever Pascal or Volta card; there will be teary eyes with disbelief and simultaneous joy Reminds me of my first proper GPU upgrade (FX5200 -> 8800GTS = shock and awe).

Nah my 970 is still pretty solid at 1080p anyways, so there is no way I'm going to pay $450 for the cheapest upgrade I can get. I was hoping to see GV104 this fall so that GV106 would come out next spring in time for Far Cry 5 (you know Ubisoft games are going to perform lousy on 2014 hardware).

I doubt 970 -> 1070 or 970 -> 2060 would even come close to my last gpu upgrade: GeForce 8400GS -> GTX 970.

http://gpu.userbenchmark.com/Compare/Nvidia-GeForce-8400-GS-vs-Nvidia-GTX-970/m7917vs2577

 

[tamz_msc]

Flagship(500$+) performance becomes mainstream(200-300$) in two generations. A GTX 780Ti is roughly around the GTX 1060 3GB now. Similarly it'll take the GV106 to get the performance of a 980Ti at 200$, assuming NVIDIA doesn't increase prices further at the corresponding segments. 1080Ti levels of performance won't be mainstream till after Volta.

 

[SteveGrabowski]

Flagship(500$+) performance becomes mainstream(200-300$) in two generations. A GTX 780Ti is roughly around the GTX 1060 3GB now. Similarly it'll take the GV106 to get the performance of a 980Ti at 200$, assuming NVIDIA doesn't increase prices further at the corresponding segments. 1080Ti levels of performance won't be mainstream till after Volta.

Pretty much what I'm looking for: 1070/980Ti performance new for $250 (I doubt the 2060 will be $200), though if there was a card with 1080Ti performance in the next gen for $400 (like GTX 970 vs GTX 780 Ti) I might jump too and buy a 4k TV to game on.

 

[eek2121]

Please stop to spread non sens and start to write better English, this forum is not a phone messenger app
Volta is even more advanced than Vega in DX12 specs. It has finer granularity in threads control (hardware based), better geometry performance and miles ahead in efficiency that its not even funny.


^^This
No more Pascal is planned. Nvidia is on full tilt with Volta from now on

I wouldn't bet on this. Nvidia won't be able to get Volta out until at least May 2018 due to GDDR6 availability. Even then, expect it to be a Titan series card. There were a few leaks a while back that suggested that Pascal will be optimized, tweaked, and rebadged as consumer Volta, while the Titan series and possibly *80 series will receive actual Volta treatment. What I do know is, Nvidia is already REALLY LATE with Volta, those of you that remember older roadmaps will remember that Nvidia had Volta pegged for last year. Pascal didn't exist.

 

[Qwertilot]

Nice definition of late here - they've already been shipping the stuff for months

They really don't need to wait for GDDR6 for all of Volta - the big cards might be fast enough to need it, GV104 down likely aren't - its just a matter of whether they want to or not. We'll find out.

 

[crisium]

They are not late. What was Volta was renamed to Pascal. What was unnamed is now Volta. They follow a 2 year cadence pretty precisely with a new arch in H1 of every even year. Even Maxwell, which had to be changed from 20nm to 28nm, still made that with Maxwell v1 750 Ti, although the v2 cards were a few months late.

Expect the first consumer Volta cards in the same Feb-May time frame next year.

 

[Qwertilot]

I wonder what 'son of Volta' is going to be called? Don't think they've done any post Volta road maps as yet.

Its a scary thought perhaps, but you can almost bet on them doing the V100 replacement on 7/10nm as soon as its even half practical. Sometime during next year I guess? Huge divergence for data center vs gaming, which makes sense of course

 

[TheF34RChannel]

Nvidia's next generation of Volta GPUs are expected to release in March 2018

Today some rumours have a emerged which states that Nvidia's gaming-oriented Volta series GPUs will release in March 2018, which makes a lot of sense given the current production schedule for GDDR6 memory from companies like SK Hynix. TSMC's new 12nm process, which Volta is based on, is also expected to be ready for large scale production by the end of the year, giving Nvidia another reason not to rush production, as higher yields and production rates are promised later.

https://www.overclock3d.net/news/gp..._gpus_are_expected_to_release_in_march_2018/1