NVIDIA Volta Rumor Thread

[Ajay]

I dont think so. It should be possible to get these numbers into 600mm², 400mm² like with Maxwell. Bouowmx already showes the scaling which are worst case numbers, but a gaming chip will be smaller. GP102 will only scale shaders up and features. V100 had Tensor Cores, 50% more nvlinks, 2mb additional l2 cache and 45% more l1/Shared Memory per SM compared to GP100. And nevertheless they increased shader count by 40% with just 33% more die size.

GV102 won't have tensor cores (maybe a few for compatibility), won't have interfaces like nvlink which need to scale, probably won't have a increased l2 cache and scaling of L1 Cache will probably be between 15-33%. Additionally we have the 12FFN process which is unknown. Just because GV100 doesn't show scaling means nothing. Building such a big chip with a bit lower density might improve yield substantially. GV102 might have a 5% higher transistor density and then you're easy under 600mm².

I don't think Nvidia wants consumer dice as large as Maxwell. AMD isn't in a good place vis-a-vis competitiveness ATM, so I don't see NV going with large dies. Again, I could be wrong, but uarch gains and perf/watt on consumer Volta GPUs will be enough - maybe will a small bump in die size. I'll be happy if I'm wrong, but NV is highly valued by Wall Street now, so profits come first.

 

[Samwell]

That had negligible effect. They increased the transistor count by 38%, and the cuda cores by 40%, which means the impact of the other systems on transistor budget was negligible. So it will take about 33% more die size to increase core counts by about 40%.

If you think NVidia is going to increase die size by 33% and not increase price by similar amount(or more), you aren't aware of the competitive landscape, or the state of Moore's law.

We are hitting a process wall, and I expect more conservative jumps in core counts going forward. 20% then 20% giving Nvidia two generations of moderate gains rather than one big one.

No the impact of the other stuff wasn't small, shaders are not everything in a chip. Shaders take up 50-60% of the chips diespace. If you just scale up shaders and leave the other stuff the same, then the size won't grow so much. A lot of size increase is coming from other stuff.

You can just have a look at Die shots . https://www.flickr.com/photos/130561288@N04/36230799276/in/album-72157650403404920/
1 GPC is around 40 mm² in Gp104, pretty sure the same in GP102. So the 1536 Shader increase needed to get to 5376 SP would take around 100mm², so a 570 mm² Chip. Now give it a slight density increase of 5% vs GP102 (V100 density increased 3% vs GP100) and you're at 540mm². 60mm² for new features which Volta added should be enough and you have your gaming chip with 600 mm².

Two generations of moderate gains only make sense if you do that by using the same gpu as cut and noncut generation. Design cost are two high nowadays, so that it doesn't make sense to design 2 chips with small difference.

I don't think Nvidia wants consumer dice as large as Maxwell. AMD isn't in a good place vis-a-vis competitiveness ATM, so I don't see NV going with large dies. Again, I could be wrong, but uarch gains and perf/watt on consumer Volta GPUs will be enough - maybe will a small bump in die size. I'll be happy if I'm wrong, but NV is highly valued by Wall Street now, so profits come first.

Nvidia doesn't know how good Vega will be at the time Volta was designed. AMD was always talking of a big architecture jump and probably they communicated the same to partners. If you're the competition with the same partners then you get some infos of them and also have the information that amd is planning a new major architecture. Do you really think that if nvidia knows that amd is planning a big architecture jump they will sit down and make a smaller jump with the possibility to lose? Look at V100, that's not their style. They'll go all in and design a chip with which they're sure to beat amd. As for price, they can launch it at 799$ für GP102 and 599$ for GV104. They'll earn enough. Finfet production cost are also going down slowly and yield increases over time. I don't think a 400 mm² GV104 in 2018 would cost much more to produce than a 315mm² GP104 in 2016.

 

[Head1985]

I dont expect much from small volta(gv104).GTX2080 will be 10-20% faster than reference1080TI(once both oc gap will be smaller).
GTX2070 10% slower than reference 1080TI(and aftermarket cards match it) but once both oc 1080TI will be faster.

Good card will be again Volta TI-barely cutdown vs full GV102 at decent price and 60+% faster than 1080TI.

 

[amenx]

I dont expect much from small volta(gv104).GTX2080 will be 10-20% faster than reference1080TI(once both oc gap will be smaller).
GTX2070 10% slower than reference 1080TI(and aftermarket cards match it) but once both oc 1080TI will be faster.

Why do you say that? For last 2 gens Nvidia have made the x70 cards = to the Ti in performance (970 = 780ti, 1070 = 980ti). Seems like a strong marketing forumla which greatly boosts the x70 appeal and success. Personally think Nvidia will continue with same, so hopefully the 2070 = 1080ti.

 

[Qwertilot]

That's definitely what I meant by 'normal' Then the xx80 a bit above and the xx80ti quite a bit above of course.

 

[Head1985]

Why do you say that? For last 2 gens Nvidia have made the x70 cards = to the Ti in performance (970 = 780ti, 1070 = 980ti). Seems like a strong marketing forumla which greatly boosts the x70 appeal and success. Personally think Nvidia will continue with same, so hopefully the 2070 = 1080ti.

Because GTX1070 is so much cutdown and slower than 1080TI, so for volta NV will need cutdown GTX2070 very little vs GTX2080(15% max vs 33% now with gtx1070) to be able match GTX1080TI whitch is around 60-70% faster than GTX1070.
Pascal GTX1070 brings only 50-55% performance gain vs GTX970 and that was on new node.So for volta 2070 bring 60-70% gain on same or only slighly better 12nm node they will need cut GTX2070 very little.And i dont believe it.They wil cut it probably again with one GPC out.
https://forums.anandtech.com/threads/gtx1070-2114-9500-vs-gtx1080ti.2517242/

 

[Rifter]

Why do you say that? For last 2 gens Nvidia have made the x70 cards = to the Ti in performance (970 = 780ti, 1070 = 980ti). Seems like a strong marketing forumla which greatly boosts the x70 appeal and success. Personally think Nvidia will continue with same, so hopefully the 2070 = 1080ti.

Im hoping for this as well, i want a good card that will max 4k and also be good for VR, so will go 2070 if it gets me there, or 2080 if i have to. Im hoping the 2070 lives up to the last few generations of performance gain. Vega just didnt deliver, hopefully Volta doesnt let me down too.

 

[Dayman1225]

The just released PGI v17.7 (compiler and tools) now supports Volta
https://insidehpc.com/2017/09/new-pgi-17-7-release-supports-nvidia-volta-gpus/

the PGI page:
http://www.pgroup.com/products/new-in-pgi.htm

 

[sonicology]

Why do you say that? For last 2 gens Nvidia have made the x70 cards = to the Ti in performance (970 = 780ti, 1070 = 980ti). Seems like a strong marketing forumla which greatly boosts the x70 appeal and success. Personally think Nvidia will continue with same, so hopefully the 2070 = 1080ti.

I may be wrong but I'm expecting a more modest gain from this generation than the last. Nvidia don't have the advantage of a big process shift (12nm will offer very limited improvements) although GDDR6 should help them out considerably.

I'd still expect the x80 part to offer a modest improvement over a 1080ti so Nvidia can cash in on it being the new benchmark champion, but I expect it will be a much smaller step up than the 1080 was over the 980ti.

 

[happy medium]

I may be wrong but I'm expecting a more modest gain from this generation than the last. Nvidia don't have the advantage of a big process shift (12nm will offer very limited improvements) although GDDR6 should help them out considerably.

I'd still expect the x80 part to offer a modest improvement over a 1080ti so Nvidia can cash in on it being the new benchmark champion, but I expect it will be a much smaller step up than the 1080 was over the 980ti.

So you expect the gtx2080 to be only 27% faster than a gtx1080ti ?

Even the gtx1070 was 12% faster than a gtx980ti when release.

 

[happy medium]

Because GTX1070 is so much cutdown and slower than 1080TI, so for volta NV will need cutdown GTX2070 very little vs GTX2080(15% max vs 33% now with gtx1070) to be able match GTX1080TI whitch is around 60-70% faster than GTX1070.
Pascal GTX1070 brings only 50-55% performance gain vs GTX970 and that was on new node.So for volta 2070 bring 60-70% gain on same or only slighly better 12nm node they will need cut GTX2070 very little.And i dont believe it.They wil cut it probably again with one GPC out.
https://forums.anandtech.com/threads/gtx1070-2114-9500-vs-gtx1080ti.2517242/

How much faster is a gtx1080ti vs a gtx1070?

 

[Head1985]

How much faster is a gtx1080ti vs a gtx1070?

Well i tested myself more than 20games.It is 62% average faster vs golden sample GTX1070 at 2114-2140/9500.Probably 70% faster vs GTX1070 at 2000/9000.Volta 2070 will need have bigger performance gain than pascal 1070 because pascal 1070 was only 55% avg faster than GTX970 and 1080TI is 60-70% faster than GTX1070.If they didnt manage even 60% gain with pascal 1070 on new 16nm, then i dont see how they manage 70% gain on almost same process(12nm is just slightly better 16nm).Only way is cutdown GTX2070 much less than 1070 so it will be very close to GTX2080.But i think they will cut it like 1070 so one GPC/33% SP down

 

[happy medium]

they didnt manage even 60% gain with pascal 1070 on new 16nm, then i dont see how they manage 70% gain on almost same process(12nm is just slightly better 16nm).[/QUOTE

.

Maxwell + clock speed and more shaders = Pascal
Volta is a new architecture with 12nm power saving improvements.

 

[Ajay]

I don't think a 400 mm² GV104 in 2018 would cost much more to produce than a 315mm² GP104 in 2016.

The more I think about it, the more I think this is a very good point. If true, GV104 will be an excellent upgrade (unless prices go way up, which NV has been hinting at). Hopefully, the mining performance will suck.

 

[Dayman1225]

Isn't the 12nm used (refined 16nm afaik) made to be cheaper to produce than the previous 16nm process? If true that point does make sense

 

[R0H1T]

I'd say it;s pretty much a given that Volta>Pascal will be less than Pascal>Maxwell, least of all because were not gonna go from 16nm to 7nm as was the case from Maxwell to Pascal i.e. 28nm to 16nm. There's also talk that consumer Volta might be more compute heavy, therefore expecting huge gains over current gaming flagships is being overly optimistic.

 

[Ajay]

Isn't the 12nm used (refined 16nm afaik) made to be cheaper to produce than the previous 16nm process? If true that point does make sense

Memory prices are still going up and will continue to do so for a while. Specialized GDDR 6 will be especially expensive.

 

[Arachnotronic]

I'd say it;s pretty much a given that Volta>Pascal will be less than Pascal>Maxwell, least of all because were not gonna go from 16nm to 7nm as was the case from Maxwell to Pascal i.e. 28nm to 16nm. There's also talk that consumer Volta might be more compute heavy, therefore expecting huge gains over current gaming flagships is being overly optimistic.

Kepler to Maxwell was a huge jump, FYI.

 

[Dayman1225]

http://news.mydrivers.com/1/548/548671.htm

Volta V100 DGX System spotted on geekbench

Edit:
Direct Links
https://browser.geekbench.com/v4/compute/1098943
https://browser.geekbench.com/v4/compute/1098991

 

[Glo.]

Does it check single GPU, or all of them(in DGX-1 there is 8 Volta GPUs).

 

[Glo.]

I'd say it;s pretty much a given that Volta>Pascal will be less than Pascal>Maxwell, least of all because were not gonna go from 16nm to 7nm as was the case from Maxwell to Pascal i.e. 28nm to 16nm. There's also talk that consumer Volta might be more compute heavy, therefore expecting huge gains over current gaming flagships is being overly optimistic.

Change from 192Cores/256 KB Register File Size to 128 Cores/256 KB RFS brought 33% IPC increase, and was reflected in both: gaming and compute benchmarks.

IPC increase from going from 128cores/256 KB RFS, to 64 Cores/256 KB RFS is 50%. Volta may be biggest jump in IPC we have seen in very long time.

In essence: 1024 CUDA core GPU using Volta arch. clocked at 1.5 GHz will have the same level of performance as 1536 CUDA core, 1.5 GHz Pascal arch. GPU.

Also bare in mind that with each generation Nvidia slightly increases core counts in their GPUs.

Kepler to Maxwell GX104 - 1536 CC -> 2048 CC.
Even if Pascal used the same arch as Maxwell in consumer space, we still got massive increase in performance thanks to increased core clocks and core counts:
Maxwell - Pascal GX104 - 2048 -> 2560 CC's.
The same pattern was everywhere.

Thats why I expect that GV104 resembling Volta architecture will have 3072 CUDA cores, with 50% higher IPC, resulting in around 65-70% higher performance than GP104 had.
GV107 - 1024 CC's, with 50% higher IPC, resulting in around 15% higher performance than GTX 1060.

Nvidia this way will be able to say to everybody: Our competition democratized VR. We are democratizing 1440p, and 4K gaming with GTX 2050 Ti, and GTX 2060.

 

[Malogeek]

Does it check single GPU, or all of them(in DGX-1 there is 8 Volta GPUs).

Apparently it will only test single GPUs.

 

[LTC8K6]

WCCF link:

http://wccftech.com/nvidia-volta-tesla-v100-gpu-compute-benchmarks-revealed/

NVIDIA Volta Tesla V100 GPU Accelerator Compute Performance Revealed – Features A Monumental Increase Over Pascal Based Tesla P100

 

[R0H1T]

Change from 192Cores/256 KB Register File Size to 128 Cores/256 KB RFS brought 33% IPC increase, and was reflected in both: gaming and compute benchmarks.

IPC increase from going from 128cores/256 KB RFS, to 64 Cores/256 KB RFS is 50%. Volta may be biggest jump in IPC we have seen in very long time.

In essence: 1024 CUDA core GPU using Volta arch. clocked at 1.5 GHz will have the same level of performance as 1536 CUDA core, 1.5 GHz Pascal arch. GPU.

Also bare in mind that with each generation Nvidia slightly increases core counts in their GPUs.

Kepler to Maxwell GX104 - 1536 CC -> 2048 CC.
Even if Pascal used the same arch as Maxwell in consumer space, we still got massive increase in performance thanks to increased core clocks and core counts:
Maxwell - Pascal GX104 - 2048 -> 2560 CC's.
The same pattern was everywhere.

Thats why I expect that GV104 resembling Volta architecture will have 3072 CUDA cores, with 50% higher IPC, resulting in around 65-70% higher performance than GP104 had.
GV107 - 1024 CC's, with 50% higher IPC, resulting in around 15% higher performance than GTX 1060.

Nvidia this way will be able to say to everybody: Our competition democratized VR. We are democratizing 1440p, and 4K gaming with GTX 2050 Ti, and GTX 2060.

I don't expect the "IPC" to go up majorly & the clocks to remain as high as say Pascal, it could be one or the other but probably a "bit" of both. The compute performance though, for HPC, has gone through the roof & as we all know that doesn't tell us much about gaming, per se.

In the end, I'd say Nvidia could simply cram in more cores to make the upcoming (Volta) gaming chips outperform the 1080Ti, there's also GDDR6 &/or high speed GDDR5x to look forward to. Like I said, I don't expect huge gains from Pascal to Volta, for games, though better compute & more cores could skew this in favor of the high end estimated that many are speculating in this thread.

In short there are many ways to skin cat, but I still don't expect Volta to be anything other than evolutionary. We'll find out how Nvidia achieves their next bit of magic, in due course of time.

 

[Head1985]

Change from 192Cores/256 KB Register File Size to 128 Cores/256 KB RFS brought 33% IPC increase, and was reflected in both: gaming and compute benchmarks.

IPC increase from going from 128cores/256 KB RFS, to 64 Cores/256 KB RFS is 50%. Volta may be biggest jump in IPC we have seen in very long time.

In essence: 1024 CUDA core GPU using Volta arch. clocked at 1.5 GHz will have the same level of performance as 1536 CUDA core, 1.5 GHz Pascal arch. GPU.

Also bare in mind that with each generation Nvidia slightly increases core counts in their GPUs.

Kepler to Maxwell GX104 - 1536 CC -> 2048 CC.
Even if Pascal used the same arch as Maxwell in consumer space, we still got massive increase in performance thanks to increased core clocks and core counts:
Maxwell - Pascal GX104 - 2048 -> 2560 CC's.
The same pattern was everywhere.

Thats why I expect that GV104 resembling Volta architecture will have 3072 CUDA cores, with 50% higher IPC, resulting in around 65-70% higher performance than GP104 had.
GV107 - 1024 CC's, with 50% higher IPC, resulting in around 15% higher performance than GTX 1060.

Nvidia this way will be able to say to everybody: Our competition democratized VR. We are democratizing 1440p, and 4K gaming with GTX 2050 Ti, and GTX 2060.

stop with this register file nonsense.Maxwell doubled Rops, 4x increase cache, brings 40% clock increase, new delta color compression, TB rasterization.This is why it was so much better than kepler.