Kitguru: US NOAA to use Pascal GPUs for Hurricane modeling

[jpiniero]

Why though? GP104 (or GP204?) is a mid-range chip and mid-range product in NV's Pascal stack. Why would we expect it to have 1TB/sec 8-16GB HBM2. I would be shocked if it does.

You are correct in that it likely wouldn't help that much performance wise. The reason it would have HBM2 would be more for the mobile parts. I think it's going to be $649 for the top bin GP104 and $449 for the cut.

 

[tviceman]

Why though? GP104 (or GP204?) is a mid-range chip and mid-range product in NV's Pascal stack. Why would we expect it to have 1TB/sec 8-16GB HBM2. I would be shocked if it does. Think about it -> 980 @ 224GB/sec but GP104 would have almost 5X the memory bandwidth? Why would it need that much bandwidth? It makes no sense. All you need is sufficient enough bandwidth not overkill bandwidth. Look at the R9 380X that has 182GB/sec vs. R9 280X that has 288GB/sec. What's that extra 100GB/sec bandwidth doing for the 280X vs. the 380X other than mostly marketing?

Exactly. I'm guessing that GP104 version 1.0 will come with a maximum 12ghz GDDR5X, equating to 384 gb/s on a 256-bit bus. That would be a 70% increase over GM204's current bandwidth, which would probably be more than enough to drive a 70+% performance increase over GTX980.

 

[jpiniero]

Exactly. I'm guessing that GP104 version 1.0 will come with a maximum 12ghz GDDR5X, equating to 384 gb/s on a 256-bit bus. That would be a 70% increase over GM204's current bandwidth, which would probably be more than enough to drive a 70+% performance increase over GTX980.

Have to figure 12 ghz GDD5X would draw way too much to be considered realistic in any part. And 70% faster than 980 for GP104 isn't anywhere near realistic unless it's like $800+ or nVidia drops Maxwell support quickly (both are possible)

 

[tviceman]

Have to figure 12 ghz GDD5X would draw way too much to be considered realistic in any part. And 70% faster than 980 for GP104 isn't anywhere near realistic unless it's like $800+ or nVidia drops Maxwell support quickly (both are possible)

GTX680 (GK104) was 70+% faster than GTX560 TI GF114 at release. GTX680 ended up being ~90% faster over the course of it's life.
GTX980 (GM204) was 70+% faster than GTX680 GK104 at release. GTX980 is currently 90% faster than GK104 and they're both on the same node.

And have you read the spec sheet on GDDR5X? Double the data rate, same power consumption. Furthermore, GTX680 beat GTX580 by ~30% and had the same MSRP as GTX580. GTX980 beat GTX780TI by ~10% and had a lower MSRP. Are you even trying to half-ass research anything you're saying? Why do you keep posting blatantly false information or expect something out of the ordinary come finfets?

 

[jpiniero]

Why do you keep posting blatantly false information or expect something out of the ordinary come finfets?

It's called multi-patterning.

ASML and even ARM have been saying there's no $/transistor benefit at 16FF for some time now. nVidia and AMD didn't even bother with 20 planar because that was much worse, and 16FF uses the same tooling. Plus the dGPU volume has been decimated. To make the money back they will have to be stingy on transistors or raise prices. In reality they will do a little bit of both.

 

[tviceman]

It's called multi-patterning.

ASML and even ARM have been saying there's no $/transistor benefit at 16FF for some time now. nVidia and AMD didn't even bother with 20 planar because that was much worse, and 16FF uses the same tooling. Plus the dGPU volume has been decimated. To make the money back they will have to be stingy on transistors or raise prices. In reality they will do a little bit of both.

Everyone and their mom always complains about the cost benefit when new nodes arrive. Nvidia has been complaining about transitioning to new nodes since the dawn of time. There is nothing particularly shocking or new about this. And hey, guess what!? The costs of wafers go down over time. Nvidia complained in corporate slides and rather abrasively about 28nm being too high in costs before their transition, queue a few years later and we have the biggest dies from both AMD and Nvidia that they've ever produced on 28nm. Weird how things work out.

As has already been said left and right, it's a foregone conclusion that costs are going up. The catch with die sizes and performance is that Nvidia cannot afford to let up on the high end and give Intel's HPC efforts momentum. Thus, since we can assume Nvidia will go full throttle with their big die, they cannot afford to leave a massive gap between their flagship and their mid-tier die. The dots... connect... something... I give up. Keep crying wolf.

 

[ShintaiDK]

Everyone and their mom always complains about the cost benefit when new nodes arrive. Nvidia has been complaining about transitioning to new nodes since the dawn of time. There is nothing particularly shocking or new about this. And hey, guess what!? The costs of wafers go down over time. Nvidia complained in corporate slides and rather abrasively about 28nm being too high in costs before their transition, queue a few years later and we have the biggest dies from both AMD and Nvidia that they've ever produced on 28nm. Weird how things work out.

As has already been said left and right, it's a foregone conclusion that costs are going up. The catch with die sizes and performance is that Nvidia cannot afford to let up on the high end and give Intel's HPC efforts momentum. Thus, since we can assume Nvidia will go full throttle with their big die, they cannot afford to leave a massive gap between their flagship and their mid-tier die. The dots... connect... something... I give up. Keep crying wolf.

Unlike previously, transistor cost didn't go down below 28nm.

Unless you want NVidia to run in the red like AMD. And for AMD digging an even deeper hole for itself. Then there is only 2 solutions. Dont increase the transistor budget, or to increase the price. This wont change before EUV arrives.

In other words, expect small dies or high prices. Or a combination.

 

[tviceman]

1. How old is that slide?
2. Of course I expect prices to go up. Of course I don't expect GP104 to be as big as GM204, but its called architecture improvements. GM204 has 40% more transistors than GK104, but is 1.6-1.9x faster in graphics and even more so in compute.

 

[ShintaiDK]

Mid 2014. But nothing changed since. Same reason why Samsung is launching 3 new SoCs, 2 being 28nm for 2016.

This is what Broadcom said mid 2015 for example:

Note the high design cost as well. Its another thing throwing companies off the curve.

 

[jpiniero]

Thus, since we can assume Nvidia will go full throttle with their big die, they cannot afford to leave a massive gap between their flagship and their mid-tier die. The dots... connect... something... I give up. Keep crying wolf.

nVidia is a very creative company... they will figure something out. But people's expectations need to be scaled back dramatically. I do think we will see Maxwell rebrands, it's just how high up the chain it goes.

 

[tviceman]

nVidia is a very creative company... they will figure something out. But people's expectations need to be scaled back dramatically. I do think we will see Maxwell rebrands, it's just how high up the chain it goes.

The only way we see a Maxwell "rebrand" at this point is if Nvidia decides to shrink various Maxwell parts down to finfets to increase both performance and perf/w, which I guess is totally plausible but I question how much money that will actually save in the long run.

So yes, I concede that point and idea to you. I can see Nvidia making just 1 Pascal part (GP100) for compute and ultra-high end desktop while shrinking GM200, GM204, and GM206 to fill out the rest of the lineup. Straight up rebrands though? No way. GM107 is getting encroached by high end IGP's so that definitely needs to be replaced at the same power envelope (GM206 shrink). But if they take the time to design the cores and layout for a low end graphics-focused card, then designing bigger chips would be a trivial cost at that point.

So if GM206 can be shrunk down to increase performance slightly (10%) while dropping the power envelope 25% then I think Maxwell shrinks & rebrands with only 1 Pascal GPU is a very real possibility.

 

[nvgpu]

https://forum.beyond3d.com/threads/nvidia-pascal-speculation-thread.55552/page-18#post-1880483

The CUDA DLL also calls equal to a few codenames:
GP100
GP102
GP104
GP106
GP107
GP10B

Pascal top to bottom lineup is coming, no rebrands.

 

[tviceman]

https://forum.beyond3d.com/threads/nvidia-pascal-speculation-thread.55552/page-18#post-1880483



Pascal top to bottom lineup is coming, no rebrands.

While I think it's more-than-likely we'll see top to bottom "true" Pascal chips, it is entirely plausible that Nvidia internally code-names Maxwell shrinks as Pascal chips. Also, there has never been a chip with a "2" or a "B" at the end, so those pascal code names are odd. It's always been 100/110, 104/204, 106/206, 107/207. Anyways, if Volta is following up in 2017 like Nvidia says, then I don't see anything personally wrong with Nvidia shrinking GM200 if they can get 10-15% more performance with the TDP of GK104/GM204.

 

[nvgpu]

GP10B is Tegra GPU, probably in the next-gen Tegra to be announced at CES 2016.
GM20B was the GPU in Tegra X1, GK20A in Tegra K1.

They are not shrinking Maxwell, shipping GPUs in 2016 without HEVC 10bit hardware decoder is just asking for market to swing to competition when Kaby Lake will have HEVC 10bit hardware decoder.

 

[jpiniero]

Pascal top to bottom lineup is coming, no rebrands.

Huh, interesting. I didn't think nVidia would do it, but there are people who would pay $1499 for a GPU that was 80% faster than Titan X. The GP102 is probally the equivalent of GM200 (gimped DP), something like 30% faster than Titan X for $999.

 

[Arachnotronic]

https://forum.beyond3d.com/threads/nvidia-pascal-speculation-thread.55552/page-18#post-1880483



Pascal top to bottom lineup is coming, no rebrands.

Excellent!

 

[ShintaiDK]

Now we just need the specs

 

[Brommer]

Mid 2014. But nothing changed since. Same reason why Samsung is launching 3 new SoCs, 2 being 28nm for 2016.

This is what Broadcom said mid 2015 for example:

Note the high design cost as well. Its another thing throwing companies off the curve.

So its pretty much a given fps/€ wont go up significantly from maxwell. Unless they can get some crazy architectual gaines

 

[ShintaiDK]

So its pretty much a given fps/€ wont go up significantly from maxwell. Unless they can get some crazy architectual gaines

That's not true.

Remember the cost of the GPU itself contra the rest. Even if you doubled the price of the GPU, the card retail price wouldn't double, far from.