Kitguru: US NOAA to use Pascal GPUs for Hurricane modeling

[Mondozei]

The news.

Since Pascal was designed from the ground up to have much more HPC capabilities than the stripped-down Maxwell, it does look like we'll see a repeat of Kepler here and a full-blown HPC-focused rollout for Big Pascal and a smaller (GP104) for consumers for the initial phase, not least because of the uarch/node improvement in the same generation. Again, last seen in Kepler.

Nothing in set in stone, of course, but the trickles just keep coming in that point to this.

 

[Mondozei]

Oh yeah, some more confirmation and details.
(Although this is mostly rehash of stuff we already know).

 

[railven]

If gtx 1080 is based of gp104 or whatever it's called and is <20% faster my 980ti after ocs I just stick with my card. I only upgrade when its >25% improvement OR I got a new monitor and old card don't cut it no more. I'd rather upgrade the girlfriends rig versus a side grade (or I just got gtx 1080 for me and give her my 980 ti. Ha-ha she said she's cool with it but id feel like a snake.)

 

[ocre]

So 5 to 10 times more accurate. Thats nothing to sneeze at

 

[Teizo]

"Pascal is expected to hit the market during the first half of 2016."

End of March, first of April?

 

[Silverforce11]

As expected, NV has to refresh its HPC stack first since Kepler still rules there.

Big Pascal with HBM2 will go where the market pay $5000+ for each until that market demand is met. Gamers will get a 104 mid-range SKU for high-end prices, as usual.

 

[cmdrdredd]

As expected, NV has to refresh its HPC stack first since Kepler still rules there.

Big Pascal with HBM2 will go where the market pay $5000+ for each until that market demand is met. Gamers will get a 104 mid-range SKU for high-end prices, as usual.

It's long been the case now that GPUs are priced according to relative performance. The fastest GPUs in a particular lineup are the most expensive regardless of the technology in the chip itself.

Anyway these things really need to be able to start pushing 4k at 60fps without turning settings down to medium. With UHD Blu-Ray on the horizon for next year a lot of people may be looking at going to 4k soon. Would be nice to have GPUs capable of pushing that resolution for games comfortably.

 

[Silverforce11]

There's recent games that SLI/CF top-GPU can't handle at 4K with MSAA. Once MSAA is not used, current stuff in dual-configs can handle it very well.

So next-gen if people want to max everything, they will probably need 2 top-GPU configs to do it. Assuming big Pascal or AI is ~80% above current big Maxwell or Fury.

 

[hawtdawg]

If gtx 1080 is based of gp104 or whatever it's called and is <20% faster my 980ti after ocs I just stick with my card. I only upgrade when its >25% improvement OR I got a new monitor and old card don't cut it no more. I'd rather upgrade the girlfriends rig versus a side grade (or I just got gtx 1080 for me and give her my 980 ti. Ha-ha she said she's cool with it but id feel like a snake.)

Its a full node shrink. Little Pascal will probably be bigger than Maxwell, and close to twice as fast.

 

[tviceman]

If gtx 1080 is based of gp104 or whatever it's called and is <20% faster my 980ti after ocs I just stick with my card. I only upgrade when its >25% improvement OR I got a new monitor and old card don't cut it no more. I'd rather upgrade the girlfriends rig versus a side grade (or I just got gtx 1080 for me and give her my 980 ti. Ha-ha she said she's cool with it but id feel like a snake.)

I think it's a foregone conclusion that Pascal's GP104 will be somewhere between 20-45% faster than a reference GM200. The big question will be the price.... $599 or $649? If it's 40% faster, I can see $649 for the top GP104 SKU, then $499 and $399 for two lesser SKU's.

 

[tviceman]

As expected, NV has to refresh its HPC stack first since Kepler still rules there.

Big Pascal with HBM2 will go where the market pay $5000+ for each until that market demand is met. Gamers will get a 104 mid-range SKU for high-end prices, as usual.

GK104 had an 11 month lead over GK110 in the consumer market. For 11 months it was Nvidia's fastest GPU for the desktop. Why in the world would Nvidia price it differently and squeeze the rest of their product stack pricing when AMD presented no reason to do so? GK104 was priced according to it's performance relative to the competition. If AMD had their act together and had good release drivers for the HD7970, or priced it more competitively, Nvidia would have responded appropriately at GTX 680's release. If AMD strikes first next generation with competitive pricing AND competitive products, Nvidia can't come to the market later with slower parts at more expensive prices as they would get ripped apart in reviews.

Nvidia's been making a good business out of their "mid-range code-named" chips beating AMD's best chips at release. Why would the sell for less when they don't have to?

 

[Silverforce11]

I think it's a foregone conclusion that Pascal's GP104 will be somewhere between 25-45% faster than a reference GM200. The big question will be the price.... $599 or $649? If it's 40% faster, I can see $649 for the top GP104 SKU, then $499 and $399 for two lesser SKU's.

How faster it will be depends on how optimistic you are.

Assume a few IC-based metrics.

Let's say GP104 will have the same transistor count as GM200 at half the die size. This is assuming perfect transistor density doubling with the node jump.

Thus, at the same transistor count, assuming perfect node performance/scaling, the performance would be the same at half the power usage.

This is where NV can be flexible in how they target the power profile to bring up or down on the target performance.

Because node jumps are rarely perfect, the gains will be lower. This is countered by any uarch performance leaps.

Overall, the ballpark is quite wide. Because we don't know the chip-metrics of GP104, its actually crazy to expect a solid performance target at this point. GP104 could land at a very wide performance target range.

If and when data is available, ie. "GP104 is 300mm2 with the ~transistor count of GM200", then we can speculate with more accuracy.

@Pricing, obviously the price will be whatever the market allows it. So if there's no competition, gamers will be owned.

 

[tviceman]

I'll stick by my prediction, which I already thought was a fairly wide performance target of 20-45% faster than GM200 and a price ceiling of $649. It's near impossible to make performance predictions based on transistor counts of different architectures. GM204 was 10% faster at release than GK110 with 27% less transistors while GK104 was 30% faster at release than GF110 with 16% more transistors. Likewise, comparing to like-predecessors also yields highly variable information. GM204 had 48% more transistors than GK104 and was much more efficient per transistor, being ~75% faster at release. GK104 had 80% more transitors than GF114 but was only 70% faster at release, making it less efficient per transistor.

 

[Pottuvoi]

If they can use FP16 pascal should be quite fast.

 

[Timmah!]

If they can use FP16 pascal should be quite fast.

I need to know so much whether the app of my choice is going to benefit from this.

Interesting thing Nvidia presented Pascal and still no word about the upcoming dual GPU Maxwell. The first concrete news were almost 2 months ago, since then nothing. Wonder what are they waiting for?

 

[Magic Carpet]

Wonder what are they waiting for?

Christmas

IMO, you better wait for Pascal. Now in the meantime, assemble another system with 2-4 780s for your GPU computing and make some dough in-between.

 

[RussianSensation]

Its a full node shrink. Little Pascal will probably be bigger than Maxwell, and close to twice as fast.

What? No, not going to happen. With a full node shrink + new architecture, NV managed 2X from Big Fermi to Big Kepler.

I am going to assume when you said Little Pascal you meant mid-range GP104, not GTX950 successor. Let's look at it another way.

@ 4K to show as far GPU limited as possible.

An after-market 980Ti is 46% faster than a reference 980. Why am I using an after-market 980Ti, because almost all $650 980Tis have big 15-17% overclocks out of the box like the MSI Gaming 980Ti.

For your statement to be true, an after-market GP104 chip succeeding 980 would need to be:

2x (158 / 108) => 2.93X faster than a reference 980 to be 2X faster than a typical 980Ti. This is literally impossible.

If 980's successor is 2X faster than a 980, it would still be 'only' 58-60% faster than a reference 980Ti, nowhere close to 2X. Mind you, that would still be incredible and unheard of from a next gen mid-range chip (680 was 'just' 35-40% faster than a 580) but your expectations are just completely unrealistic.

I'll stick by my prediction, which I already thought was a fairly wide performance target of 20-45% faster than GM200 and a price ceiling of $649.

Seconded for GP104, but I'll add 20-45% faster specifically at 1440P and 4K; and our basis point is a reference 980Ti, not an after-market 980Ti. At 1080P, it's going to be much less due to major CPU limitations in many console ported PC games. I am also expecting NV to keep raising prices. 20%-45% faster than GM200 at $649 is going to sound like great value against the 980Ti, regardless if they will first release a mid-range GP104 in 2016 and GP100/110 in 2017.

 

[jpiniero]

With Intel launching Knights Landing, the first product out almost has to be Tesla... so it would make sense there would be a Titan as the first product.

Remember:
- No $/transistor benefit from 28 -> 16FF
- Going from gimped DP on GM200 to full DP with GP100
- HBM2 has got to be pricey if they are launching with it. Boy will people be pissed if GP100 is HBM2 but GP104 is GDDR5X
- Yield on something big is going to be crap and too expensive anyway, the GP100 Tesla product is probably multiple dies

So don't get your hopes up.

Interesting thing Nvidia presented Pascal and still no word about the upcoming dual GPU Maxwell. The first concrete news were almost 2 months ago, since then nothing. Wonder what are they waiting for?

AMD to release the Fury X2?

 

[RussianSensation]

Boy will people be pissed if GP100 is HBM2 but GP104 is GDDR5X

Why? If Pascal delivers 2X the perf/watt over Maxwell, an after-market 980 successor could be up to 2X faster than a 980, which would deliver a massive increase in performance over even the 980Ti.

980 = 7Ghz GDDR5 @ 256 bit = 224GB/sec
Pascal GP104 = 14Ghz GDDR5X @ 256 bit = 448GB/sec

Also, what if they use 4th generation memory compression? If they do, they might not even need to double the memory bandwidth on GP104 to get 80-100% more performance over the 980.

If GP104 is 80-100% faster than the 980 and has 8GB of GDDR5X, even at $649 gamers will buy it in droves.

- Yield on something big is going to be crap and too expensive anyway, the GP100 Tesla product is probably multiple dies

So don't get your hopes up.

You can make the same argument every generation going back 20 years. There are plenty of ways for NV to sell GP100 in 2016 outside of the professional market. The first is simply to raise prices on the Titan series to $1299-1499, etc. The second is to wait until Q3-4 2016 when yields improve. The third is to release a partial yielding Titan chip, similar to what they did with the OG Titan. Then they could follow up with the Titan Black fully unlocked model in 2017.

I have no way of knowing what they are going to do as we can only speculate at this point but there are lots of possibilities how it's possible that we may still see GP100 in the consumer space. Is it going to be cheap? Of course not but modern GPUs have dramatically increased in price and gamers still keep buying. GTX980Ti/Fury X are still $600+ for the most part and we should have $650 GPUs 30% faster than that in 2016, maybe even faster than that.

I don't know why there is so much fear mongering on Pascal that started after 980Ti launched. It's as if some gamers don't want Pascal to wipe the floor with Maxwell and Fury X. Oh but it will, in all key metrics, price/performance, perf/watt, VRAM amount, features, etc. It might not hit price/performance out of the park right off the bat but a generation lasts 2 years.

 

[tviceman]

- HBM2 has got to be pricey if they are launching with it. Boy will people be pissed if GP100 is HBM2 but GP104 is GDDR5X

The only people who will be disappointed are the people who buy graphics cards for written specs instead of proven performance. If someone is buying a graphics card based on die size or memory type instead of real-world in game frames per second, then they are truly ignorant.

 

[jpiniero]

Why?

I think people would be pissed that it doesn't have it esp since full GP104 is going to be so expensive. Even if the bandwidth benefits aren't realized the lower power draw will be.

 

[RussianSensation]

I think people would be pissed that it doesn't have it esp since full GP104 is going to be so expensive. Even if the bandwidth benefits aren't realized the lower power draw will be.

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?

What matters is effective bandwidth and how much you need before it's overkill. I mean you should know this considering Fury X has 512GB/sec vs. 336GB/sec for the 980Ti.

I think gamers will be more pissed if GP104 costs even more than $550. I know I will be (even though I fully expect it to be more expensive as NV keeps raising prices). I expect HBM2 to be used on the GP100/110 level products. If GP104 is 20-45% faster than GM200 and it has 8GB of GDDR5X, I am perfectly fine with it. Would I have preferred the more efficient HBM2? Sure, but not if it raises the prices even higher and the extra bandwidth is worthless. In that case, let them use GDDR5X on mid-range and HBM2 on flagship cards.

 

[nvgpu]

http://www.gputechconf.com/

APRIL 4-7, 2016 SILICON VALLEY

Can't wait for April 4, when Nvidia will finally demo Pascal to customers at GTC 2016.

 

[Arachnotronic]

http://www.gputechconf.com/

APRIL 4-7, 2016 SILICON VALLEY

Can't wait for April 4, when Nvidia will finally demo Pascal to customers at GTC 2016.

Nice!

 

[ShintaiDK]

I think people would be pissed that it doesn't have it esp since full GP104 is going to be so expensive. Even if the bandwidth benefits aren't realized the lower power draw will be.

No. The only thing that matters is how it performs. Powerdraw could also be a lot less if the chip was bigger and downclocked. Demanding HBM2 is just technological purist and silly.

The only people who will be disappointed are the people who buy graphics cards for written specs instead of proven performance. If someone is buying a graphics card based on die size or memory type instead of real-world in game frames per second, then they are truly ignorant.

Exactly!