NVIDIA Pascal Thread

[ShintaiDK]

GP100 is not going to be more than 500mm2 as peak, if it supports 4 stacks of HBM2. Because HBM2 is quite big and needs to e on an interposer.

This may cost Nvidia good chunks of the HPC segment to Knights Landing.

 

[jpiniero]

So with all the talk how HBM2 parts wont be ready before the end of year, you expect exactly the one product to be launched, which is the most likely to feature them?

HBM2 is available now. Not in volume by any means, but enough for their HPC contracts.

This may cost Nvidia good chunks of the HPC segment to Knights Landing.

500 mm2 should be enough to get it well above the 3 TF DP of KNL. The rumor is 4 TF DP but we shall see.

 

[caswow]

HBM2 is available now. Not in volume by any means, but enough for their HPC contracts.



500 mm2 should be enough to get it well above the 3 TF DP of KNL. The rumor is 4 TF DP but we shall see.

you have to be able to package into a finished product. who knows if nvidia and their partners can do it without problems. amd already has a product and a functioning supply chain.

 

[Cloudfire777]

AMD will launch 2 Polaris GPUs and 2 Vega GPUs and compete against the GP107/GP106/GP104/GP100 lineup.

GP107- Polaris 11
GP106- Polaris 10
GP104- Vega 11
GP100 - Vega 10

If AMD is in the business of going bankrupt, well thats a good plan for that....

You are suggesting that AMD will launch the competitor to GP104, which launch in June, in 2017 (according to AMDs own roadmap).
Are you not aware how much potential they would lose by letting Nvidia have almost all the cake for 6 months + ?

No, what will happen is that GP104 will be a small upgrade from GM200, so that GTX 1080 will beat GTX 980Ti by 20% ish.
AMD will follow with Polaris 10 which will also be 20% ish faster than GTX 980Ti.

You dont have to go longer back than look at 7870 and 6970 to see similarities.

 

[Glo.]

Cloudfire, you do realize that Polaris 10 is Pitcairn replacement? R7 470X for example.

 

[xpea]

GP100 is not going to be more than 500mm2 as peak, if it supports 4 stacks of HBM2. Because HBM2 is quite big and needs to e on an interposer.

This may cost Nvidia good chunks of the HPC segment to Knights Landing.

No you are wrong. GP100 can be close to 500mm2 even with HBM2.
TSMC already produces much bigger interposer (775mm2) for XILINX than the toy UMC is doing for AMD... and since 2011...
see here (interposer size on last page):
http://www.xilinx.com/programmable/about/research-labs/3-D_Architectures.pdf

 

[Cloudfire777]

Cloudfire, you do realize that Polaris 10 is Pitcairn replacement? R7 470X for example.

I think Polaris 10 is replacing R9 390X all the way down to R9 380.
By replacing, I mean the GPUs AMD will launch will beat the 300 cards.

Remember that 7870 beat 6970 by 10%. I expect similar with Polaris 10 and Fury X as well. Which fits since AMD demo`d Polaris 10 where it had more FPS in Hitman than Fury X.
Doubt Nvidia`s GP104 to be that much more powerful. I think GTX 1080 will beat Polaris 10 by 10-20% but AMD will make up for it with cheaper Polaris 10 GPUs since the chip is only 232mm2

 

[ShintaiDK]

No you are wrong.
TSMC already produces much bigger interposer (775mm2) for XILINX than the toy UMC is doing for AMD... and since 2013...
see here (interposer size on last page):
http://www.xilinx.com/programmable/about/research-labs/3-D_Architectures.pdf

The interposer for AMD(Fiji) is much bigger than 775mm2. 1011mm2 to be precise.
http://www.anandtech.com/show/9390/the-amd-radeon-r9-fury-x-review/3

Finally, as large as the Fiji GPU is, the silicon interposer it sits on is even larger. The interposer measures 1011mm2, nearly twice the size of Fiji. Since Fiji and its HBM stacks need to fit on top of it, the interposer must be very large to do its job, and in the process it pushes its own limits. The actual interposer die is believed to exceed the reticle limit of the 65nm process AMD is using to have it built, and as a result the interposer is carefully constructed so that only the areas that need connectivity receive metal layers. This allows AMD to put down such a large interposer without actually needing a fab capable of reaching such a large reticle limit.

 

[Timmah!]

HBM2 is available now. Not in volume by any means, but enough for their HPC contracts.



500 mm2 should be enough to get it well above the 3 TF DP of KNL. The rumor is 4 TF DP but we shall see.

Now honestly i doubt this is more than your assumption. I dont know what number of GPU they are supposed to supply as part of their HPC contracts, but i would hazard a guess it gets into thousands, if not tens of thousands, in which case thats already a "volume".

 

[jpiniero]

Now honestly i doubt this is more than your assumption. I dont know what number of GPU they are supposed to supply as part of their HPC contracts, but i would hazard a guess it gets into thousands, if not tens of thousands, in which case thats already a "volume".

Yeah, tens of thousands. As opposed to consumer GPUs which would need a million+.

 

[Timmah!]

Yeah, tens of thousands. As opposed to consumer GPUs which would need a million+.

Yeah, and even for those tens of thousands you need a facility to be going, its not like Jen Hsun Huang can produce them on his knee in his garage, cause its not million+.

Bottom line, this is from January:

http://www.extremetech.com/extreme/...ass-production-of-next-generation-hbm2-memory

and you think they are going to launch GP100 on Tuesday? Fine, but then the rumour about HBM2 not before the end of the year is clearly false.

 

[Glo.]

I think Polaris 10 is replacing R9 390X all the way down to R9 380.
By replacing, I mean the GPUs AMD will launch will beat the 300 cards.

Remember that 7870 beat 6970 by 10%. I expect similar with Polaris 10 and Fury X as well. Which fits since AMD demo`d Polaris 10 where it had more FPS in Hitman than Fury X.
Doubt Nvidia`s GP104 to be that much more powerful. I think GTX 1080 will beat Polaris 10 by 10-20% but AMD will make up for it with cheaper Polaris 10 GPUs since the chip is only 232mm2

In performance, yes. But not in market segment. Secondly, you compare two different transition in nodes. It may not be the case.

 

[raghu78]

If AMD is in the business of going bankrupt, well thats a good plan for that....

You are suggesting that AMD will launch the competitor to GP104, which launch in June, in 2017 (according to AMDs own roadmap).
Are you not aware how much potential they would lose by letting Nvidia have almost all the cake for 6 months + ?

No, what will happen is that GP104 will be a small upgrade from GM200, so that GTX 1080 will beat GTX 980Ti by 20% ish.
AMD will follow with Polaris 10 which will also be 20% ish faster than GTX 980Ti.

You dont have to go longer back than look at 7870 and 6970 to see similarities.

Nobody knows when Nvidia will launch GP104. Right now there is just rumours. We have to wait and see if Nvidia has any news on Pascal GPUs at GTC 2016.

 

[Silverforce11]

You are suggesting that AMD will launch the competitor to GP104, which launch in June, in 2017 (according to AMDs own roadmap).

Are you not aware how much potential they would lose by letting Nvidia have almost all the cake for 6 months + ?

1. We still don't know when GP104 SKUs are releasing in retail.

2. The "cake" for GPU market is ~$300 and under. Even you should be aware of that. High priced GPUs are the icing on the cake.

3. If you go by the Roadmap, Vega has most of the bar falling in late 2016.

4. Polaris 11 and 10 is for the entry and mainstream market, where prices are typically ~$120-150 for entry, and ~$200-249 for mainstream.

Unless NV is launching GP106 around the same time as Polaris, there's actually no competition as they are in different market segments.

The worse case scenario for AMD is GP104 is released in June 2016, priced very well, ie. the 970 successor will ~$330, and Vega arrives in early 2017. That gives NV a 6 months period where the well priced 2nd tier GP104 will encourage gamers who normally spend for mainstream GPUs to spend extra and get GP104.

However, on a new node that has not matured yet for big GPU dies, it will be more normal for GP104 to be more expensive.

There's this other scenario, don't discount it, because Raja was quite public when he made the claims that AMD will be several months ahead of the competition on launching next-gen FF GPUs.

http://arstechnica.com/gaming/2016/01/amd-confirms-high-end-polaris-gpu-in-development-for-2016/

"We believe we're several months ahead of this transition, especially for the notebook and the mainstream market," said Koduri. "The competition is talking about chips for cars and stuff, but not the mainstream market."

Raja's scenario: Pascal is paper launched in June when Polaris is released, retail availability a few months later.

 

[JDG1980]

You are suggesting that AMD will launch the competitor to GP104, which launch in June, in 2017 (according to AMDs own roadmap).
Are you not aware how much potential they would lose by letting Nvidia have almost all the cake for 6 months + ?

AMD and Nvidia are targeting different markets with the first generation of FinFET products, IMO. AMD wants laptop and AIO contracts, which is why Polaris has a small die size and focuses on energy efficiency and not absolute performance. Nvidia wants to win the HPC war with Intel, and sell more megabuck Quadro cards at the high end. Thus we'll be seeing Nvidia start off from the top down, and AMD work from the bottom up. Both companies may actually benefit from this since they won't be directly competing with each other in 2H 2016.

No, what will happen is that GP104 will be a small upgrade from GM200, so that GTX 1080 will beat GTX 980Ti by 20% ish.
AMD will follow with Polaris 10 which will also be 20% ish faster than GTX 980Ti.

Only 20% over GTX 980 Ti would be a pathetic showing on Nvidia's side. Just a raw die-shrink of GM200 would produce a ~300mm^2 chip, which is at the low end of where a "4"-series Nvidia chip usually lands in terms of die size. It's going to be using GDDR5X instead of standard GDDR5, so the bus width can be cut from 384-bit to 256-bit while actually increasing memory bandwidth. Then add the fact that you can crank up clock speeds substantially on FinFET compared to the older planar process, while keeping power consumption the same. That should be enough to give >30% performance increase over GTX 980 Ti (which, remember, isn't a full GM200) by itself.

Polaris 10 is a Pitcairn-sized chip. Faster than GTX 980 Ti is probably overly optimistic. Maybe in the most compute-focused titles like AOTS, but not across the board. This chip is really designed to offer performance on par with or slightly exceeding Hawaii and GM204 with much lower power consumption.

 

[jpiniero]

nVidia loves messing with AMD so a paper launch of desktop GP104 at Computex would be something they would totally do, especially if desktop Polaris hasn't been released by then. The thing is there really isn't anything blocking the release unless the chips simply aren't ready - TSMC increased it's 16FF wafer count in March, yields should be good and they are using GDDR5.

 

[JDG1980]

nVidia loves messing with AMD so a paper launch of desktop GP104 at Computex would be something they would totally do, especially if desktop Polaris hasn't been released by then. The thing is there really isn't anything blocking the release unless the chips simply aren't ready - TSMC increased it's 16FF wafer count in March, yields should be good and they are using GDDR5.

I would be very surprised if GP104 used standard GDDR5. Assuming a 256-bit bus width (which has been the case on every "4"-series Nvidia GPU up until now), it would be completely hamstrung by memory bandwidth limitations and not running anywhere near its full potential. This chip isn't going to be viable without GDDR5X.

 

[jpiniero]

it would be completely hamstrung by memory bandwidth limitations and not running anywhere near its full potential. This chip isn't going to be viable without GDDR5X.

GDDR5X isn't going to be realistic until the end of the year really if not into 2017. They aren't waiting that long. Besides, it gives them an easy way to refresh the lineup in 2017.

 

[LTC8K6]

I think we will see GDDR5X cards much earlier than most people expect.

 

[JDG1980]

GDDR5X isn't going to be realistic until the end of the year really if not into 2017. They aren't waiting that long. Besides, it gives them an easy way to refresh the lineup in 2017.

I think you're overestimating how long the lead times are. Just because GDDR5X only became public knowledge relatively recently doesn't mean that Nvidia and AMD haven't been working on it for a long time. And it's specifically designed to be an easy drop-in upgrade for GDDR5, so the earliest prototype cards could have used legacy GDDR5 until the newer memory was at least available for sampling. Micron is planning on starting mass production for GDDR5X this summer, which means risk production comes sooner (and we know sampling has already begun). GPUs can be, and have been, released with risk production parts before mass production officially starts.

If GP104 actually did use a 256-bit bus with standard GDDR5, then it's not going to be beating GM200 at all; it's going to be lagging behind because the needed memory bandwidth just won't be there. I don't think Nvidia is going to release a product like that.

 

[Silverforce11]

GPUs can be, and have been, released with risk production parts before mass production officially starts.

Sure, low volume, Fermi Teslas was the most recent one where it had to go premature with low yields and specs. Even Fury X, before TSV went into mass production, but again it was plague by out of supply despite the high prices on release for a top-end low volume SKU.

GP104 for mid-range with a market that has volume going GDDR5X on risk production? Nah.

No rush, folks will wait it out because they know Pascal is coming anyway and Polaris is not mid-range+ so it's not a competitor.

 

[Arachnotronic]

GDDR5X isn't going to be realistic until the end of the year really if not into 2017. They aren't waiting that long. Besides, it gives them an easy way to refresh the lineup in 2017.

Yep, any card coming in May/June will most likely not have GDDR5X.

 

[Sweepr]

I would be very surprised if GP104 used standard GDDR5. Assuming a 256-bit bus width (which has been the case on every "4"-series Nvidia GPU up until now), it would be completely hamstrung by memory bandwidth limitations and not running anywhere near its full potential. This chip isn't going to be viable without GDDR5X.

Agreed. It's either 384-bit GDDR5 with actual availability following Computex (fits the GM200 EOL rumour) or 256-bit GDDR5X paper launched at the event and little to no cards till sometime in H2.

BTW the smallest GPU with 384-bit GDDR5 right now is Tahiti at 352mm².

 

[RussianSensation]

If AMD is in the business of going bankrupt, well thats a good plan for that....

You are suggesting that AMD will launch the competitor to GP104, which launch in June, in 2017 (according to AMDs own roadmap).
Are you not aware how much potential they would lose by letting Nvidia have almost all the cake for 6 months + ?

No, what will happen is that GP104 will be a small upgrade from GM200, so that GTX 1080 will beat GTX 980Ti by 20% ish.
AMD will follow with Polaris 10 which will also be 20% ish faster than GTX 980Ti.

You dont have to go longer back than look at 7870 and 6970 to see similarities.

If Polaris 10 can beat 980Ti by 20%, but it only has 2304-2560 shaders, that would be the most revolutionary architecture from AMD in years. Sounds like fairy tales. Even a 35% increase in IPC coupled with a 30% increase in GPU clocks is not enough for such a low shader chip to beat 980Ti by 20%. Also, AMD disclosed Polaris 11 against 950; so clearly a budget offering with a sub-140mm2 die size. After that, they have hinted at bringing minimum VR spec to the masses. That means Polaris 11 is clearly not a 970/290 replacement. That brings us back to Polaris 10. When I think of the masses and where R9 390 sits @ $300 US market price and lower, there is no way I am going to believe that Polaris 10 is going to be = 980Ti+20% at $300 and lower (the mass market).

The way I am reading the rumors is that Polaris 10 and GP104 are not competitors. 1070/1080 are 970/980 replacements --> $330-550 market segment. By definition Polaris bringing 970/290/290X minimum spec to the masses is a completely different market for next gen than where 1070/1080 will be. That's my take on it. That's why I insist that AMD has no plan for most of 2016 to counter 1070/1080 other than lower prices on Nano/Fury/Fury X.

Those Fiji chips are going to be a year old soon and AMD already dropped the Nano from $649 to $499:
http://www.anandtech.com/show/9938/amd-cuts-price-of-radeon-r9-nano-to-499

Market prices for the Nano are $459-479. AMD will need to drop them even more once 1070 comes out. With 28nm and 1 year older HBM1 tech, Nano/Fury @ $399 and Fury X @ $499 is what AMD could do. However, I think it's not going to be enough of a price drop. This is why AMD is keeping prices on 390/390X/Fury/Fury X close to MSRP -- maximize profits while they can since once GP104 comes out, the party is over until Vega 11. AMD also probably has no resources* to launch Vega + Polaris so they chose <$300 market over $300+ market. From a market share point of view, it makes perfect sense. For enthusiasts, ya GP104 is going to be the way to go for most of 2016. That's my prediction for now. AMD's recent track record of 290/290X and Fury/Fury X suggest it needs more time to pump out more complex chips. Roadmap for Vega suggests this isn't changing any time soon.

*Another possibility is AMD is taking a massive risk of launching higher end part later than GP104, but going for Gold long term. If GP104 has GDDR5X, Vega is likely HBM2. That means AMD may give up the first 6 months of $350-$550 market segments but then go for the win over the next 1.5 years of that generation. Having said that, it's possible the Polaris 10 leaks we are seeing are for a cut-down part. If Polaris 10 beats Fury X, I just don't see how a 232mm2 2304 shader 256-bit GCN chip can do it. Sounds unreal. It would also mean automatic discontinuation of the Nano/Fury/Fury X lines. We heard a rumor of 980Ti and 390/390X being discontinued but none related to Fiji. Just connecting the dots. Moreso, AMD just paper launched $1500 Radeon Pro Duo. If Polaris 10 > Fury X, and uses less power, why in the world would AMD not launch a dual chip air cooled Polaris Pro Duo in June/July? Instead they went through all of this trouble of complex water cooling. THAT tells me straight up that Polaris 10 is slower than Fury X and it is not a competitor to Pascal GP104.

Bringing this back to GP104. If this is how AMD will play the game in 2016, NV could return to a 3-tier mid-range 660Ti/670/680 line from the same GDDR5X GP104 die.

Btw, I really don't want such a market scenario in 2016; yet Raja's continued emphasis on AMD having just 2 distinct Polaris chips for most of 2016 leads me to conclude the above. NV having AMD's roadmap with Vega way out means as consumers, we are probably screwed. It pretty much means NV can once again raise prices of the mid-range x04 series. Alternatively, if NV puts 1070 @ $330 and it ~ 980Ti's, I have no clue what AMD is going to do.

 

[RussianSensation]

Kind of sad that would be considered crazy since that's pretty much where we were just under a year and a half ago with the 290 and 290x.

Ya, but the general market didn't care. Here in Canada, after-market 290s could barely sell even for $280-300 CAD. The same product sells for $400+ as a 390. Look at the current U.S. prices for 390X/980 -- generally they are $390+.

R9 390X = $390+
http://m.newegg.com/ProductList?keyword=R9%20390x&sortField=PRICE

980 = $450+
http://m.newegg.com/ProductList?keyword=gtx980%20graphic%20card&sortField=PRICE

These prices are absurd when 290X was $250-280 but the market didn't care. So, in that context, Polaris 10 hitting $199-299 in a 110-120W TDP with 390X/980 performance is ground-breaking. It also means if I were NV, no way would I price 1070 @ $329 if it ~ 980Ti's performance when I was just selling 980 for $450 and AMD's Vega is out in late 2016/early 2017. What would you do if you were NV? Honest question -- I can sell 980 for $450+ in 2016 so why the heck would I sell 1070 ~ 980Ti for $329?

What about 1080? "I already raised mid-range 560Ti from $249 to $499 680. Next gen I bumped that up again to $550. Most high end buyers bought it anyway. For my 3rd gen, I'll release a card 20% faster and go up to $599-649. Why the hell not?! Fastest card on the market today costs $600. Isn't that reasonable? I am giving you 8GB of VRAM, lower power usage, more features and driver support."

Just saying