[Kitguru]Nvidia`s big Pascal GP100 have taped out - Q1 2016 release

[Silverforce11]

kitguru fail. Hynix reported HBM2 availability in Q2, but now somehow Pascal is Q1. Either Pascal wont be HBM2 (which would be actual news and should be the title of the article) or this is 100% clickbait garbage.

HBM2 production starts in Q2 2016, volume expected in Q3 2016.

From reading GloFo's & Amtrak's materials, IIRC, interposer & stacking integrating the GPU & HBM takes a few months for the entire process.

So maybe NV rush to Pascal with HBM1, but a BIG Pascal with HBM1? How well do you think that would serve the main market for big Pascal, Tesla with 4GB?

As a reminder, the last time NV went to a new node on a big die first = FERMI, late, worse than expected and terribly low yields. They've learnt from that mistake with Kepler & Maxwell.

ps. I didn't need to click on that clickbait to know its chance of being correct is slim to none. If NV is peddling about its upcoming Pascal this early, it suggests Titan X is going to get creamed by Fury hard.

 

[hawtdawg]

no way in hell does Nvidia have a 16/14nm GPU out before fall of next year.

 

[jpiniero]

If NV is peddling about its upcoming Pascal this early, it suggests Titan X is going to get creamed by Fury hard.

I would think it's Knights Landing that has nVidia worried.

 

[Silverforce11]

I would think it's Knights Landing that has nVidia worried.

Tesla with HBM1 isn't going to be a threat in HPC. It's only purpose would be to reclaim the performance crown in the consumer 3d gaming market.

 

[ShintaiDK]

kitguru fail. Hynix reported HBM2 availability in Q2, but now somehow Pascal is Q1. Either Pascal wont be HBM2 (which would be actual news and should be the title of the article) or this is 100% clickbait garbage.

Be careful, its actually a production start roadmap and not when they ship actual products. Assuming no delays and we are already into Q3. As you say, I doubt we see HBM2 cards before Q4/2016 or Q1 2017.

 

[RussianSensation]

Be careful, its actually a production start roadmap and not when they ship actual products. Assuming no delays and we are already into Q3. As you say, I doubt we see HBM2 cards before Q4/2016 or Q1 2017.

Are you suggesting NV won't launch Pascal in 2016, or they will use HBM1/GDDR5 with first generation 16nm Pascal products?

 

[AnandThenMan]

Would Nvidia really shelve a new generation of GPU until the desired memory tech was ready? I doubt it, so either they go with HBM v1.0 or GDDR5, or both?

 

[dark zero]

I think that Pascal might come with HBM1, BUT... made in Intel fabs at 14 nm FinFet.

Yeah... some people might think that Intel would buy nVIDIA eventually, and seems logic since it can get the perfect performance in everything

 

[AnandThenMan]

Intel minting a chip for Nvidia that is in direct competition with their own products, yea I don't see that as likely.

 

[Rvenger]

Man, I would love to see Pascal with HBM2. I can see it right now. Running AMD Zen and Pascal as my new rig. I really hope it all comes true to be honest.

 

[raghu78]

Are you suggesting NV won't launch Pascal in 2016, or they will use HBM1/GDDR5 with first generation 16nm Pascal products?

for once shintaidk is not far off from the mark. HBM2 production ramps in Q2 2016. It takes 4-5 months to go through the various steps of 2.5D manufacturing as they have longer lead times due to more processing steps. So I would say Sep-Oct 2016 sounds right. Anybody expecting 16/14nm GPUs with HBM2 before that is setting themselves up for disappointment. :whiste:

btw the launch volume will be low and demand will be sky high as we saw at HD 7970 and GTX 680 launch. I expect AMD and Nvidia to push their 300 sq mm FINFET GPU prices upto USD 700. The price gouging by retailers who will want to cash in on the demand supply imbalance will be the icing on the cake. :biggrin:

 

[JDG1980]

We don't know when TSMC's 16nm FinFET+ process will become viable. Frankly, at this point, I'll believe it when I see it; TSMC has demonstrated that their marketing promises are worthless.

If it turns out that FinFET+ is ready before HBM2, then it wouldn't surprise me if Nvidia decided to launch Pascal by releasing a small chip using standard GDDR5. Although it will level off eventually, HBM2 will be more expensive at first, and that may not fit well with a GPU designed to sell in mid-range laptops and $100-$150 discrete cards. GM107 (GTX 750/750 Ti) is a 148 sq. mm. chip with 1.9 billion transistors. If we assume double density from the node shrink, then this should allow Nvidia to cram nearly 3.8 billion transistors into a 16nm FinFET+ chip of the same size. That is enough to handily beat the GM206-based GTX 960 (2.9 billion transistors, 227 sq. mm.) - even without taking architectural improvements into account. The TDP would presumably be about on par with existing GM107. The GTX 750 Ti was a big success because it was (and is) the most powerful card you can get without needing an external power connector. It also fits well into laptop designs. It's safe to assume that a hypothetical "GP107" would be successful for the same reasons.

A design pattern like this would mean that Nvidia wouldn't need to tackle a new node, new architecture, and new memory subsystem all at the same time. They could do the first two while proceeding with work on the new memory controller.

 

[videogames101]

So HBM is just a bunch of stacked DRAM, and uses a silicon interposer rather than a PCB to allow for massive IO width and multiple independent channels? Cool. I wonder how much cleaner your data eye is compared to traditional DDR modules with PCB traces. Cadence will already sell you a model too, man I wish I had could play with that for a few months.

As far as re-writing your memory controller/subsystem for something like this, I don't think it would be that bad. As far as I can tell it's wider and has similar latency compared to GDDR5? The signal is practically the same. Sure there are probably new optimizations to be made, like with this single bank refresh, data-bit inversion, extra channels and such, but IMO Nvidia should be able to hammer out at the very least a sub-optimal controller without any problems.

If you want to know when Pascal will be released, I doubt it will be held back by the HDL guys at Nvidia. I think the difficult part would be getting the interposer/PHY right. TSV's and interposers are old hat in research facilities, but not so much for high-yield and high-volume consumer products (as far as I'm aware). Look at TSMC and Hynix schedules, that's probably the best indicator. I would be surprised if any new memory sub-system isn't already working well in simulations, assuming they are actually going with HBM or HBM2.

 

[Innokentij]

About time, god knows how old this outdated 28nm is now. \o/

Edit: Dec 22, 2011 AMD 7970 first 28nm GPU MY GOD THATS OLD!

 

[Gikaseixas]

ps. I didn't need to click on that clickbait to know its chance of being correct is slim to none. If NV is peddling about its upcoming Pascal this early, it suggests Titan X is going to get creamed by Fury hard.

...and that's the intent of this thread, you nailed it.

 

[Alatar]

So in this site they, including me, are all shills, a big thank you to help us differentiate between the good and the bad, your arguments are indeed indiscutable...

I never said SA is filled with shills lol. I compared it to AMD zone which was an AMD fan page that constantly got almost everything wrong about intel / NV and was hopelessly optimistic about everything that is AMD.

That's not what shills are but a lot of people here seem to think that this pascal rumor comes from NV through some shills.

 

[dangerman1337]

So much of this rumor is unbelievable. For it to be real, there are three major and unlikely IF's that need to line up.

1) Finfets are getting better than expected yields at TSMC
2) Pascal GP100 isn't as big as either GK110 or GM200
3) Nvidia isn't leading a new family of GPU's with the small die first
.

Yeah that board being held up is obviously a mock up for marketing. That and the other information.

Looks like another junk information clickbait article.

Eyrines is the same guy who first said that 2nd generation Maxwell will be on 28nm and said that the GM204 would be out late Q3 and he was right compared to WccfTech trash articles keep saying that Maxwell would be on 20nm. The only one thing that is making me doubt is HBM2 but I'll go ask him that, I think this is going to be kind of a repeat of Maxwell with a GP2xx coming later in the year with HBM2.

 

[NTMBK]

Bring on the 14nm cards. I've been sat on this HD7770 since 2011, and I could do with an upgrade to handle all these next-gen games.

 

[flopper]

Must be magic when a card tapes out when HBM2 seems to start production 2016.

 

[maddie]

So HBM is just a bunch of stacked DRAM, and uses a silicon interposer rather than a PCB to allow for massive IO width and multiple independent channels? Cool. I wonder how much cleaner your data eye is compared to traditional DDR modules with PCB traces. Cadence will already sell you a model too, man I wish I had could play with that for a few months.

As far as re-writing your memory controller/subsystem for something like this, I don't think it would be that bad. As far as I can tell it's wider and has similar latency compared to GDDR5? The signal is practically the same. Sure there are probably new optimizations to be made, like with this single bank refresh, data-bit inversion, extra channels and such, but IMO Nvidia should be able to hammer out at the very least a sub-optimal controller without any problems.

If you want to know when Pascal will be released, I doubt it will be held back by the HDL guys at Nvidia. I think the difficult part would be getting the interposer/PHY right. TSV's and interposers are old hat in research facilities, but not so much for high-yield and high-volume consumer products (as far as I'm aware). Look at TSMC and Hynix schedules, that's probably the best indicator. I would be surprised if any new memory sub-system isn't already working well in simulations, assuming they are actually going with HBM or HBM2.

Do you think having the microbump connections to the interposer allow better signal routing to various parts of the die? Could this be one significant difference between memory controller designs.

 

[maddie]

Eyrines is the same guy who first said that 2nd generation Maxwell will be on 28nm and said that the GM204 would be out late Q3 and he was right compared to WccfTech trash articles keep saying that Maxwell would be on 20nm. The only one thing that is making me doubt is HBM2 but I'll go ask him that, I think this is going to be kind of a repeat of Maxwell with a GP2xx coming later in the year with HBM2.

This is a much more believable scenario. Small die with HBM1, then GP204 + hbm2 towards end of year.

 

[happy medium]

This is a much more believable scenario. Small die with HBM1, then GP204 + hbm2 towards end of year.

Yes I would believe a slower gtx 970/980 level chip with 4gb HBM1 first, then faster chips with HBM2 later. Its possible Nvidia will time Pascals arrival will HBM 2's availability.

 

[Cloudfire777]

Eyrines is the same guy who first said that 2nd generation Maxwell will be on 28nm and said that the GM204 would be out late Q3 and he was right compared to WccfTech trash articles keep saying that Maxwell would be on 20nm. The only one thing that is making me doubt is HBM2 but I'll go ask him that, I think this is going to be kind of a repeat of Maxwell with a GP2xx coming later in the year with HBM2.

He never said HBM2. He didnt even mention HBM. HBM is on Nvidia's roadmap for Pascal. Thats all we know.

There are also slides that show 2GB stacks for HBM1. 4 of them is one way to get 8GB. Who knows how far Hynix are down that route and if they can offer it to Nvidia in Q1.
Another possibility is eight 1GB stacks around the die instead of four of them.

Disregarding that Pascal wont be available in Q1 because of HBM2 like someone here tries to say, is just silly when Eyrines didnt even mention it and that its not even a requirement.

He says this about going GP100 first

This is a change from Kepler and Maxwell where the smallest chip (GK107 and GM107 respectively) taped out first. Maybe the experience with 20nm was enough for NV to go back to their usual big die first strategy (Fermi, Cloudfire)

 

[happy medium]

HBM is on Nvidia's roadmap for Pascal. Thats all we know.

I think they also said HBM2 for Pascal, am I right?

 

[Cloudfire777]

I think they also said HBM2 for Pascal, am I right?

They said it supports HBM2.
You know, the usual tease/hype without promising they will actually use it on all Pascal GPUs.

http://www.kitguru.net/components/graphic-cards/anton-shilov/ceo-of-nvidia-cant-wait-to-tell-you-about-future-pascal-products-in-the-pipeline/