Nvidia GeForce GTX 1080 rumor mill just blew up.

Exciting developer tools for deep learning. However, wish I had more "deep learning" about my next SLI setup.
 
Yeah no consumer cards were talked about, but it was pretty cool. Although, he said that GP100, at 600mm^2 is the biggest finfet chip ever made, WRONG, Knights Landing is, at ~683mm^2 on Intel's 14nm, which is superior to TSMC's 16nm FF+ as well, plus I believe KNL is already shipping.
 
Yep, since they went with HBM2 on the tesla, the only way pascal is hitting consumers this year at remotely reasonable prices is if they created a whole separate chip with GDDR. Which seems unlikely.

Huge opportunity window for AMD here.
 
trudude said:
Looks like nay sayers were wrong about the new Titan having HBM2.
Click to expand...

No., they're just going to be shipping what is essentially HBM2 samples. And that's why they've only announced these supercomputer clusters at an outrageous price - to get whatever they can before they can source these cards in quantity.

HBM2 will be available in 2016, just not in quantity. Expect Big Pascal sometime next year.

Smaller Pascal will not be using HBM, that's already known. That should be released "soon."
 
You think P204 will be on GDDR? They'll split their lines like that? How is that "known"?
 
schizo said:
You think P204 will be on GDDR? They'll split their lines like that? How is that "known"?
Click to expand...

It's what AMD did last generation. How fucking hard do you think it is?

Even on the exact same chip, AMD proved you could support two different memory standards (HD 4850 = GDDR3, HD 4870 = GDDR5).

It's not hard. There are tons of Nvidia desktop GDDR5 GPUs (i.e. GTX 750 Ti) that use DDR3 on the notebook side to save cost/power.
 
No they didn't. The 390 series was all rebrands, only the furys were new.

HBM needs to be ON the chip. It's a whole different ballgame.
 
If they were releasing consumer Pascal cards "soon" then why didn't they mention it?

Why no working silicon if it's so soon?
 
schizo said:
No they didn't. The 390 series was all rebrands, only the furys were new.
Click to expand...

Go back to school because you missed the mark by a mile there.

Tonga was half of a Fury. SAME ARCHITECTURE, SAME FEATURES, DIFFERENT MEMORY CONTROLLERS.

Tonga = GDDR5.
Fury = HBM1
 
spintroniX said:
If they were releasing consumer Pascal cards "soon" then why didn't they mention it?

Why no working silicon if it's so soon?
Click to expand...
It's gotta be far off, he didn't even have a board or GPU to show. Very telling.
Although I cut off the stream right after he showed their new super computer.

It's very impressive but also 8+ MONTHS away.

Tahiti was half of a Fury. SAME ARCHITECTURE, DIFFERENT MEMORY CONTROLLERS.

Tahiti = GDDR5.
Fury = HBM1
Click to expand...
Tahiti...?
Tonga is a better example. Fury X is literally two 380X's duct taped together.
 
There is no need to mention it yet, why? That isn't the point of this conference, its not about gaming its all about HPC and Compute. If they already have p100's in mass production, you don't think p104's are going to be harder to produce? If they are releasing software that is specific for Pascal in June, don't people need those cards to use that software?
 
TaintedSquirrel said:
It's gotta be far off, he didn't even have a board or GPU to show. Very telling.
Although I cut off the stream right after he showed their new super computer.


Tahiti...?
Tonga is a better example. Fury X is literally two 380X's duct taped together.
Click to expand...

You're right, I mistyped. I meant Tonga!

Tonga = 1/2 Fiji :D
 
You're right, I forgot fury was still GCN 1.2.

Back to topic, P104s-- if they were coming, wouldn't they make the keynote?
 
If they are in mass production then surely he could have gotten his hands on one, right?
It's fluff. The GP100 is so impressive because it doesn't exist. This is a pre-emptive strike to steal AMD's thunder.

I would not be surprised at all if he was flat out lying.
 
schizo said:
You're right, I forgot fury was still GCN 1.2.

Back to topic, P104s-- if they were coming, wouldn't they make the keynote?
Click to expand...

razor1 said:
There is no need to mention it yet, why? That isn't the point of this conference, its not about gaming its all about HPC and Compute. If they already have p100's in mass production, you don't think p104's are going to be harder to produce? If they are releasing software that is specific for Pascal in June, don't people need those cards to use that software?
Click to expand...
 
TaintedSquirrel said:
If they are in mass production then surely he could have gotten his hands on one, right?
It's fluff. The GP100 is so impressive because it doesn't exist. This is a pre-emptive strike to steal AMD's thunder.
Click to expand...


There is no way IBM, HP, and others making their servers around fluff.... They would have needed pascal well before this. OEM verification of a server takes several months as each piece as to be verified and then as the components are put together, another verification step takes place. So for a Q1 launch of those products, they would need it now or earlier. And nV would not have put their names up on this if they were talking about fluff, because if they did, they got permission from those companies to talk about what they are doing.
 
schizo said:
You're right, I forgot fury was still GCN 1.2.

Back to topic, P104s-- if they were coming, wouldn't they make the keynote?
Click to expand...

I dunno, why steal the show with P104 when you've already stolen the show with P100? There's plenty of shows still to come before summer starts, and they could stick it in any of those.

Note, I said "soon" because we don't know how long it will be. Could be weeks. Could be months. The only solid date we had for a Pascal part announcement was GDC :(
 
Yep, we have no clue how long.

Should be a solid window of opportunity for AMD.
 
schizo said:
Yep, we have no clue how long.

Should be a solid window of opportunity for AMD.
Click to expand...


I agree. Let's see what they do with it. You can win with a horse in a one horse race, but it's much harder to do if that horse starts running the wrong way on the track.
 
defaultluser said:
They're saying a ~6x speedup in deep learning performance in a cluster (fp16) versus the older cards (I scaled the Pascal cluster down to 4 to be comparable).

There's only a 75% increase in compute performance over Titan X.

So this must be due to the new NVLINK interconnect plus the much faster memory bus? Guess Maxwell compute was truly bandwidth-limited...but we knew that already :D
Click to expand...
This gpu has 10.75tflops of fp32 throughout, ~21 tflops fp16.

Unless they release a totally modified Gp100 with reduced fp 64 cores and more fp32 I think I'm sticking to my 8.5tflop monster

Yet to see if fp16 capabilities will appear in consumer line, if so I'll be more than happy to buy gp104

Also seems like they chose to produce a lower density chip, they can fit ~20B transistors on 600mm easily

I think the big increase in fp32 comes from clock, gm200 throughput was based on base clocks not boost. At 1500mhz a 980ti does 8.5T
 
Last edited:
Ieldra said:
This gpu has 10.75tflops of fp32 throughout, ~21 tflops fp16.

Unless they release a totally modified Gp100 with reduced fp 64 cores and more fp32 I think I'm sticking to my 8.5tflop monster

Yet to see if fp16 capabilities will appear in consumer line, if so I'll be more than happy to buy gp104

Also seems like they chose to produce a lower density chip, they can fit ~20B transistors on 600mm easily

I think the big increase in fp32 comes from clock, gm200 throughput was based on base clocks not boost. At 1500mhz a 980ti does 8.5T
Click to expand...

Oh damn you're right, I FORGOT!

I thought Maxwell already had double-speed FP16, but I was thinking of the Tegra X1 announcement months ago.

This means half of that 6x performance improvement can be attributed to the double fp16 throughput. So It's a lot more reasonable to place the 6x speedup on 3x higher fp16 compute plus the extra bandwidth :D

I guess I shouldn't try to think and post when I'm following something like this.
 
The surprise for me is the size, took years for 28nm node to produce 600mm dies

Despite the
defaultluser said:
Oh damn you're right, I FORGOT!

I thought Maxwell already had double-speed FP16, but I was thinking of the Tegra X1 announcement months ago.

This means half of that 6x performance improvement can be attributed to the double fp16 throughput. So It's a lot more reasonable to place the 6x speedup on 3x higher fp16 compute plus the extra bandwidth :D

I guess I shouldn't try to think and post when I'm following something like this.
Click to expand...
Having said that if thus overclocks anything like maxwell it will be truly monstrous
 
Well this is a half node, so I don't think it was as hard of a transition as one would expect.

Yeah if they can get a 50% increase from overclocking, like what Maxwell 2 gets that would be crazy.
 
razor1 said:
Well this is a half node, so I don't think it was as hard of a transition as one would expect.

Yeah if they can get a 50% increase from overclocking, like what Maxwell 2 gets that would be crazy.
Click to expand...

Why do you call this a half node? Every node change bears a number that does really resemble the actual gate width, but there are also several other metrics you can use.

Based on glofo info on their "14nm" transistor density can reach 3.3x 28nm numbers, and operating such a density of transistors at 2400mhz~, can't remember exact number, we'd hit 8x thermal density!

8x!!!

I strongly suspect overclocking will run into very hard thermal limits in gp100

Also, ive been wondering, is hbm factored into die area? Is the interposer? Memo y controller?
 
Well the masks and metal layer are similar to 20nm.

HBM memory controller is part of the die space, but it is much smaller that traditional GDDR, which HBM is around 30% smaller.

yes I expect a hard limit as well, with power concerns with OC'ing, Fin Fet's transistor tolerances to frequency increase tend to scale up fast once out of its ideal ranges.
 
razor1 said:
Well the masks and metal layer are similar to 20nm.

HBM memory controller is part of the die space, but it is much smaller that traditional GDDR, which HBM is around 30% smaller.

yes I expect a hard limit as well, with power concerns with OC'ing, Fin Fet's transistor tolerances to frequency increase tend to scale up fast once out of its ideal ranges.
Click to expand...

Yeah leakage goes out the window at high frequency, look at how fucking hot skylake gets and Intel's 14nm is more mature and feature size is actually smaller in several areas
 
They are already clocking it pretty high. I highly doubt you will be able to get much beyond the 1500Mhz that we are seeing with Maxwell chips anyways.

In any case, if anyone is wondering about the ACTUAL size of the current FinFet process nodes, here ya go:

Cell-SizeComparison.png



So, it will be very interesting to see what TSMC does for their 10nm node, will they make up for the lost scaling this time? Or just scale a normal amount and and up having a 'bigger' 10nm node than everyone else?
 
noobthegreat said:
Boo! No consumer card timeline. :mad:
Click to expand...

I'm now guessing "little Pascal/Polaris" is late Q3 this year (school season release hints) and "big Pascal/Vega" will be late Q1 to mid Q2 next year. From what has been presented by NVIDIA, no real compelling reason to move from my R9 390 Dual Core (seems that a custom OC GTX 980 Ti (AMP, Classified, Lightning, HOF, etc. will get around 30% over stock 980 Ti performance, and little Pascal looks like "maybe" a 10-20% performance boost on the 980 Ti, on stock/mild overclocks. Not really worth getting at all, IMO, unless power constraints are a big issue).
 
noobthegreat said:
Boo! No consumer card timeline. :mad:
Click to expand...

They don't need to have a timeline for consumer card releases. All they needed to do was show off a professional card with super stats, and everyone will hold out for the consumer version of it. No matter how long it takes to come to market or the fact that the super card might be $5,000 - $10,000 because consumers want their next purchase to be associated with that professional super card.

Simple Marketing 101.
 
Paper launch? To head off Intel? Nvidia maybe on the verge of loosing the HPC market or start the downward spiral. Not sure if AMD is seriously targeting this market yet.

If Nvidia has P100 working then P104 should not be far behind and this event does not look appropriate to discuss gaming cards that cost pennies compared to these bad boys.
 
noko said:
Paper launch? To head off Intel? Nvidia maybe on the verge of loosing the HPC market or start the downward spiral. Not sure if AMD is seriously targeting this market yet.

If Nvidia has P100 working then P104 should not be far behind and this event does not look appropriate to discuss gaming cards that cost pennies compared to these bad boys.
Click to expand...

Still, GP100 specs give us a good foundation for estimation re GP104 :) 2560 cores for the full chip seems reasonable
 
spintroniX said:
If they were releasing consumer Pascal cards "soon" then why didn't they mention it?

Why no working silicon if it's so soon?
Click to expand...

Well a big reason could be that Maxwell was never a good card for the HPC market and so it is a lucrative area that has been missing a GPU model for awhile now.
This had to be addressed, and analysts awhile ago were mulling over that from a business strategy NVIDIA needed to focus back on the HPC type market first and then consumer.
That said I am surprised they are even at this stage with a large die/high power/high clock Pascal and I look forward to seeing what response there will be from Samsung-GF/AMD in next few months (if at all).

Does it exist?
Well they mention the Alexnet performance was averaged from 20 runs.
Furthermore from a business practice it really makes sense for them to maximise their profit margin by only allowing these GPUs to be provided direct and in the market willing to pay top dollar for single card HPC performance such as this.
This may be compounded that some collaborations with NVIDIA such as IBM want to use the NVLINK as a complete unified GPU-CPU architecture and so there will be projects that have access to these cards first.
I really doubt there will be enough cards to go around for awhile, IMO the achievement by TSMC should not be underestimated though.

Also I am not surprised they omitted any consumer schedules, it would tank current sales of their popular mid-to-upper Maxwell cards, but then as I mention their priority needs to be HPC type markets as they have been "ignored" since Keplar.
But still this could be a good indicator that we could see 970/980 replacements "soon".
Cheers
 
They launched the Titan X at GDC last year during the deep learning / self driving car keynote. They also talked about Pascal. That's why I was hopeful we'd at least get an update.
 
They mentioned it but didn't go into any details if I remember correctly at least outside of its compute capabilities.
 
You must log in or register to reply here.
Back
Top