Nvidia : Pascal is 10X Faster Than Maxwell – Features Mixed Precision, 3D Memory and NV-Link

[Xsilent(X)]

[Deletive]

I'm more excited about volta even though we'll probably never see consumer GPU's from it.

[Doobeedoo]

It better be priced like how flagship hardware was prices years ago... like freaking half of today, and actually offered more.

[Lynx]

isn't NVlink only available on server and those kind of stuff?

[patrickjp93]

The only way Nvidia gets 10x faster at deep-learning AI is if they pick up branch prediction in the GPU, something that is thermally expensive (and the reason Intel and IBM are the two biggest AI researchers. AI is based on heavily branching code (you can TRY to reduce it to mathematical functions, but they break the moment you allow evolution and not just simple learning), and even parallelized there are lots of race conditions and branches to deal with. Providing a short-circuit call-response with neuron testing combined with even 50% accurate branch prediction (Haswell branch predictor is 96% accurate) would be a significant upgrade, but at what cost? 16nmFF+ is a great process shrink which should help power and heat a lot, especially if it's FD-FFS-SOI (Fully-Depleted FinFet Silicon On Insulator) which is what IBM went with to try to keep their Power 8 chips in heat envelopes just within the realm between obnoxious and insane (230W for the 12-core flagship).

 

That said, how many branch predictors and at what accuracy? If you wonder why x86 doesn't scale down super easily to lower heat envelopes it's a combination of ALU count, out of order processing, and branch prediction. To attempt this on something as massive as a GPU... It sounds like Fermi all over again...

[Texhnlyze]

isn't NVlink only available on server and those kind of stuff?

 

For now but when pascal comes around, it will be part of GPUs.

[zappian]

 Nvidia's beastly new Titan X graphics card wows you—and it should—well, hold onto your pants: The company said it's next-gen GPU technology, codenamed Pascal, will offer roughly ten times the performance of the Titan X, at least in compute tasks.

The next-generation GPU from Nvidia would also offer a host of new technologies, including 3D-stacked memory and NVLink, said CEO Jen-Hsun Huang during the GPU Technology Conference's keynote in San Jose. The GPU will essentially be a super computer, Huang said, and Nvidia is shooting for it to be ten times faster than today's Maxwell-based GPUs. 

 

Well I for one am EXTREMLY cynical about this.

We never see this kind of performance leap from one generation to another.

Only time will tell but I seriously doubt it.

 

Original Article: http://www.pcworld.com/article/2898175/nvidias-next-gen-pascal-gpu-will-offer-10x-the-performance-of-titan-x-8-way-sli.html

[Toddwjp]

key word was compute tasks. 

[givingtnt]

8 way sli ? find me a motherboard that supports that and im in

 

also, those mobile gpu that will come out of it will be SICK

[DIV1D3]

10 times the performance of a Titan X hey? So in reality 5 times the performance, and 20 times the price, right?

[ADZ_123_!"£]

The world isn't ready for this kind of stuff, remember the consoles that are still lagging behind...

[Michael McAllister]

10x performance?  Bullshit.

[CtW]

8 way sli ? find me a motherboard that supports that and im in

 

also, those mobile gpu that will come out of it will be SICK

Im pretty sure this would be done with 4 dual gpu cards...

[zappian]

key word was compute tasks. 

 

I noticed that lol.

Even then you rarely see this drastic increase in performance.

Maybe the arquitecure is THAT efficient at those tasks , we will see.

[Arty]

Something like this?

 

thats not sli

thier just in thier

[carolkarine]

-SNIP-

[Thought]

I don't see this as being impossible. There have been leaps and bounds in performance and efficiency with rendering tasks in the past few years up to twenty years. When they say they'll have "Roughly X performance" then I'll believe them because they've gotten very close to their targets before. Nvidia is more likely to hit that 10 GHz mark than Intel with how they're going too.

 

Something that hasn't been mentioned is whether or not Jen-Hsun said these are dedicated compute cards or consumer cards. I would hazard a guess saying dedicated.

[carolkarine]

I noticed that lol.

Even then you rarely see this drastic increase in performance.

Maybe the arquitecure is THAT efficient at those tasks , we will see.

 

 

key word was compute tasks. 

 

 

 

considering the DP compute on titan X is gimped to hell this shouldn't be terribly hard. it's a 1/32 of the SP because they had to cut out the DP compute units to save die space. now that they can get a much smaller process, they'll be able to ramp back up the DP compute perf. if they didn't in fact, it would be kinda sad. 10x is a minimum of what I'd expect.

 

too bad it'll still be overpriced.

[Gibbons]

I have a feeling a lot of prosumers are going to misinterpret  Pascal and think NVlink is going to quadruple their video-games FPS when they aren't even close to maxing out PCIe 3.0.

[zappian]

I don't see this as being impossible. There have been leaps and bounds in performance and efficiency with rendering tasks in the past few years up to twenty years. When they say they'll have "Roughly X performance" then I'll believe them because they've gotten very close to their targets before. Nvidia is more likely to hit that 10 GHz mark than Intel with how they're going too.

 

Possible ? yes

Unlikely ? Most definitely.

Nvidia has a track record of lie after lie , I will believe it when they have a prototype or when the final product comes out.

[Guest Kloaked]

@Prastupok

 

very vroom ram

 

many compute

[givingtnt]

Something like this?

 

thats not sli. do you see a bridge ?

[SurvivorNVL]

So, we can probably expect at max a 20-30% increase in FPS for games from the Titan X, maybe.  Can we just lose sli-bridges?  That'd be nice.

[givingtnt]

He was asking just about the motherboard.

its a mining rig. there are splitters. and or probably 2 motherboards

[asim1999]

8 way sli ? find me a motherboard that supports that and im in

 

also, those mobile gpu that will come out of it will be SICK

4 dual GPUs in SLI