NVIDIA GP100 Silicon Moves to Testing Phase

[btarunr]

NVIDIA's next-generation flagship graphics processor, codenamed "GP100," has reportedly graduated to testing phase. That is when a limited batch of completed chips are sent from the foundry partner to NVIDIA for testing and evaluation. The chips tripped speed-traps on changeover airports, on their way to NVIDIA. 3DCenter.org predicts that the GP100, based on the company's "Pascal" GPU architecture, will feature no less than 17 billion transistors, and will be built on the 16 nm FinFET+ node at TSMC. The GP100 will feature an HBM2 memory interface. HBM2 allows you to cram up to 32 GB of memory. The flagship product based on GP100 could feature about 16 GB of memory. NVIDIA's design goal could be to squeeze out anywhere between 60-90% higher performance than the current-generation flagship GTX TITAN-X.



View at TechPowerUp Main Site

 

[15th Warlock]

Tripping speed-traps already? Sounds like this is gonna be one fast processor!! ba bum tss!!

Seriously though, goodbye 28nm, you shall not be missed, about time we moved to a smaller process

 

[The Von Matrices]

Seriously though, goodbye 28nm, you shall not be missed, about time we moved to a smaller process

28nm will still be around for a long time for mid-range and low-end ASICs. The cost incentive to move to a process smaller than 28nm just great enough to incentivize pure die shrinks (i.e. die shrink with no architectural changes). While we might not see any new 28nm designs, the existing 28nm chips will likely remain on the market for years, albeit relegated to lower price points.

 

[FordGT90Concept]

Holy crap, $3100 USD each?

I really do hope they are 16nm parts. It's long past due.

 

[btarunr]

Holy crap, $3100 USD each?

I really do hope they are 16nm parts. It's long past due.

That's just something they scribbled for customs.

 

[64K]

It's going to kick ass! This will not be the normal GPU upgrade that we've come to expect. Next up. Arctic Islands.

 

[Xzibit]

That's just something they scribbled for customs.

It certainly isn't going to be lower than $650 ><

It will cost them buyers more to pull a Titan/GF Ti on GP100

 

[Eroticus]

Funny how people are thinking they would get normal price for new gen gpu and hbm2 memory .. this would be limited only to 1 high end card Titan and maybe later "1080 Ti" only cuz new AMD card ...

 

[mastrdrver]

I didn't think the HBM2 Specification had been finalized.

 

[the54thvoid]

If people want to make statements about the pricing in a negative manner it seems churlish given the recent history going back to what.... 8800 Ultra days? Flagship = most expensive (unless of course your flagship card under performs your main flagship card but you still sell it for $650 anyway).

Neither brand produces 'affordable' flagships these days. Unfortunately.

What is more important is how the architecture stacks out as AMD do have a bit of a laurel to sit on for DX12. Pascal has been touted as 'mixed' compute but that doesn't mean too much without knowing what the mix is. It needs heavy parallelism to match GCN's ability to render lots of disparate info queues. All those transistors will be less meaningful if Pascal doesn't address DX12's bare metal language.

 

[Xzibit]

If people want to make statements about the pricing in a negative manner it seems churlish given the recent history going back to what.... 8800 Ultra days? Flagship = most expensive (unless of course your flagship card under performs your main flagship card but you still sell it for $650 anyway).

Neither brand produces 'affordable' flagships these days. Unfortunately.

What is more important is how the architecture stacks out as AMD do have a bit of a laurel to sit on for DX12. Pascal has been touted as 'mixed' compute but that doesn't mean too much without knowing what the mix is. It needs heavy parallelism to match GCN's ability to render lots of disparate info queues. All those transistors will be less meaningful if Pascal doesn't address DX12's bare metal language.

From the Pascal presentation it means more FP16 (Half-Precision).

 

[FordGT90Concept]

That's just something they scribbled for customs.

I figured but there has to be some logic behind the number. I assume that's about how much it would cost them to make a replacement chip should they be lost/stolen/damaged. Prototypes are always more expensive to produce than production models.

From the Pascal presentation it means more FP16 (Half-Precision).

Who uses that? Silly NVIDIA. Even FP32 isn't adequate in a lot of situations. The focus should be on FP64, not FP16.

 

[the54thvoid]

From the Pascal presentation it means more FP16 (Half-Precision).

I'm still scratching my head - better go to work..... I'll continue scratching it when I return. What I didn't want to say in this thread for obvious reasons is what about asynchronous capability?

 

[FordGT90Concept]

Remember, asynchronous compute has been around since Direct3D 11 and apparently not many video game developers use it (or to the extent that GCN can handle anyway).

 

[Xzibit]

Who uses that? Silly NVIDIA. Even FP32 isn't adequate in a lot of situations. The focus should be on FP64, not FP16.

Well they been leaning out ever since Fermi. I suspect its less complexity at faster rate. We just don't know how that will translate to game performance.

 

[FordGT90Concept]

The only place I know of that games use FP16 is in images.

 

[ZeDestructor]

I figured but there has to be some logic behind the number. I assume that's about how much it would cost them to make a replacement chip should they be lost/stolen/damaged. Prototypes are always more expensive to produce than production models.


Who uses that? Silly NVIDIA. Even FP32 isn't adequate in a lot of situations. The focus should be on FP64, not FP16.

FP32 is enough for general/gaming graphics. FP64 is pretty much exclusive for professional graphics (CAD and the like) and HPC. FP16 is in a bit of a niche spot, but when it can be used, it's a healthy doubling of perf, for tiny amounts of extra die area. right now, FP16 is pretty much exclusively used by mobile games, and some incredibly niche HPC spots.

 

[Xzibit]

The only place I know of that games use FP16 is in images.

Post-processing effects

 

[Ebo]

Get a grib guys, its only ingeneering samples there no need to go bonkers yet.

Pascal is a long way from being finished.

We will all know what it can do not at launch, but when review samples is tested.

 

[Parn]

60 - 90% performance improvment. Well this certainly looks promising, but I guess the price will also be TITAN like.

Anyway I'll be waiting for a full GP104 based product.

 

[HumanSmoke]

The only place I know of that games use FP16 is in images.

Well, you can add Neural Nets (Deep Learning) to your knowledge base (Pascal specificity in this article)
FWIW, I think you'll find that AMD will also target mixed compute modes (FP16/32/64) for the same reasons that Nvidia and ARM are integrating it. Not every workload requires the power budget or restraints of FP32 or FP64.

 

[bug]

Get a grib guys, its only ingeneering samples there no need to go bonkers yet.

Pascal is a long way from being finished.

We will all know what it can do not at launch, but when review samples is tested.

Well, at this point it certainly seems Nvidia is further along than AMD. Which is still not Earth-shattering, but it's still worth noting.

 

[okidna]

Who uses that? Silly NVIDIA. Even FP32 isn't adequate in a lot of situations. The focus should be on FP64, not FP16.

That's not for the graphic/gaming department, what you see on the slide that Xzibit posted is the performance comparison between Maxwell and Pascal for their cuDNN framework (CUDA framework for deep learning).

The full slide : http://on-demand.gputechconf.com/gtc/2015/presentation/S5715-Keynote-Jen-Hsun-Huang.pdf

The only place I know of that games use FP16 is in images.

And you're right, right now the best deep learning architecture for NVIDIA GPU and cuDNN is the deep CNN (Convolutional Neural Network) which most researcher uses for image (2D) classification and detection.

 

[deemon]

So what about async compute in Pascal? Is it fixed now? or not?

 

[rtwjunkie]

So what about async compute in Pascal? Is it fixed now? or not?

We won't know the answer to that (the extent of async compute support) for many months!