NVIDIA Pascal Thread

[Glo.]

So is it 3584 cores or 3584 + 1792?

At the same page you have comparison with Kepler and Maxwell. Does biggest GPUs from both architectures come with 2880 + 960 cores, and 3072 + 4 cores?

No. It is ONLY 3584/3840 CUDA core GPU.
If there would be separate FP64 units, then ok. But they are part of SMM's.

 

[sontin]

Hitting the die size on a node is entirely a question of economics. I'm sure any ASIC vendor could have pushed the limits of nearly any node in recent memory. It just has never made economic sense to do so, because the cost per unit is very high with such low yields. That's why 600mm2 came last for 28nm for consumers -- only then did it make financial sense.

What? GK110 was introduced at the GTC2012 and was shipped to customers in June 2012.

That was 1 1/2 years before AMD introduced Hawaii.

 

[airfathaaaaa]

Hmm, 64 Cuda Cores per SM for Pascal.

GCN also has 64 Shaders per CU.

who was the one shouting a week ago that pascal will be on the level of gcn 1.0 but nothing more? yeah i guess he was the one that fall into it spot on i guess

 

[Timmah!]

The whole Chip has 60 Cluster with 64 FP32 Cores and 32 FP64 Cores.

So based on this and that comparison table posted on that Nvidia blog, does that mean that GK110 had actually 3840 cores (2880 + 960)? Thus actually more overall than GM200?

 

[Kris194]

It seems that Pascal will smash Maxwell cards Just like I expected.

 

[sontin]

So based on this and that comparison table posted on that Nvidia blog, does that mean that GK110 had actually 3840 cores (2880 + 960)? Thus actually more overall than GM200?

Yes. But only the FP32 or FP64 cores can be used.

 

[AtenRa]

Yes. But only the FP32 or FP64 cores can be used.

Dont think that is how it works.

 

[Qwertilot]

The haters were right about one thing - no public demo does kind of make it seem like a launch in June unlikely. September might be a more realistic timeframe for any consumer GPU.

I wonder slightly. One of the big questions was (esp with AMD's very small die first strategy) whether the processes could actually do a big/high TDP chip. Think we've got an answer to that now

Might even be HBM2 bottlenecking availability here, with that all being new too.

The only real obstacle to GP104 seems to be GDDR5X/coping with all the engineering etc. We'll see I guess.

 

[Erenhardt]

Yes. But only the FP32 or FP64 cores can be used.

Titan had an option to enable FP64 cores in drivers. Those were dedicated units.
Pascal looks like beefed up kepler.

 

[USER8000]

Look what NV did on 28nm with GK110 (high compute) -> GM200 (low compute focus on gaming). I would guesstimate that with 610mm2 die size, they will repeat the same strategy with GP100 -> Volta. Being stuck on the same node means to get another 50-75% boost probably means going all in on lean gaming focused Volta. Then GP100 would serve as their compute backbone until 7/10nm. If they create a straight up FP32 gaming Pascal, with the performance characteristics you describe, they will have nothing to sell during Volta generation.

This could also explain the aggressive 2017/2018 span between Pascal and Volta. NV could neuter compute and improve the architecture. Pretty surprising turn of events after some people were predicting 3500-4072 CUDA GP104. This means if GP104 has 3072 CUDA cores and it overclockes better than GP100, the gap between them could be 25% or less. A max overclocked Titan X/980Ti smashes 980 OC by 30-40%. Interesting times ahead.

Except the GK110 based Titan appeared nearly 6 months after we first started hearing about initial shipments to ORNL and that was months after they first started talking about the GK110.

May 2012 mention about GK110:

http://arstechnica.com/gaming/2012/...e-gpu-for-streamed-desktops-and-cloud-gaming/

First indications of shipments of GK110 based cards to customers to like ORNL in September 2012:

http://www.hpcwire.com/2012/09/10/o...batch_of_kepler_gpus_for_titan_supercomputer/


First GK110 based Titan cards in Feburary 2013:

http://www.anandtech.com/show/6760/nvidias-geforce-gtx-titan-part-1

So easily 8 to 9 months.

Then a few more months later for the cheaper GTX780 to be launched:

http://www.anandtech.com/show/6973/nvidia-geforce-gtx-780-review

So 12 months.

Then it took another two years,for costs to go down for the gaming orientated GM200 based cards with a similar die size to appear.

The GM200 cards did not need to subsidised anywhere as much by professional sales.

I think you are too optimistic for a 600MM2 gaming orientated die to be coming anytime soon.

Going by 28NM it took three years for it to happen.

That would mean something like 2019 for a dedicated SP die around 600mm2 appearing for gaming cards.

 

[Timmah!]

Dont think that is how it works.

How does it work then?

@Sontin: If what you say is true, why is GM200 as big as GK110, even bigger if it has actually less cores? I understand there are other parts of GPU design beyond shaders, lies the difference between them there?

 

[antihelten]

Titan had an option to enable FP64 cores in drivers. Those were dedicated units.
Pascal looks like beefed up kepler.

Beefed up Kepler, with the shader to SM ratio rebalance from Maxwell.

 

[Headfoot]

What? GK110 was introduced at the GTC2012 and was shipped to customers in June 2012.

That was 1 1/2 years before AMD introduced Hawaii.

What does that have to do with anything?? OT...

 

[xpea]

Guys, it's pretty obvious that GP100 is exclusively an HPC chip... only by the fact that it has no ROPS ! This thing will never be put inside a consumer board.
And that's why this new GP102 makes sens. Remove most of FP64 ALUs, integrate as much as possible FP32 ones, remove NVlink, replace HBM by GDDR5(X), reduce the oversized 14MB(!!!) register file (but good for compute) and you can get the ultimate gaming chip for Titan and xx80Ti boards we love

 

[SteveGrabowski]

Was today the only presentation Jen is scheduled to give? So no chance of hearing about GP104?

 

[Genx87]

This is a massive upgrade over Maxwell in this market. Holy crap Nvidia is taking Intels Phi seriously and came up with this monster.

Would be nice to know what the consumer gaming parts will look like. Will they go with something like Kepler\Pascal or Maxwell?

 

[Kris194]

No, consumer cards should be announced at Computex.

 

[Arachnotronic]

This is a massive upgrade over Maxwell in this market. Holy crap Nvidia is taking Intels Phi seriously and came up with this monster.

Would be nice to know what the consumer gaming parts will look like. Will they go with something like Kepler\Pascal or Maxwell?

Ayup. This looks like a very effective Phi killer.

 

[Stuka87]

Not sure why everybody is so happy about this. This is a Tesla card, it has *NO ROPS*, it is impossible to use this GPU in any sort of consumer card. Yes it will be awesome for those that need big compute power in their super computers, but pretty sure nobody here falls under that category.

 

[xpea]

This is a massive upgrade over Maxwell in this market. Holy crap Nvidia is taking Intels Phi seriously and came up with this monster.

NV is destroying Phi
Especially when we know how badly Phi performs on real HPC applications (contrary to the simple linpack stuff intel likes to show). Even more important, with the massive FP16 score and huuuuuge register file, GP100 has no competition for Deep learning (biggest trendy market)

 

[USER8000]

This is a massive upgrade over Maxwell in this market. Holy crap Nvidia is taking Intels Phi seriously and came up with this monster.

Would be nice to know what the consumer gaming parts will look like. Will they go with something like Kepler\Pascal or Maxwell?

Emm,you mean Kepler?? The GK210 was used in those markets not the GM200.

NV is destroying Phi
Especially when we know how badly Phi performs on real HPC applications (contrary to the simple linpack stuff intel likes to show). Even more important, with the massive FP16 score and huuuuuge register file, GP100 has no competition for Deep learning (biggest trendy market)

Intel can afford to throw billions at the market - you only have to look at how they have managed to sneak into the top 10 list of the TOP500.

 

[Kris194]

Not sure why everybody is so happy about this. This is a Tesla card, it has *NO ROPS*, it is impossible to use this GPU in any sort of consumer card. Yes it will be awesome for those that need big compute power in their super computers, but pretty sure nobody here falls under that category.

It doesn't matter. Look at the Tesla M40, it has 3072 CUDA cores, just like Titan X Also it shows that Nvidia is all about increasing performance while keeping/slightly increasing TDP. That's exactly what I wanted to see.

 

[Genx87]

Not sure why everybody is so happy about this. This is a Tesla card, it has *NO ROPS*, it is impossible to use this GPU in any sort of consumer card. Yes it will be awesome for those that need big compute power in their super computers, but pretty sure nobody here falls under that category.

It gives us a glimpse into what Nvidia has in store for Pascal.

 

[Erenhardt]

Guys, it's pretty obvious that GP100 is exclusively an HPC chip... only by the fact that it has no ROPS ! This thing will never be put inside a consumer board.
And that's why this new GP102 makes sens. Remove most of FP64 ALUs, integrate as much as possible FP32 ones, remove NVlink, replace HBM by GDDR5(X), reduce the oversized 14MB(!!!) register file (but good for compute) and you can get the ultimate gaming chip for Titan and xx80Ti boards we love

LOL WUT. It can't run crysis... D:

 

[Adored]

LOL WUT. It can't run crysis... D: