NVIDIA GeForce 20 Series (Volta) to be released later this year - GV100 announced

[Glo.]

That's not good news, if true. It kind of lends credibility to that "only 16,000 Vega units" rumor.

If there is 16 GB version of the GPU, where are 8 GB stacks in the product stack of SK Hynix?

 

[Saylick]

If there is 16 GB version of the GPU, where are 8 GB stacks in the product stack of SK Hynix?

You're not suggesting that there's another Vega SKU with 4 x 4GB HBM2 stacks, are you?

 

[Glo.]

You're not suggesting that there's another Vega SKU with 4 x 4GB HBM2 stacks, are you?

No, since the architecture can work with up to 2048 bit memory controller.
At least for current version of the design.

 

[LTC8K6]

IIRC, GDDR5X was not available either when the 1080 popped up with it on board...

 

[Det0x]

Not so great news for Nvidia Volta V100 as Google revealed some details of it's TPU2.
180 TFLOP/s both for training and inferencing.

 

[Magic Hate Ball]

If there is 16 GB version of the GPU, where are 8 GB stacks in the product stack of SK Hynix?

Apparently on the Vega Frontier Edition

 

[Bacon1]

If there is 16 GB version of the GPU, where are 8 GB stacks in the product stack of SK Hynix?

Did they list the HBM 1 ones AMD used? Does Samsung list the HBM2 ones Nvidia uses?

AMD said multiple times that Vega is using 16GB

 

[Cookie Monster]

Not so great news for Nvidia Volta V100 as Google revealed some details of it's TPU2.
180 TFLOP/s both for training and inferencing.

They are not direct competitors. The use case for V100 is vastly different to the TPU (they happen to overlap in Tensor based work). The TPU is pretty much the definition of an ASIC (application specific integrated circuit), in that it just does Tensor based work and thats it.

The V100 on the otherhand can be used for a variety of computations which also includes Tensor based work (it can do this stuff in parallel with FP64/FP32 work for instance). Perhaps post processing with the CUDA cores while running machine learning algorithm via Tensor cores. Not just deep learning but it can do other things which makes it different to the TPU.

And to my knowledge Google happens to be one of nVIDIAs HPC customers.

 

[Ajay]

IIRC, GDDR5X was not available either when the 1080 popped up with it on board...

Yeah, what's on a public data book isn't the same as what is available to a high volume customer willing to pay for the latest and greatest DRAM.

 

[raghu78]

Yeah, what's on a public data book isn't the same as what is available to a high volume customer willing to pay for the latest and greatest DRAM.

very true. Samsung does not have any databook in their website which mentions HBM2 but we know they have been supplying Nvidia with HBM2 for P100 for around a year now. Similarly Hynix is AMD's HBM2 supplier and I am pretty sure there are agreements between these two companies for HBM2 supply just as would be the case between Nvidia and Samsung.

 

[Glo.]

Did they list the HBM 1 ones AMD used? Does Samsung list the HBM2 ones Nvidia uses?

AMD said multiple times that Vega is using 16GB

I suggest reading the context of that post, you responded .