Nvidia Ampere Discussion [2020-05-14]

NVIDIA clearly stated TSMC will produce the majority of 7nm orders, Samsung will get minor orders. The silliness is thinking NVIDIA will bifurcate their design across two foundries with two different fabrication processes.

 

No it IS. A100 shares the same arch as Geforce Ampere, same CUDA cores, caches, Tensor core and everything, NVIDIA will add new generation RT cores, reduce the number of Tensor cores and call it a day, you don't use two different fabrication processes for this.

 

And if this is 350W on TSMC 7nm then it means Ampere is garbage. I'd rather believe it's that high power because it's on Samsung.

 

That's funny, did you by any chance get to know what kind of performance you are getting?

 

My guess is that GA102 & GA104 will be fabbed @TSMC 7nm, and the smaller GA106 & GA107 @Samsung 8nm or 7nm.

 

The thinking was "gaming" ampere would not be 7nm , but 8nm from samsung, so this quote would still be right. And everybody was fighting over TSMC capacities at 7nm... A cheaper but good 8nm process could be a logical solution.

If it's silly to you, ok I guess...

 

Again why? If those are two different archs then that's fine I guess, but they are one arch with different configuration, you don't bifurcate your design like that unless the fabrication processes are identical or similar enough.

 

It seems TSMC has strict NDA and non-compete (NDA'd team cannot communicate with another foundry for multiple years) clauses in their contracts for use of their bleeding edge nodes. So, either NVidia has signed on the dotted line and there's a "single design" (as @DavidGraham asserts) or NVidia is multi-fab and is using multiple teams, hence multiple designs.

If NVidia has done any work at Samsung, then all that work is toast if NVidia is now 100% TSMC. Or the TSMC element of that work was on non-bleeding-edge TSMC.

3 stark choices. Take your pick.

 

Wanna bet on the "same CUDA cores" part?
GA-variants won't have FP64 CUDA cores which are separate from FP32 CUDA cores but in the same SM. Then there's of course rumors regarding different FP32/INT32 split, but they could be just rumors and nothing more. And there's more than just RT-cores missing from A100, which will be in GA-series (not in the SM units though, ROPs are outside for example)

 

Nvidia confirmed, that GA100 is a complete graphics chip. Not sure about the RT cores though, but ROPs, Raster and Display Engines are all there.

edit: re-read it, RT cores are not part of GA100.
Wrote about it here, but you'd have to take my word for it anyway that I did not make that up out of thin air.


edit2: And please let me state the perfectly obvious in saying that this would mean no GA100-Titan cards this round.

 

NV has been using both TSMC and Samsung for quite a while now, guys.

 

It seems to me that the power requirements for the high end ampere cards, to the point of making up a new 12pin power connector, are an indicator that the rumors of nvidia going with Samsung 8nm (equivalent to TSMC 10nm in performance IIRC) are true.

We know from Microsoft's latest presentation that large 7nm SoCs are very expensive even if yields are good, and we also know nvidia got comfortable with making very large chips for different performance brackets with Turing.

This doesn't mean nvidia won't win in absolute performance on their $1500 halo products (or they wouldn't have bothered with GDDR6X). But this could leave some room for 7nm AMD GPUs to finally get some notebook dGPU market, as suggested before.

 

320W is disappointing for the 3080 if it has the same number of SMs as a 2080 ti with a less than 200 MHz clock increase. That’s a 70W increase in power (30%) over the 2080 ti with only a 11% clock boost, on a node shrink. Only the that makes sense is the SMs are wider or the dumped most of that power into RT and tensor cores. AMD looks positioned to be able to smoke them on raw raster performance, but I guess we’ll see.

 

2070 looks very similar to 2080 super but 30W LESS. Having a 100W gap between 3070 and 3080 doesn’t make a lot of sense to me. I wonder if 3080 will be the dud in this lineup.

 

The 3090 really has to be very fast since it already has 135% 2080Ti FE TDP. Given the node jump, it *should* be pretty fast assuming nV wants to at least maintain the efficiency ratio.

 

Bet.

 

I’m really curious now if the 3070 is on a better node (TSMC) or if it has a lower ratio of RT and tensor cores. Otherwise I don’t understand the 100W difference.

 

I think it’s fair to assume GDDR6X is going to account for a significant chunk of the power budget. It would explain the relatively svelte 220W TGP for the 3070.

I wouldn’t bet on AMD smoking the 3080 In pure raster. Best case Big Navi will come with a 384-bit bus at 16Gbps which puts it at exactly the same bandwidth as the 3080. It should make for a very fun matchup though especially if RDNA2 clocks high.

 

3070 is GA104, 3080 is GA102 this time around.