Read more.Quote:
Meanwhile AIB Manli registers GA104-400 GPU and GeForce GTX 2080/2070 names.
Printable View
Read more.Quote:
Meanwhile AIB Manli registers GA104-400 GPU and GeForce GTX 2080/2070 names.
https://i.imgur.com/ioLzQAi.jpg
https://i.imgur.com/zeWjFEF.jpg
The bottom PCB is for a GTX1080TI.
Looking at the size of the GPU pads,this looks smaller than than the GP102 in the GTX1080TI,so is under 471MM2.
I estimate around 400MM2 to 420MM2 in size,just comparing relative sizes(roughly).
Interestingly the GM204 in the GTX980 was around 398MM2.
If the 1180 has the same number of shaders as a 1080ti but on a 12nm process, then that could be about right.Quote:
Originally Posted by CAT-THE-FIFTH
TSMC 12NM isn't any denser than 16NM AFAIK but has improved leakage,so you build bigger GPUs or clock them higher(?),but I suspect moving to a 256 bit memory controller,means less space needed(which is why you are seeing a smaller chip?),and the shaders might be clocked higher.Quote:
Originally Posted by DanceswithUnix
However,its interesting that both the GTX1180 and GTX1080TI seem to use the same power connector arrangement,so it will be interesting to see how power consumption compares to the GTX1080 though!
It was true that AMD didn't get a density improvement on Ryzen, but that was their choice. Wider tends to be better than higher clocked on GPUs, so it wouldn't surprise me if here they kept the clocks and went for more shaders.Quote:
Originally Posted by CAT-THE-FIFTH
Nvidia has gone the other way - every "new" chip has pushed clockspeeds higher and higher. AMD is the one who adds more and more shaders,until they hit the limit of what GCN can apparently handle.Quote:
Originally Posted by DanceswithUnix
Remember,the GP104 chip in the GTX1080 has 10% less transistors than the chip in the Titan Xm,and 20% less shaders. The GTX1080 had no problems beating a Titan Xm. It did it through more capable shader units and more clockspeed.
I mean it is quite possible Nvidia depreciates the shaders even more by making them use less transistors,and then packs more in,but since this is still a 16NM/12NM GPU,and its not like they have a whole ton of extra transistors to chuck at the problem.
Also,look at the power requirements,for a part with more efficient GDDR6,it does seem to need a 6 pin and 8 pin power connector. My GTX1080FE uses one 8 pin one.
I was about to say, AMD seemed to hit a 4096 shader limit ages ago. I hope they have that sorted for 7nm, though if they don't that might help explain why current 7nm parts are AI use only as drivers will be very different.Quote:
Originally Posted by CAT-THE-FIFTH
That's the thing,AMD kept adding more shaders,and Nvidia concentrated on moar clockspeed. AMD ended up hitting issues quicker and then tried to add moar clockspeed which caused problems.Quote:
Originally Posted by DanceswithUnix
Also another thing,look at the Titan V:
https://www.gamersnexus.net/guides/3...r-volta/page-2
The DX12/Vulkan titles enabled it to push ahead of the GTX1080TI,but DX11 showed issues. So like with AMD,who has issues getting good usage of those 4096 shaders in DX11 titles,Nvidia will start to hit similar issues.
Plus,OFC,if Nvidia is pushing RTX,that needs extra functionality,so again I think its improved shaders,etc but clocked higher.
Anyway,we will soon find out who has made the correct predictions!! :p
I think comparing a Titan to a 1080 is problematic too. The shrink from 28nm to 16nm is massive, the die size is half.Quote:
Originally Posted by CAT-THE-FIFTH
Looking at Titan X Maxwell to Pascal, it went wider on shaders as well as smaller die. Girth matters ;)
They dropped transistors from 8 billion on the GM200 to 7.2 billion on the GP104.Quote:
Originally Posted by DanceswithUnix
So 20% less shaders,improved shaders and higher clockspeeds,still made it beat the GM204.
The GP102 has 12 billion transistors,so the 50% increase in shaders came from a 50% increase in transistors.
Since there is no real die shrink to give Nvidia more transistors to add much more shaders,and its a smaller chip than the GP102,it means either Nvidia makes each shader more depreciated,and adds more(think HD3000 series to HD4000 series),or they simply improve IPC,and clock it higher. I think the latter as they need to add RTX in there somewhere! ;) It seems some tech channels are now dropping hints about RTX support now.
Plus you are not considering some other issues - more shaders will need more bandwidth,and any bandwidth improvements from GDDR6 will be negated by a narrower memory bus.
Also,leaks have indicated less shaders:
https://www.techpowerup.com/245657/n...tnamese-stores
Thats says 3584 shaders and less bandwidth.
Remember this too:
https://tpucdn.com/reviews/NVIDIA/Ti..._2560_1440.png
The Titan X(not the 3840 shader Xp) is barely 25% faster than a GTX1080,despite 40% more shaders as it is TDP limited. It can clock higher but it hits power consumption issues.
Its also why the GTX1080TI is so close to the Pascal based Titan X and Titan Xp or even beats them sometimes.
Hence,less shaders than the Xp,with more IPC clocked higher within the same TDP will probably match a GTX1080TI or beat it.
Anyway,probably enough from me in this thread about a GPU I am probably not going to buy! ;)