Rumor: NVIDIA GeForce Ampere to be fabbed at 10nm, all cards RTX ?
Click here to post a comment for Rumor: NVIDIA GeForce Ampere to be fabbed at 10nm, all cards RTX ? on our message forum
Venix
True or not does not really matter , the cards will be judged based on their performance /price/ consumption ... If they are on 7 nm or 10nm .....hell if it delivers will not matter if it is on 1mm 😛, but for the love of god Nvidia i hope the 3060 will get at least 8gb ram .... On the 1060 6gb where fine 2060 questionable but ...ok on a 3060 6gb will not allow the card to age that well
RooiKreef
Not Impossible given the fact that AMD is loading the 7nm nodes of TSMC and Samsung up with the end of the year console and new GPU line launches. But with that said, 10nm sounds odd given the fact that they are already on the 12nm node with a GPU that has 5120 CUDA Cores and 640 Tensor Cores. So why move the node if you already producing a good product of similar kind on a 12nm node?
Astyanax
Why are you owed that information?
Leakers can't leak if you can trace them.
UnrealGaming
THE CURSED 10nm https://cdn.frankerfacez.com/emoticon/214681/1
AMD might have a chance to catch up after all
Netherwind
We've been let down the last few iterations anyway so one more time around wouldn't be a surprise. I expect the usual 20-25%. Let's just hope we don't get another leap in pricing, that last one was sour AF.
nevcairiel
Leakers also can't be trusted if they could be any random troll with no mention of sources or evidence, or even a confirmed history of leaks. (and "same person with new account" is not proof of anything. How do you know thats the same person? Coming back to my original question.)
kings
No.
If AMD, Nvidia and Intel were to reply to all the rumors that come across the internet, they would do nothing else.
slyphnier
delete
asturur
How does 10nm compare to a 1080ti and with a console?
I mean AMD is at 7nm and no AMD card is a current upgrade over a 1080ti, also big navi is still 7nm and could be an upgrade of some sort.
Why people give so much importance to nm?
AMD 8 core CPU at 7nm are as good as intel 8 core at 14nm.
craycray
Intel 8 Core are 'Just as good' at 70-100% more power consumption.
Hyperus
That is not quite correct, INT32 units are not as prevalent or used as it may seem, Nvidias slides were especially misrepresenting in that regard, altough I think that was not their intention.The reason we have them seperated now is that it is a silicon cheap way to increase performance. FP32 is still the most important part about shading, and rays need shading, I actually think that RTX is mainly held back by shading performance, the first leaks by CorgiKitty made more sense, and now he claims that "double the FP32 units does not mean twice the CUDA Cores"
which is exactly what it means. It would be probably more correct to specify wether or not we name Int32 units also CUDA cores as they could be considered as such, but still, generally speaking we mean the amount of FP32 Units in a GPU.
His first leaks suggested that Ampere would double the FP32 Units per SM, a statement which he deleted claiming that that might happen for Hopper and, well, what I wrote above.
Generally speaking, whenever I play BFV with RTX on ultra, I see a 100% 3D GPU Usage in my task manager, suggesting that the RT cores really are not the limit here.
Astyanax
the actual bottleneck in raytracing is the availability of Int processing.
pharma
Could you get the stated performance increase (70-75%) in Indiana's Big Red 200 if Ampere were in fact at 10nm instead of 7nm?
[SPOILER]https://wccftech.com/nvidia-next-gen-ampere-gpu-75-percent-faster-existing-gpus/[/SPOILER]
Hyperus
No, its really not.
To know why, we need to look at what an Integer is.
An integer is a whole number, there is no fraction there, it lacks accuracy to a high degree and is generally not all that used in graphics processing, simply due to accuracy constraints.
I can not imagine a way to just use integers for RT shading the entire time, it just does not make sense.
I am guessing you have seen this https://i1.wp.com/www.techarp.com/wp-content/uploads/2018/09/NVIDIA-RTX-OPS-calculation-01.jpg ?
This would suggest exactly what you wrote about, but it doesn't matter, it has only been represented in that way to signal the fact that Turing can do Int32 in parallel with FP32
Andrew LB
The leather jacket was pretty clear that Ampre would be made using TSMC 7nm.
Astyanax
DXR (and RT cores) are implemented as (fixed) Integer math, not floating point, only the resulting shading of the post calculated scene is float point.
Pascal can't do int and fp in parallel which is where their performance tanks, Volta can which is why it can do RTX without RT cores,
Volta's general purpose Int cores have to handle both BVH and pre-calculation (like pascals) but the concurrent execution permits it without significant overhead.
Hyperus
No, that is simply not correct, you can not just use integer operations for tasks that simply require the accuracy. The performance tanks, yes, but not that bad.
Tensor cores are the other thing that sets Volta apart, we can do vector math in 3D space with matrices.
Integer math is not just a different type of math, it is a whole different type of data type, it can not simply replace floating point data.
BVH is just shooting rays against a less complex set of geometry, usually boxes, as the intersection with every triangle in a scene would have to be tested without it.
Venix
sure if you take out completely the power consumption !
Astyanax
The tensor cores are not used at all in DXR on Volta, and are not exposed to the system outside of CUDA and NVAPI Denoising.
Hyperus
They do not need to be exposed, we are talking about a driver emulation here, NV drivers can do with the hardware whatever they want.
