The RTX 3080/3090 have 80.7% greater transistor density than the RTX 2080 Ti

[g67575]

If you are wondering how Nvidia has increased the shader count so much, I believe the 80.7% increase in transistor density is one of the main reasons.

sources:
https://www.techpowerup.com/gpu-specs/nvidia-ga102.g930
https://www.techpowerup.com/gpu-specs/nvidia-tu102.g813
https://www.techpowerup.com/gpu-specs/nvidia-ga100.g931

The die shrink from TSMC 12nm to Samsung 8nm wouldve helped to increase transistor density, but if you look at the Ampere A100 gpu with 7nm TSMC transistors, it looks like Nvidia couldve gone even further to increase transistor density, but chose not to maybe due to high costs or perhaps, low production yields on TSMC 7nm.

The ampere A100 has a transistor density that is 165.5% greater than the RTX 2080 TI! I think this shows that Nvidia could produce even higher spec 7nm consumer Ampere GPUs in the future, but maybe they won't need to since AMD is still catching up.

 

[g67575]

I've just worked out something else thats quite interesting. The die size of the 7nm TSMC Xbox series X gpu is approx 171mm mm² (excluding other parts of the SOC). If you divide the 15.3 billion transistors it has by the die size (171), you get a transistor density of 89.4 mm²! This is similar to the transistor density of 91.2 mm² already achieved by TSMC's 7nm lithography.

That would be more than double the transistor density of the Navi 10 based 5700 XT (from 41 mm², thats a 118% increase).

So, I think a doubling of transistors to 21 billion (or more) for Navi 2X is likely, we know AMD has managed 15.3 billion transistors already on RDNA2, with around 200w TDP.

Since Nvidia managed to more than double the shader unit count of the RTX 2080 TI with 'just' a 50.5% increase in transistor count, I dont see a reason why AMD couldnt get similar results with an even larger increase in transistors.

Assuming a doubling of shader units, a 19-20 TFlop count should be doable for the top RDNA 2 GPU, similar to the RTX 3070.

Increasing the transistor count by more than 50% could result in more than double the shader unit count of RDNAv1, I guess its unknowable at this point


This theory contradicts the speculative Navi 21 specs listed here, which instead suggests a very large GPU die size of 505 mm², and similar transistor density (41.6M / mm²) to RDNAv1, for the top RDNA 2 GPU.
Link here:
https://www.techpowerup.com/gpu-specs/amd-navi-21.g923

 

[labbby]

If that was the case wouldn’t AMD start dropping random leaks before ampere goes on sale showing exactly that ?

sure, they’ve got a few weeks to do this but unless they start dropping some performance hints a lot of people will just jump on the ampere train

 

[Grim5]

If you are wondering how Nvidia has increased the shader count so much, I believe this is one of the main reasons.

souces:
https://www.techpowerup.com/gpu-specs/nvidia-ga102.g930
https://www.techpowerup.com/gpu-specs/nvidia-tu102.g813
https://www.techpowerup.com/gpu-specs/nvidia-ga100.g931

The die shrink from TSMC 12nm to Samsung 8nm wouldve helped to increase transistor density, but if you look at the Ampere A100 gpu with 7nm TSMC transistors, it looks like Nvidia couldve gone even further to increase transistor density, but chose not to maybe due to high costs or perhaps, low production yields on TSMC 7nm.

The ampere A100 has a transistor density that is 165.5% greater than the RTX 2080 TI! I think this shows that Nvidia could produce even higher spec 7nm consumer Ampere GPUs in the future, but maybe they won't need to since AMD is still catching up. Maybe a 7nm based Titan?

What happened dude? You were predicting quite strongly rtx3000 would only get to 18Tflop and instead we got 36Tflop

 

[g67575]

what happened is, Nvidia increased the transistor count by 50.5% for the 3080/3090, resulting in a massive performance increase! This performance increase was not reflected with the Ampere A100 (which has 54.2 billion transistors), which is what I based my prediction on

die shrinks matter, I dont think AMD has utilized the benefits of 7nm yet with RDNAv1 GPUs (AMD engineers said there were power efficiency problems with RDNAv1). It looks like the 7nm transistor density is much higher with RDNAv2

 

[g67575]

Ofc, talking about transistor densities improvements just isnt very 'sexy', AMD / NVIDIA marketing teams prefer to talk about easier to communicate general performance stats like TF count.

 

[g67575]

Is anyone gonna upgrade from a 5700 XT to the top RDNA 2 GPU?

 

[nutsnut]

I'm waiting to see what RDNA2 looks like and actual bench marks for the RTX3080.

Then i'll decide whether i'll stick with the 5700XT which is still a very capable 1440p card or upgrade.

 

[willhub]

I've just worked out something else thats quite interesting. The die size of the 7nm TSMC Xbox series X gpu is approx 171mm mm² (excluding other parts of the SOC). If you divide the 15.3 billion transitors it has by the die size (171), you get a transistor density of 89.4 mm²! This is similar to the transistor density of 91.2 mm² already achieved by TMSC's 7nm lithography.

That would be more than double the transistor density of the Navi 10 based 5700 XT (from 41 mm², thats a 118% increase).

So, I think a doubling of transistors to 21 billion (or more) for Navi 2X is likely, we know AMD has managed 15.3 billion transistors already on RDNA2, with around 200w TDP.

Since Nvidia managed to more than double the shader unit count of the RTX 2080 TI with 'just' a 50.5% increase in transistor count, I dont see a reason why AMD couldnt get similar results with an even larger increase in transistors.

Assuming a doubling of shader units, a 19-20 TFlop count should be doable for the top RDNA 2 GPU, similar to the RTX 3070.

Increasing the transistor count by more than 50% could result in more than double the shader unit count of RDNAv1, I guess its unknowable at this point


This theory contradicts the speculative Navi 21 specs listed here, which instead suggests a very large GPU die size of 505 mm², and similar transistor density (41.6M / mm²) to RDNAv1, for the top RDNA 2 GPU.
Link here:
https://www.techpowerup.com/gpu-specs/amd-navi-21.g923

It's a gamble, I'm going to get an RTX3080 as I want my desktop up and running as it's GPUless currently, but I'll be holding onto the edge of my seat when Navi benchmarks are release hoping it's nothing I regret.

 

[Kaapstad]

Can this go in the main ampere thread.

 

[five8five]

There's a rumour that the biggest RDNA2 card has been having heat issues and could be delayed. AMD don't have the best track record scaling GPUs up so there could be some substance to it

 

[g67575]

Nah, if they can release the console RDNA GPUs in November 2020, with average / good cooling in a compact space at sufficiently cool temps, I dont think excessive heat will prevent a desktop GPU launch, which should have higher end cooling (AMD stated they will use a 2/3 fan design on reference RDNA 2 graphics cards).