TU104 Big Die

[T4C Fantasy]

GTX 280 has a 576mm2 Die (GT200) 65nm
GTX 285 has a 470mm2 Die (GT200B) 55nm
GTX 480 has a 529mm2 Die (GF100) 40nm
GTX 580 has a 520mm2 Die (GF110) 40nm
GTX 680 has a 294mm2 Die (GK104) 28nm
GTX 780 has a 561mm2 Die (GK110) 28nm
GTX 780 Ti has a 561mm2 Die (GK110B) 28nm
GTX 980 has a 398mm2 Die (GM204) 28nm
GTX 980 Ti has a 601mm2 Die (GM200) 28nm
GTX 1080 has a 314mm2 Die (GP104) 16nm
GTX 1080 Ti has a 471mm2 Die (GP102) 16nm
RTX 2080 has a 553mm2 Die (TU104) 12nm
RTX 2080 Ti has a 754mm2 Die (TU102) 12nm

 

[qubit]

I'd love to see what the TU100 is like.

 

[T4C Fantasy]

I'd love to see what the TU100 is like.

hhahah ikr
https://www.techpowerup.com/gpudb/3224/geforce-rtx-2080
https://www.techpowerup.com/gpudb/3305/geforce-rtx-2080-ti
for full chips
RTX 2070 will be TU106

 

[Vya Domus]

I'd love to see what the TU100 is like.

Should be exactly the same as TU102.

 

[T4C Fantasy]

Should be exactly the same as TU102.

i think it would be bigger for the hbm controller

hbm would still be in demand over gddr6 for servers

 

[dj-electric]

Deep learning and RT units absolutely beef this core up

 

[Vya Domus]

i think it would be bigger for the hbm controller

hbm would still be in demand over gddr6 for servers

Previously, Quadros used exactly the same chips as Teslas, given that the current Quadro series use GDDR6 it's unlikely they'll break the norm now. They'll likely skip HBM with Turing, plus the way Nvidia talked about all this made it sound as if V100 isn't going anywhere. HBM is expensive and still relatively low volume, I don't see them giving Turing HBM.

 

[T4C Fantasy]

Previously, Quadros used exactly the same chips as Teslas, given that the current Quadro series use GDDR6 it's unlikely they'll break the norm now.

Quadro GP100 Tesla P100
Quadro GV100 Tesla V100 all HBM versions of the 102s
100s are always bigger

GP102 471mm2
GP100 610mm2
GV100 815mm2

 

[Vya Domus]

Quadro GV100 Tesla V100 all HBM versions of the 102s

Exactly, what makes you think this time around the Tesla variant of Turing will have HBM ?

 

[T4C Fantasy]

Exactly, what makes you think this time around the Tesla variant of Turing will have HBM ?

because how else are they going to make DGX boards? in that small form factor for servers

 

[Vya Domus]

because how else are they going to make DGX boards? in that small form factor for servers

A Turing based DGX system is again just speculation, what I am saying is Quadros always used the same chips as Teslas. Given that we already know that the new Quadros have GDDR6 logic dictates so will the Teslas, although who is to say there will even be a Turing based Tesla ? Nvidia doesn't have unlimited manufacturing capacity on 12nm, 3 different 700mm^2+ would be a lot, even for them.

 

[T4C Fantasy]

A Turing based DGX system is again just speculation, what I am saying is Quadros always used the same chips as Teslas. Given that we already know that the new Quadros have GDDR6 logic dictates so will the Teslas, although who is to say there will even be a Turing based Tesla ? Nvidia doesn't have unlimited manufacturing capacity on 12nm.

the teslas and hbm products usually come last anyways, we wont know now. but its just likely as nvidia is definitely into the workstation scene, right now GV100 is the biggest die and it lacks ray tracing and ray tracing would be very useful in the rendering scene for teslas and quadros, a render farm of GV100s or TU102s all the money would be with TU102/100 thats why i speculate it.

 

[bug]

I'd love to see what the TU100 is like.

I doubt there's a TU100. TU102 is supposedly already as big as the fab equipment allows.

 

[lavalake]

Yes. The 100 for this series is GV100. It's obvious by now that the CUDA cores in Turing are identical to Volta. TU102 is basically just a smaller GV100 with DP replaced with RT. RT cores have no relevance, while DP is very important to the markets 100 series caters too.

With GV100 maxing out at 800 mm^2, we'll have to wait for 7nm for the next big 100 series chip. AM100, anyone?

 

[qubit]

I doubt there's a TU100. TU102 is supposedly already as big as the fab equipment allows.

I wonder. I'd have thought that if TU102 was the biggest they'd have called it TU100. We'll know for sure in time, it's all speculation for now. I'll bet the TU100, if it exists, will go into some overpriced Titan for 10 grand or something.

 

[bug]

I wonder. I'd have thought that if TU102 was the biggest they'd have called it TU100. We'll know for sure in time, it's all speculation for now. I'll bet the TU100, if it exists, will go into some overpriced Titan for 10 grand or something.

My guess is that either there won't be a TU100 or if they plan one it will be the same as TU102 with less stuff disabled. But that just my guess based on what I've read so far.

 

[qubit]

My guess is that either there won't be a TU100 or if they plan one it will be the same as TU102 with less stuff disabled. But that just my guess based on what I've read so far.

I'm not aware of any disabled bits though. If they are, then yeah, that's probably the difference.

 

[Caring1]

Serious question because I don't know, Why don't they stack chips instead of making them bigger on a single plane all the time?