NVIDIA TU116 GPU Pictured Up Close: Noticeably Smaller than TU106

[btarunr]

Here is the first picture of NVIDIA's 12 nm "TU116" silicon, which powers the upcoming GeForce GTX 1660 Ti graphics card. While the size of the package itself is identical to that of the "TU106" on which the RTX 2060 and RTX 2070 are based; the die of the TU116 is visibly smaller. This is because the chip physically lacks RT cores, and only has two-thirds the number of CUDA cores as the TU106, with 1,536 against the latter's 2,304. The die area, too, is about 2/3rds that of the TU106. The ASIC version of TU116 powering the GTX 1660 Ti is "TU116-400-A1."

VideoCardz scored not just pictures of the ASIC, but also the PCB of an MSI GTX 1660 Ti Ventus graphics card, which reveals something very interesting. The PCB has traces for eight memory chips, across a 256-bit wide memory bus, although only six of them are populated with memory chips, making up 6 GB over a 192-bit bus. The GPU's package substrate, too, is of the same size. It's likely that NVIDIA is using a common substrate, with an identical pin-map between the TU106 and TU116, so AIC partners could reduce PCB development costs.



View at TechPowerUp Main Site

 

[cucker tarlson]

192-bit,not 256-bit.That'd have to be 8GB.

 

[ArbitraryAffection]

Similar core config in terms of CUDA cores, backend and memory bus as GP106, 12nmFFN only slightly improved density (if any?) over 16nmFF, throw in a turing features like seperate INT cores, Tensors, but minus RT cores and i say this chip is aorund 250-300mm2 and 6-7b transistors. I guess someone could work it out? If £250 actually might be an attractive card. I have the really bad itch right now. I love my 570 but. Itch.

Also 590 will need a price cut to £200 to compete I think.

 

[Fabio]

Similar core config in terms of CUDA cores

20% less cuda core is quite a bit

 

[champsilva]

This is almost same size as GP104

Using some photoshop and scale
GP104 = 314 mm²
TU116 = 310 mm²
TU106 = 445 mm²

 

[cucker tarlson]

20% less cuda core is quite a bit

yeah 2060 with 1920 is around 1080,this will compete with 1070.

 

[lexluthermiester]

192-bit,not 256-bit.That'd have to be 8GB.

Based on those pictures though, doing an 8GB/256bit version would be a trivial effort.

 

[dj-electric]

Similar core config in terms of CUDA cores, backend and memory bus as GP106, 12nmFFN only slightly improved density (if any?) over 16nmFF, throw in a turing features like seperate INT cores, Tensors, but minus RT cores and i say this chip is aorund 250-300mm2 and 6-7b transistors. I guess someone could work it out? If £250 actually might be an attractive card. I have the really bad itch right now. I love my 570 but. Itch.

Also 590 will need a price cut to £200 to compete I think.

I do, firmly believe RX 500s need price re-structure.
At least for their 8GB models, would love to see RX 570 at 159$, RX 580 at 189$, RX 590 at 209$. I don't see how otherwise they are going to move sizable amount of units after this.

 

[ArbitraryAffection]

20% less cuda core is quite a bit

16.6% *

Based on those pictures though, doing an 8GB/256bit version would be a trivial effort.

I don't think TU116 has a 256-bit bus. They likely chopped off the memory channels to make the chip smaller, but reused the PCBs from TU106 (2070).

So similar size to GP104 but somewhat significantly lower graphics performance, wow NVIDIA really did bet on those Tensors being useful on this chip. huh.

 

[Darmok N Jalad]

16.6% *

I don't think TU116 has a 256-bit bus. They likely chopped off the memory channels to make the chip smaller, but reused the PCBs from TU106 (2070).

So similar size to GP104 but somewhat significantly lower graphics performance, wow NVIDIA really did bet on those Tensors being useful on this chip. huh.

With GDDR6, it won’t even need 256bit to exceed the bandwidth of the 1070 Ti. Since it seems to be a replacement to the 1070 Ti, there’s not much reason to believe we’ll see anything better unless it’s to increase margins.

 

[cucker tarlson]

With GDDR6, it won’t even need 256bit to exceed the bandwidth of the 1070 Ti. Since it seems to be a replacement to the 1070 Ti, there’s not much reason to believe we’ll see anything better unless it’s to increase margins.

this supposedly has 12gbps ddr6 tho. but still higher than ddr5 256-bit.

 

[Imsochobo]

I do, firmly believe RX 500s need price re-structure.
At least for their 8GB models, would love to see RX 570 at 159$, RX 580 at 189$, RX 590 at 209$. I don't see how otherwise they are going to move sizable amount of units after this.

Here converted as tax and all usually = * 10 so divide for typical US prices(comparable!) for my country (Norway) see RX580 at 2200 kr for 8gb and 2000 for 4gb.
RX570 1829 - Often down at 1500-1600

GTX1050TI cheapest is almost 1800 same as RX570!!
Cheapest GTX1060 3GB is 2200, same as 8gb RX580.

RX 590 is 2800 so that is a bit expensive but next performance level is RTX2060 at 3700.

What the hell is wrong with amd's pricing ?
You want them to sell them with loss, even with GTX1060 6GB performance you pay the same as a high end GTX1050ti... yay, still Nvidia outsells amd for some hilarious reason in taht segment, price is not how amd should compete cause it's proven it doesn't work!

 

[Darmok N Jalad]

Here converted as tax and all usually = * 10 so divide for typical US prices(comparable!) for my country (Norway) see RX580 at 2200 kr for 8gb and 2000 for 4gb.
RX570 1829 - Often down at 1500-1600

GTX1050TI cheapest is almost 1800 same as RX570!!
Cheapest GTX1060 3GB is 2200, same as 8gb RX580.

RX 590 is 2800 so that is a bit expensive but next performance level is RTX2060 at 3700.

What the hell is wrong with amd's pricing ?
You want them to sell them with loss, even with GTX1060 6GB performance you pay the same as a high end GTX1050ti... yay, still Nvidia outsells amd for some hilarious reason in taht segment, price is not how amd should compete cause it's proven it doesn't work!

I went with a used RX 480 instead. Being the Nitro+, it was essentially clocked to match an RX 580 already. I don’t think the 590 is really worth it, but to each their own, I suppose.

 

[ianatikin]

So GTX1660Ti vs RTX2060 has 20% less Cuda cores, around 17% less memory clocks and around 100% less RT cores. Yet it costs 20% less than RTX2060? What does It tell You? For Me it means that even nVidia doesnt believve that RT cores are worth anything.

 

[Turmania]

It is off topic but was wondering will Nvdia follow up a Gtx 1050Ti successor with same 75w power usage ?

 

[Fluffmeister]

Looks like the rx 590 is about to take a beating.

 

[Imsochobo]

Looks like the rx 590 is about to take a beating.

Not if they price it too high as everything below high end for nvidia.
If priced the same RX590 beating commence, power effeciency isn't That important, it's not 20$ worth for people.
But nvidia sells anyways

 

[Miguel2013]

low power usage means they can overclock higher correct? cause you can cool them easier right?

 

[lexluthermiester]

I don't think TU116 has a 256-bit bus. They likely chopped off the memory channels to make the chip smaller, but reused the PCBs from TU106 (2070).

The traces are there, but who knows.

 

[eraser666]

The traces are there, but who knows.

pin to boradLike GM206 use GM204's PCB, GP106 use GP104's PCB

TU116 is an original 192bit GPU.

TU116 can use PG160.PG160 is 2060/2070's PCB.

GTX1660Ti
~1950MHz Board Power 130W
FSE 75xx TS 64xx

 

[londiste]

GP104 = 314 mm²
TU116 = 310 mm²
TU106 = 445 mm²

1536 vs 2304 CUDA cores. ~43% bigger die with 50% more cores.
Looks like RT Cores are very minor in terms of die size.

 

[sam_86314]

1536 vs 2304 CUDA cores. ~43% bigger die with 50% more cores.
Looks like RT Cores are very minor in terms of die size.

Do we know if the 1660 Ti's chip is the full TU116, or a cut-down version? There could be inactive CUDA cores in it. Maybe a future 1770 or something could use a less cut-down version with 1920 or 2048 CUDA cores.

 

[londiste]

Do we know if the 1660 Ti's chip is the full TU116, or a cut-down version? There could be inactive CUDA cores in it. Maybe a future 1770 or something could use a less cut-down version with 1920 or 2048 CUDA cores.

You are right, we don't. 2048 is unlikely though - GP104 has 2560 and Turing CUDA cores are far larger than Pascal's.
When trying to do high-level theorycrafting on SM sizes it seems that Volta SM is ~47% larger than Pascal's. Turing's SM is ~9% bigger than Volta's and ~60% bigger than Pascal.

 

[XiGMAKiD]

It needs to perform at GTX 1070 level performance in most games to worth buying and maybe GTX 1070 Ti in some games

 

[Vayra86]

So similar size to GP104 but somewhat significantly lower graphics performance, wow NVIDIA really did bet on those Tensors being useful on this chip. huh.

Bingo

Its quite an achievement to squeeze lower performance out of the same die space. But no worries, you can use DLSS at 4K to make use of all those cores and achieve nothing substantial with it.