Nvidia Turing Speculation thread [2018]

[JasonLD]

Didn't notice this posted yet: http://www.expreview.com/63600.html

TU102 is apparently being made by both TSMC ("Taiwan", 12nm) and Samsung ("Korea", 14 or 11nm? they don't have 12nm)
Will be interesting to see if there's a difference between the two in die size or performance (overclocking potential)

No, Silicon itself would be fabbed on TSMC with 12nm. Name of country is being printed where final packaging is made.

For example with Vega

Vega is all fabbed on Globalfoundaries but name of the country is different depends on where the final packaging was made. You have to actually look at those codes to find out if they are made on different foundaries but it is very unlikely Turing is fabbed on foundary other than TSMC.

 

[entity279]

The only mining algorithm that really matters is Ethereum - it's 80% of the GPU mining revenue. [...]

That said the entire market has crashed, so mining should have 0 effect on the launch:
https://bitinfocharts.com/comparison/ethereum-mining_profitability.html#log

Both arguments are misleading in the context. While I'd say it's not likely Turing will impact GPU mining significantly, it's certainly not impossible - you're being too dismissive

RT cores are there on the board because they're significantly more efficient than the standard CUDA cores at their given task.
Now, if any given mining algorithm could take advantage of these cores, it will also become more efficient on the Turing architecture.

Hence all the profitability discussions would then go to the bin - you may be only able to mine an obscure coin profitably but who cares ? So long as you'd need say 4 Pascal cards to achieve the same results as 1 Turing card, you'd take most of the revenue , and the profitability will be way higher.

We can take Vega as an example, and how it was distruptive for XMR mining. And it didn't even have special cores.

 

[Kaotik]

No, Silicon itself would be fabbed on TSMC with 12nm. Name of country is being printed where final packaging is made.

For example with Vega

Vega is all fabbed on Globalfoundaries but name of the country is different depends on where the final packaging was made. You have to actually look at those codes to find out if they are made on different foundaries but it is very unlikely Turing is fabbed on foundary other than TSMC.

I've been under the impression if it's printed on the chip it marks where the actual chip was made, rather than assembled?

 

[Deleted member 2197]

Some more Editor day leaks ....
Relative shader performance Turing vs Pascal"

https://videocardz.com/77696/exclusive-nvidia-geforce-rtx-2080-ti-editors-day-leaks

 

[BRiT]

Img doesnt want to inline, but is it the one about shading performance?

The fourth leak for today is showcasing Turing shading performance. This is a relative comparison versus Pascal. If this chart stands true, then each Turing core could be ~50% faster than Pascal core in applications such as 3DMark. This might be very fruitful information for further comparisons.

 

[Deleted member 2197]

Yea, relative shader performance Turing vs Pascal using some games/benchmark apps as examples.

 

[Malo]

Img doesnt want to inline, but is it the one about shading performance?

Images from videocardz you need to remove the s from https to inline them.

 

[A1xLLcqAgt0qc2RyMz0y]

I've been under the impression if it's printed on the chip it marks where the actual chip was made, rather than assembled?

No, the country printed on the package is where it is packaged.

Intel has no foundry in Vietnam yet this Xeon states Vietnam:
https://drive.google.com/open?id=1ca26YMLPN4y3kb98-df_C6Llxk5jdCYL

Intel does have a Test/Assembly site in Vietnam:
https://www.intel.com/content/www/us/en/architecture-and-technology/global-manufacturing.html

 

[McHuj]

According to chiphell. TU-104 is around 530mm (It's not square and I think a little bigger actually). That's a big increase over GP104 (314mm).

https://www.chiphell.com/thread-1902726-1-1.html

 

[Malo]

Well TU104 has 80% of the RT and Tensor core count of TU102? That's seemingly a lot of die area according to the layout Nvidia has showed so far.

 

[Kaotik]

No, the country printed on the package is where it is packaged.

Intel has no foundry in Vietnam yet this Xeon states Vietnam:
https://drive.google.com/open?id=1ca26YMLPN4y3kb98-df_C6Llxk5jdCYL

Intel does have a Test/Assembly site in Vietnam:
https://www.intel.com/content/www/us/en/architecture-and-technology/global-manufacturing.html

That's not printed on the actual chip (die, whatever you're supposed to call it) though

 

[no-X]

According to chiphell. TU-104 is around 530mm (It's not square and I think a little bigger actually). That's a big increase over GP104 (314mm).

https://www.chiphell.com/thread-1902726-1-1.html

If the GDDR6 chips are 14×12 mm and the proportions of GPU and GDDR6s in the Nvidia's image weren't altered, than the GPU could be even bigger, around 560 mm².

 

[giannhs]

Both arguments are misleading in the context. While I'd say it's not likely Turing will impact GPU mining significantly, it's certainly not impossible - you're being too dismissive

RT cores are there on the board because they're significantly more efficient than the standard CUDA cores at their given task.
Now, if any given mining algorithm could take advantage of these cores, it will also become more efficient on the Turing architecture.

Hence all the profitability discussions would then go to the bin - you may be only able to mine an obscure coin profitably but who cares ? So long as you'd need say 4 Pascal cards to achieve the same results as 1 Turing card, you'd take most of the revenue , and the profitability will be way higher.

We can take Vega as an example, and how it was distruptive for XMR mining. And it didn't even have special cores.

rt cores
how exactly do we know that they are actually a new tech for a technology that is older than jesus and not just glorified tensor cores with some limitation and name?

 

[silent_guy]

rt cores
how exactly do we know that they are actually a new tech for a technology that is older than jesus and not just glorified tensor cores with some limitation and name?

You can’t accelerate the actually process of tracing rays with tensor cores. Tracing rays (ray/triangle intersection) is a completely different set of operations than vanilla bulk matrix multiplications (which is what tensor cores do.)

With tensor cores, you can only reduce the amount of rays by filling in the gaps.

If RT cores were tensor cores, one wouldn’t be able to run demos on a single Turing at the same speed as on 4x Volta.

 

[giannhs]

You can’t accelerate the actually process of tracing rays with tensor cores. Tracing rays (ray/triangle intersection) is a completely different set of operations than vanilla bulk matrix multiplications (which is what tensor cores do.)

With tensor cores, you can only reduce the amount of rays by filling in the gaps.

If RT cores were tensor cores, one wouldn’t be able to run demos on a single Turing at the same speed as on 4x Volta.

no offence but that whole "one turing card is faster than a single dxg system" is a complete lie
everyone knows he was reffering to the pica pica demo
but history is a bitch

https://twitter.com/x/status/977225599805669376

the demo was running perfectly on a single card and nvidia insisted to use a dgx for pr and now we know why

 

[Ike Turner]

no offence but that whole "one turing card is faster than a single dxg system" is a complete lie
everyone knows he was reffering to the pica pica demo
but history is a bitch

https://twitter.com/x/status/977225599805669376

the demo was running perfectly on a single card and nvidia insisted to use a dgx for pr and now we know why

Just to be clear: Brian Karis (Epic) was referring to the Star Wars Reflections demo (which also didn't really require 4 Voltas but was definitely more demanding than PICA PICA) While Colin (SEED/EA) was indeed talking about PICA PICA.

 

[CSI PC]

They messed up the FE pricing again like they did with the 1080 so the MSRP prices are pointless, no vendor will ever sell the Ti at $999 when the FE goes for $1199 and that's reflected in the prices. Typical Nvidia scummy practices here. They really like that FE tax and to make their partners look bad because they don't sell at MSRP.

Previous gen I agree but from an AIB perspective the new Turing FE using fans with a full vapor chamber/pipe system could be a nightmare if it was undercutting them in price while also overclocked just as well.
I think partners would be happier either with the FE price being higher to be less competitive or not sold at all (which is not going to happen), especially as it seems they have also improved phase-power design on the reference models that was an area of design-marketing by AIB partners selling Nvidia GPUs historically (will still do so but how much it matters *shrug*).

 

[silent_guy]

no offence but that whole "one turing card is faster than a single dxg system" is a complete lie

There is absolutely no reason to think that HW ray tracing units are less than 5x faster than a traditional SM based solution for ray to triangle intersections.

So, no.

everyone knows he was reffering to the pica pica demo

Claiming that “everyone knows this” doesn’t make it. During the presentation, Pica Pica wasn’t even mentioned.

nvidia insisted to use a dgx for pr and now we know why

In a conspiratorial brain, yes.

Whatever...

It’d be extremely simply to make a ray tracing demo that conclusively proves whether RT cores are real... or not. Because, once again, tensor core math simply doesn’t map to ray to BB and ray to triangle intersection.

Meanwhile, you can see here: https://videocardz.com/77696/exclusive-nvidia-geforce-rtx-2080-ti-editors-day-leaks

RT cores: 72
Tensor cores: 576

But that’s probably all just part of the grand deception as well.

 

[A1xLLcqAgt0qc2RyMz0y]

rt cores
how exactly do we know that they are actually a new tech for a technology that is older than jesus and not just glorified tensor cores with some limitation and name?

72 RT Cores
576 Tensor Cores

 

[A1xLLcqAgt0qc2RyMz0y]

no offence but that whole "one turing card is faster than a single dxg system" is a complete lie
everyone knows he was reffering to the pica pica demo
but history is a bitch

https://twitter.com/x/status/977225599805669376

the demo was running perfectly on a single card and nvidia insisted to use a dgx for pr and now we know why

This clearly shows that the Ray Tracing is using a separate part of the Turing to do the Ray Tracing (Green) and that the Tensor Cores (Purple) are on a different part of the chip.

This disproves you conspiracy theory. 4X Volta's took 55 ms to render a Ray traced frame whereas one RTX Turing only took 45 ms. That means that a single RTX Turing is 4.89 times faster than a single Volta.

The Pascal above is a GTX 1080 Ti and is 6.8 times slower than a RTX Turing.