Nvidia Turing Speculation thread [2018]

[Voxilla]

What about Tensor cores for this next generation?

It could be usefull for inferencing images and voice.

Voice for example by teacher apprentice learning from Wavenet like NN.

 

[Alexko]

I was under the impression that it was meant as a gaming architecture, so Tensor cores seem unlikely. Also, it should be Ampère/Ampere, not Ampre... ;-)

 

[Voxilla]

I was under the impression that it was meant as a gaming architecture, so Tensor cores seem unlikely. Also, it should be Ampère/Ampere, not Ampre... ;-)

Sometimes, you need to be able to think outside of the box.

 

[CSI PC]

I was under the impression that it was meant as a gaming architecture, so Tensor cores seem unlikely. Also, it should be Ampère/Ampere, not Ampre... ;-)

Optix-RTX is also part of Gameworks though, combined with some kind of version being implemented within latest Metro game.
Seems a lot of effort just for Titan V, from perspective of both framework and a studio if it is meant to be just professional segment technology.
https://developer.nvidia.com/gameworks-ray-tracing
Near bottom of the page shows its context, which were part of the demonstation with Volta-RTX-and post processing Tensor:

The upcoming GameWorks SDK — which will support Volta and future generation GPU architectures — enable ray-traced area shadows, ray-traced glossy reflections and ray-traced ambient occlusion.

https://developer.nvidia.com/ray-tracing-games-nvidia-rtx-pdf

Not suggesting it would find its way trickle down to most models, but who knows how far Nvidia could implement this.

 

[Deleted member 2197]

Nvidia Turing: GTX 1180 specs leaked, plus everything we know so far

According to information obtained by completely unknown means by Wccftech, the Nvidia GeForce GTX 1180 may well bear a closer resemblance to today’s GTX 1080Ti than the GTX 1080. Powered by a GPU currently known as the GT104, the GTX 1180 will allegedly have a whopping 3584 CUDA cores, a 256-bit GDDR6 memory interface and between 8-16GB of GDDR6 memory, clocked at 16Gbps.

If true, this would mean the GTX 1180’s memory is significantly faster than either the 8GB of GDDR5X memory currently in the GTX 1080 or any known form of HBM2 memory (2nd Gen high-bandwidth memory). This in turn would give the GTX 1180 a huge memory bandwidth of 512GB/s. For comparison’s sake, the GTX 1080 only offered 320GB/s, while the GTX 1080Ti provides 484GB/s.
....
The graphics card itself, on the other hand, would be clocked somewhere in the region of 1.6GHz, with a max boost clock of around 1.8GHz. The thermal design power (TDP), meanwhile, is expected to be somewhere between 170-200W.
....
Those very same leaked specs also contain a potential price for the GTX 1180 as well: a rather eye-watering $699. That’s $100 more than the GTX 1080’s launch price, but given current graphics card prices (even if they are, in fact, starting to come down now), that may well actually be a good thing rather than bad.

https://www.rockpapershotgun.com/20...phics-cards-specs-release-date-price-rumours/

 

[Voxilla]

So it's Ampere AKA Turing

 

[Voxilla]

Ampere might be a bit too much publicity for:
https://amperecomputing.com/products-overview/

 

[Voxilla]

Just changed the thread title.

 

[Voxilla]

So it will be faster as Titan Xp (as has become usual with NV)
https://wccftech.com/nvidia-gtx-1180-custom-cards-by-august-september/
(first release with a 256-bit GDDR6 bus.)

So Titan XT will presumably be 384-bit GDDR6 bus.

 

[Deleted member 13524]

If true, this would mean the GTX 1180’s memory is significantly faster than either the 8GB of GDDR5X memory currently in the GTX 1080 or any known form of HBM2 memory (2nd Gen high-bandwidth memory). This in turn would give the GTX 1180 a huge memory bandwidth of 512GB/s. For comparison’s sake, the GTX 1080 only offered 320GB/s, while the GTX 1080Ti provides 484GB/s.

You can tell rock-paper-shotgun isn't remotely decent for hardware commentary the moment they come up with a statement like this.
Even the Titan V with a cut-down memory sub-system goes up to 652GB/s, and AFAIK the widest HBM2 implementation is on NEC's Tsubasa with 6 stacks and 1.2TB/s.

 

[Voxilla]

You can tell rock-paper-shotgun isn't remotely decent for hardware commentary the moment they come up with a statement like this.
Even the Titan V with a cut-down memory sub-system goes up to 652GB/s, and AFAIK the widest HBM2 implementation is on NEC's Tsubasa with 6 stacks and 1.2TB/s.

Right, but Titan XT still will be faster as Titan V (excluding double precision, question mark for tensor cores).
Also memory bandwith with 384 GDDR6 bus will likely be higher as from 3 stacks HBM2.

 

[Deleted member 13524]

Right, but Titan XT still will be faster as Titan V (excluding double precision, question mark for tensor cores).
Also memory bandwith with 384 GDDR6 bus will likely be higher as from 3 stacks HBM2.

I won't question that the next Titan will be faster in games than the previous, nor the bandwidth provided by the clock values of GDDR6 already announced by Samsung.

 

[Voxilla]

Funny there is speculation about Turing in the Volta thread

 

[Kaotik]

Funny there is speculation about Turing in the Volta thread

What's now being speculated as "Turing" was first speculated as "Volta" because there was no reason to believe NVIDIA wouldn't bring gaming versions of Volta until the relatively recent rumors of Ampere and later Turing (reportedly same thing) - in fact it's still possible that the whole Ampere/Turing is bogus rumors and it's really Volta, or it could be just Volta renamed for whatever reason with some tweaks, likely cutting Tensors etc

edit:
And to add to this, Volta is still the newest architecture NVIDIA has mentioned on their roadmaps

 

[Voxilla]

What's now being speculated as "Turing" was first speculated as "Volta" because there was no reason to believe NVIDIA wouldn't bring gaming versions of Volta until the relatively recent rumors of Ampere and later Turing (reportedly same thing) - in fact it's still possible that the whole Ampere/Turing is bogus rumors and it's really Volta, or it could be just Volta renamed for whatever reason with some tweaks, likely cutting Tensors etc

edit:
And to add to this, Volta is still the newest architecture NVIDIA has mentioned on their roadmaps

There are no gaming versions of Volta, there is a reason Titan Volta is called Titan V and not Titan XV.
Volta is a non gaming architecture HBM2 (gaming will use GDDR6), double precision, NV link, high tensor core mixed precision TFLOP/s (for AI/ML training) Volta can be used for gaming though but from a price/performance point of view it is not optimal.

 

[Alexko]

There are no gaming versions of Volta, there is a reason Titan Volta is called Titan V and not Titan XV.
Volta is a non gaming architecture HBM2 (gaming will use GDDR6), double precision, NV link, high tensor core mixed precision TFLOP/s (for AI/ML training) Volta can be used for gaming though but from a price/performance point of view it is not optimal.

Remove all these things from Volta and it's a pretty great gaming architecture. Is this what Turing/Ampère is? If so, does it warrant a new name? Does that make it a new architecture? I fear we may be venturing into semantics, here.

 

[Deleted member 13524]

Remove all these things from Volta and it's a pretty great gaming architecture. Is this what Turing/Ampère is? If so, does it warrant a new name? Does that make it a new architecture? I fear we may be venturing into semantics, here.

Well, AMD just made the architecture naming thing a whole lot looser with Vega M...

 

[Kaotik]

Well, AMD just made the architecture naming thing a whole lot looser with Vega M...

Technically AMD never said "Vega M" uses "Vega-architecture" though

There are no gaming versions of Volta, there is a reason Titan Volta is called Titan V and not Titan XV.
Volta is a non gaming architecture HBM2 (gaming will use GDDR6), double precision, NV link, high tensor core mixed precision TFLOP/s (for AI/ML training) Volta can be used for gaming though but from a price/performance point of view it is not optimal.

But are any of those "architecture defining features"? Previously we've had chips based on same architecture with and without fast double precision, with several different memory standards (including HBM2) and so on.

 

[Deleted member 13524]

Technically AMD never said "Vega M" uses "Vega-architecture" though

(¬_¬)

 

[Picao84]

(¬_¬)

One more reason why using Architecture name as series name is a terrible idea. NVIDIA has used in the past chips from previous architectures under the same series, albeit the very low end ones.