Possibility of Tensor Cores in Gaming Volta Graphics Cards

[nibir2011]

We already know that the GV100 has the tensor cores which greatly increases the performace in deep learning applications.

And for most part it seems to me that the volta architecture should have some number of tensor cores in each SKUs. However as we already know that the gaming architecture will be called ampere. What is the possiblity that those card will have some tensor cores?

I can not think future smaller Quadro and Tesla cards not having tensor cores. But will nVidia create two completely different architecture and basically different chips for Gaming and Computing? it never happened before. All the compute cards were in fact gaming cards with some features disabled in bios or driver. But to have different die for different product category it seems quite a big jump.

I am not sure myself, personally i am asking this question because in the past most of the nVidia cards has been good enough for deep learning/machine learning. However it seems to me this time either deep learning researchers will get a great push in performance from the tensor cores or they will need deep pockets to get those gains.

What are you thoughts? I know a lot of people here uses nVidia for similar stuff in tomshardware

thanks in advanvce

 

[photonboy]

1) I'd say it's a 0% chance of having Tensor Cores in a gaming GPU (short term at least). Gaming software doesn't have support for it, and the space would be better served with more CUDA cores etc.

If gaming software started to use Tensor Cores in a logical way then we could see them added, but NVidia in particular doesn't tend to waste silicon space for things that may or may not be of benefit in the near future.

2) I also don't know that Ampere is confirmed as the next gaming GPU architecture. AFAIK, Volta is still the next gaming architecture.

3) Volta for gaming (or Ampere) is based on Pascal but has more ASync Compute and some other features.

4) FYI, the V100 (Volta prosumer card) was tested at Gamers Nexus and got 2x the FPS in some situations compared to a GTX1080. It did so at a LOWER FREQUENCY which is interesting because it has exactly 2x the CUDA cores but normally:
a) CUDA vs FPS doesn't scale linearly, and
b) ROPS were lower (proportionately)

I don't want to overanalyze these results, but at a quick glance it appears that ASync Compute may have helped the Volta GPU more fully use it's CUDA CORES.

(The Tensor Cores aren't used in the gaming benchmarks. So again, in at least ONE test at 2560x1440 the V100 with 2x the CUDA CORES got 2x the FPS at a lower frequency. Neat.)

 

[nibir2011]

1) I'd say it's a 0% chance of having Tensor Cores in a gaming GPU (short term at least). Gaming software doesn't have support for it, and the space would be better served with more CUDA cores etc.

If gaming software started to use Tensor Cores in a logical way then we could see them added, but NVidia in particular doesn't tend to waste silicon space for things that may or may not be of benefit in the near future.

2) I also don't know that Ampere is confirmed as the next gaming GPU architecture. AFAIK, Volta is still the next gaming architecture.

3) Volta for gaming (or Ampere) is based on Pascal but has more ASync Compute and some other features.

4) FYI, the V100 (Volta prosumer card) was tested at Gamers Nexus and got 2x the FPS in some situations compared to a GTX1080. It did so at a LOWER FREQUENCY which is interesting because it has exactly 2x the CUDA cores but normally:
a) CUDA vs FPS doesn't scale linearly, and
b) ROPS were lower (proportionately)

I don't want to overanalyze these results, but at a quick glance it appears that ASync Compute may have helped the Volta GPU more fully use it's CUDA CORES.

(The Tensor Cores aren't used in the gaming benchmarks. So again, in at least ONE test at 2560x1440 the V100 with 2x the CUDA CORES got 2x the FPS at a lower frequency. Neat.)

Hmm but their compute cards should have tensor cores right? i so will they make two almost exactly same gpu one( quadro and tesla ) with tensor cores and one( geforce ) without. Thats why the Ampere argument comes in i think there is an announcement that nVidia says Ampere is next architecture for games. any ideas?

 

[photonboy]

Old post and there's a SIGGRAPH announcement later today but I just found out rumors are that we'll see next-gen go roughly like this:

RTX2080
RTX2070
GTX2060
GTX2050
etc.

So it appears like the new top-end cards may have TENSOR CORES for ray-tracing and the lower cards will not hence the new naming being RTX in the future?

All rumors and even if true we don't know if the "GTX2060" is simply a Pascal rebadge or is a Turing-based architecture that simply doesn't have the Tensor cores.

If I was forced to guess I'd say that Turing is the new product stack so the only main difference besides the normal NUMBERS of processing units, memory etc is the lack of TC's on the GTX stack and the inclusion of TC's on the RTX stack.

I think Turing in general would add:
FP16,
improved ASync Compute

With CONSOLES using AMD GPU's now and next-gen (modified NAVI?) NVidia can't deviate too far off from AMD architecture despite their domination of the PC platform.

So long story short I think that the RTX2080 will be Turing and essentially:
1) based on Pascal
2) improve ASync Compute
3) add VEGA features like FP16, and
4) add Tensor Cores, and
5) add VR connector (USB3 Type C I think it was)

 

[renz496]

to be honest i don't think mid range and lower end GPU will not going to have tensor core. DXR is part of DirectX spec. nvidia most likely want all their new lineup from top to bottom to be compliant with DX spec even if those low end cannot handle ray tracing due to lack of raw performance. just like how low end cards like HD5450 still have it's tessellation hardware even if the card does not have the grunt to enable tessellation in games.

nvidia most definitely improving their async compute performance but i'm not really sure about FP16. the usage is pretty limited on modern games and for developer the optimization can also be more complicated due to mixed usage between FP16 and FP32.

other than that we probably going to see big changes to the architecture since maxwell (gaming pascal is pretty much optimized maxwell for higher clock with some added stuff that did not exist in maxwell). one of the talks from board partner is "breakthrough in performance". but it might not be what many people think it of (massive performance jump). instead it might refer to more efficient architecture enabling more performance at lower power than what pascal already capable of. there are rumors floating around that nvidia intend to charge their GPU based on how much power their GPU is rated on. right now Gx104 usually rated around 160w-180w and sold for $500-$600. Gx106 which usually rated at 120w occupy the $300 segment and below. they said nvidia want to move their 120w GPU towards $500 price range. and this is where it gets interesting. some take it that 2060 (that will replace 1080/1070 performance wise) will going to cost $500. meaning after more than two years later 1080 level of performance still going to cost you $500. only with more lower power usage than 1080. or it could be the next x80 card will be roughly 1080ti performance but instead of 180w the card will only consume 120w. meaning the Gx106 chip now being pushed as the new x80 part because of it's performance.