NVIDIA ‘Ampere’ 8nm Graphics Cards

[4K8KW10]

https://videocardz.com/newz/nvidia-registers-hopper-and-aerial-trademarks

"a Twitter user called kopite7kimi shared information that Hopper is NVIDIA’s upcoming MCM architecture"

The same user said:

"GA101 cancelled|"

So, what's the current GPU roadmap going to look like?

Pascal|Volta -> Turing -> Ampere* (??) -> Hopper

Unless they wanted an absolute killer of a generation they won't go that big. There's no need for them to even if it is within the realms of possibility.

That card is for the data centre business. If they go smaller, that means the relative performance improvement over the 815 mm^2 GV100 would be either small or very small or none.

 

[Kaapstad]

GV100 is 815 mm^2 in the Tesla V100 PCIe 32 GB.
800 mm^2 would be a normal die size for Tesla A100...

The second quote sounds quasi accurate, the facts point to it being not false because:
- nvidia just recently announced Tesla V100S https://wccftech.com/nvidia-tesla-v100s-server-gpu-16-tflops-over-1tb-hbm2/
- we expect RTX 2080 Ti Super
- there is absolutely zero credible evidence that Ampere exists. Normally, there are benchmark leaks of unknown, unrecognized new GPUs, now there are none.

Only forecasts by some analysts using data based on previous launches and the periods between them. Meh.

The was no creditable evidence that the Titan V existed before launch and its arrival took all the clickbait sites totally by surprise.

 

[KillBoY_UK]

The was no creditable evidence that the Titan V existed before launch and its arrival took all the clickbait sites totally by surprise.

back in the day I remember when the press was caught off guard by the old Pentium 4 Extreme Editions in a similar manner

 

[DELETED_195422]

https://www.tomshardware.com/news/n...ter-than-turing-at-half-the-power-consumption

Lets hope they get their **** together on pricing...

 

[scubes]

https://www.tomshardware.com/news/n...ter-than-turing-at-half-the-power-consumption

Lets hope they get their **** together on pricing...

it will be the first 2 grand gpu lol

 

[robfosters]

2nd half of 2020 though

 

[robfosters]

it will be the first 2 grand gpu lol

Unlikely with the new gen consoles coming out. I think they'll try to keep it reasonable.

 

[Grim5]

Unlikely with the new gen consoles coming out. I think they'll try to keep it reasonable.

Agreed. High GPU prices in 2020 will drive millions of PC gamers to buy a Playstation/Xbox instead.

 

[DELETED_195422]

it will be the first 2 grand gpu lol

Wouldn't surprise me. Worse than that, people would pay it.

Price it decently and everyone will be like...

 

[Spook187]

I'm not playing cyberpunk 2077 until the 3080ti is out, I really hope it's not the second half of this year

 

[shamus21]

You will not get both a 50% performance boost and power reductions

 

[Grim5]

You will not get both a 50% performance boost and power reductions

might be a misquote or translating issue. I belive what they were trying to say is that Ampere has double the Performance Per Watt of Turing.

 

[Panos]

You will not get both a 50% performance boost and power reductions

I will be surprised to see how also, because physically neither the 2080 nor the 2080ti can be shrank to 7nm. Also NV said RT performance will get 50% perf not raster perf. And that's easy if rt and tensor cores work at higher clock speeds, since now are way slower than the normal cores.

 

[D.P.]

You will not get both a 50% performance boost and power reductions

Depends a bit on the exact node specifics they are using but Nvidia is skipping 7nm DUV for 7nm ++ EUV which is a surpsignly big jump given the naming, somewhat equivalent to the 12/14/16nm down to 7nm DUV.

My guess would be about a 50% performance increase in rasterization, 80-150% increasing ray-tracing, and power might drop 5%, e.g. 10-15w.

 

[Panos]

Depends a bit on the exact node specifics they are using but Nvidia is skipping 7nm DUV for 7nm ++ EUV which is a surpsignly big jump given the naming, somewhat equivalent to the 12/14/16nm down to 7nm DUV.

My guess would be about a 50% performance increase in rasterization, 80-150% increasing ray-tracing, and power might drop 5%, e.g. 10-15w.

Some times I feel this forum is WCCFTECH and logic doesn't exist. Explain to me HOW NV can achieve that?
To do so needs to make a chip bigger than current Turing, which is physically impossible to manufacture at 7nm already. (over 520mm2). And even if NV manage to squeeze it magically to 429mm2 (maximum size TSMC & Samsung can make 7nm chips) the chip will have very low yields.

AMD with the Navi has space to grow, given the chip is 251mm2 at 7nm. Yet still expect any bigger chips to be expensive because yields will be low. Hell AMD had to cut it's bus to make the chip small enough to be profitable at it's MSRP, which everyone deemed too high....

 

[Grim5]

Depends a bit on the exact node specifics they are using but Nvidia is skipping 7nm DUV for 7nm ++ EUV which is a surpsignly big jump given the naming, somewhat equivalent to the 12/14/16nm down to 7nm DUV.

My guess would be about a 50% performance increase in rasterization, 80-150% increasing ray-tracing, and power might drop 5%, e.g. 10-15w.

I think people are getting too caught up in the numbers.

Some people are saying 50% perf and 50% less power draw is impossible for 12nm to 7nm EUV. The quote from Taiwan said these figures are for Ampere - as in they already include IPC changes. So part of the number is from 12nm to 7nm EUV amd some will be from architecture changes. AMD got 25% more IPC from its latest architecture why can't Nvidia do the same

 

[Grim5]

Some times I feel this forum is WCCFTECH and logic doesn't exist. Explain to me HOW NV can achieve that?
To do so needs to make a chip bigger than current Turing, which is physically impossible to manufacture at 7nm already. (over 520mm2). And even if NV manage to squeeze it magically to 429mm2 (maximum size TSMC & Samsung can make 7nm chips) the chip will have very low yields.

AMD with the Navi has space to grow, given the chip is 251mm2 at 7nm. Yet still expect any bigger chips to be expensive because yields will be low. Hell AMD had to cut it's bus to make the chip small enough to be profitable at it's MSRP, which everyone deemed too high....

AMD 5900xt is 503mm2 and presumably built by TSMC on 7nm EUV

 

[Panos]

AMD 5900xt is 503mm2 and presumably built by TSMC on 7nm EUV

5900XT is just speculation without facts. In my argument we have FACTS from existing products and what TSMC has on spec sheets can make.

 

[4K8KW10]

5900XT is just speculation without facts.

5900 XT as a name is a speculation but Navi 21 being a 505 mm^2 is a leak, reported by many, actually everyone talks about it now:

https://hothardware.com/news/amd-radeon-rx-navi-21-gddr6-rdna-2
https://www.digitaltrends.com/computing/amd-navi-21-twice-fast-5700-xt/
https://wccftech.com/amd-radeon-rx-navi-21-gpu-2x-performance-5700-xt-die-size-rumor/
https://www.ptt.cc/bbs/PC_Shopping/M.1573129117.A.216.html
https://www.reddit.com/r/Amd/commen...hend/?utm_medium=android_app&utm_source=share

Depends a bit on the exact node specifics they are using but Nvidia is skipping 7nm DUV for 7nm ++ EUV which is a surpsignly big jump given the naming, somewhat equivalent to the 12/14/16nm down to 7nm DUV.

My guess would be about a 50% performance increase in rasterization, 80-150% increasing ray-tracing, and power might drop 5%, e.g. 10-15w.

Yup, and the more they invest in RTX speed, the less raster speed increase they will get.

That would put big Ampere on par with big Navi. If the former is 50% faster than RTX 2080 Ti, and the latter is 80-100% faster than the RX 5700 XT.

You will not get both a 50% performance boost and power reductions

might be a misquote or translating issue. I belive what they were trying to say is that Ampere has double the Performance Per Watt of Turing.

Yeah, it is either 50% more speed or the same speed in 50% of the TDP, not both.

 

[JediFragger]

/deleted