gtx 1180 leaks and infofmation from wccftech

[BasicallyAMod]

what does the size of the chip matter. it says that this is 12nm and i heard AMD navi is going to be 7nm. why does that matter?

[Humbug]

A GTX 1180 performing slightly ahead of the previous gen flagship GTX 1080ti is pretty much in line with Nvidia's usual pace of improvement correct?

In which case I guess the story is believable.

[Humbug]

10 minutes ago, BasicallyAMod said:

what does the size of the chip matter. it says that this is 12nm and i heard AMD navi is going to be 7nm. why does that matter?

This refers not to the size of the chip.

Rather to the manufacturing process used in creating the stuff on the chip. The smaller it is in general the more resources you can stuff in there, the more power efficient it is and the better it clocks. This applies all other factors being equal. So for example the architecture and the maturity of the process node etc also affects how well things clock or how efficient they are. So currently Nvidia's designs are more power efficient than AMD's because of the architecture.

 

Yes AMD's Navi will be 7nm but that will not be ready for another year. According to this article the GTX 1180 is coming this year.

[Dylanc1500]

9 minutes ago, BasicallyAMod said:

what does the size of the chip matter. it says that this is 12nm and i heard AMD navi is going to be 7nm. why does that matter?

Node "sizes" aren't as they are labeled.

[BasicallyAMod]

so basically its being able to fit more things in the chip?

 

that makes sense.

[michaelocarroll007]

1 hour ago, SC2Mitch said:

wccftrash so take it with a dumpster of salt 

Eh i mean its pretty believable at the guess of 13tflops.  The 1080 is 9tflops 45% improvement in 2 years and they already have the newest titian at 12tflops so it comes in just a Tad higher then this. Id honestly hope it hits 15tflops as 13 is pretty lack luster even more so when you consider since pricing is always going up each launch

[Humbug]

1 minute ago, michaelocarroll007 said:

Eh i mean its pretty believable at the guess of 13tflops.  

Yep, very believable. Given Nvidia's track record of generational improvements...

 

I guess launching this year is the only part which sounds doubtful.

[laminutederire]

4 minutes ago, Humbug said:

Yep, very believable. Given Nvidia's track record of generational improvements...

 

I guess launching this year is the only part which sounds doubtful.

It could be insightful though of what's next.

[Brooksie359]

43 minutes ago, Camoxide said:

"GT104" 

 

Wouldn't it be GV104? So they're either rebranding the Volta architecture or this is a separate architecture. 

 

They've previously used "GT" for chips that were based on the Tesla architecture. 

I mean that was a long time ago so it probably doesn't matter tbh.

[BasicallyAMod]

i hope they change it or it'll be confusing, but their titan cards have been confusing before soooo.....

[Coaxialgamer]

This actually makes a LOT of sense if next gen is not much more than a pascal refresh . Given the density and power improvements brought from TSMC 12nm vs the older node + a smaller IMC , ~400mm² for a 3584 Cuda core gpu seems right in line considering gp 102's 472mm² .

[Coaxialgamer]

34 minutes ago, BasicallyAMod said:

what does the size of the chip matter. it says that this is 12nm and i heard AMD navi is going to be 7nm. why does that matter?

7nm and 12nm are just technology nodes , roughly telling of how small the individual transistors in the chip will be  , and thus how many transistors/performance you can fit in a given amount of space.

The chip die size tells you how large the actual chip size is , in the case roughly 400mm² , which is fairly typical for a gpu . In theory , the true cost of a chip is determined with the die size , and stays the same regardless of the technology node ; ie the cost per mm² is the same .In practice , the cost of area has been increasing in the last few years , and a 7nm , 400mm² chip should cost more than an equivalent 12nm one . But that should be offset by the increased integration provided by the more advanced node .

also , it's worth noting that TSMC's 12nm is in production now . It's not even their most advanced tech , as they are also producing 10nm chips and are working on 7nm . There are currently no production-ready 7nm process nodes , and they won't be available until well within H2 , at the earliest . 7NM will be at the very bleeding edge , and as such presents a much riskier proposition to AMD , as well as a greater reward

14 minutes ago, michaelocarroll007 said:

Eh i mean its pretty believable at the guess of 13tflops.  The 1080 is 9tflops 45% improvement in 2 years and they already have the newest titian at 12tflops so it comes in just a Tad higher then this. Id honestly hope it hits 15tflops as 13 is pretty lack luster even more so when you consider since pricing is always going up each launch

TFLOPS basically mean nothing . There is a reason it isn't compared across architectures , and it's because the actual calculation doesn't take it into account . I can tell you exactly how many flops a 3ghz nvidia turingator gpu with 9600 cuda cores will put out ( 57.6TFLOPS ) , and that's because the formula is :

clocks * ALU_count(cores) * 2 = peak floating point throughput

[laminutederire]

18 minutes ago, Coaxialgamer said:

This actually makes a LOT of sense if next gen is not much more than a pascal refresh . Given the density and power improvements brought from TSMC 12nm vs the older node + a smaller IMC , ~400mm² for a 3584 Cuda core gpu seems right in line considering gp 102's 472mm² .

What's worrying for performance enthusiasts is that we won't really get any benefit from it at the announced price except better perf /w. It's a 1080ti at approximately the same price it's supposed to be at around that time with the same perf, and a lower tdp.

But it also shows that Nvidia do not seem to have anything new architecturally speaking. Which means that and only has Pascal to beat for the foreseeable future, so we might get slightly better than what we already have for the next 2-3 years.

[Razor01]

23 minutes ago, Coaxialgamer said:

This actually makes a LOT of sense if next gen is not much more than a pascal refresh . Given the density and power improvements brought from TSMC 12nm vs the older node + a smaller IMC , ~400mm² for a 3584 Cuda core gpu seems right in line considering gp 102's 472mm² .

 

The architecture is different its not a pascal refresh.  This is something we do know, nV stated Volta's architecture will be in the cards that come after Volta.

[Coaxialgamer]

Just now, laminutederire said:

What's worrying for performance enthusiasts is that we won't really get any benefit from it at the announced price except better perf /w. It's a 1080ti at approximately the same price it's supposed to be at around that time with the same perf, and a lower tdp.

But it also shows that Nvidia do not seem to have anything new architecturally speaking. Which means that and only has Pascal to beat for the foreseeable future, so we might get slightly better than what we already have for the next 2-3 years.

We might see architectural improvements , we might not . I was just reflecting that it would make sense given what we already know . It all depends on how Nvidia plans to exploit AMD's un-competitiveness for the time being . Although the upside is that it could leave nvidia open to a zen-like blow i, just like how intel was thoroughly unprepared on the cpu front.

 

that being said ; if this chip is actually what we get , it is actually fairly small on a now-mature process ( given that 12nm is just an improved version of their existing tech , just like GF's 12nm ). We don't know what clocks the chip will reach at max , but both size and power have plenty of wiggle room . We could easily see a traditional titan or 1180ti class chip pushing 600mm² with a 250-300w TDP like usual . The extra cores would improve performance significantly .

[Valentyn]

1 hour ago, Humbug said:

Yep, very believable. Given Nvidia's track record of generational improvements...

 

I guess launching this year is the only part which sounds doubtful.

13 TFLOPs would be a higher expectation; given the Titan V is at 14.9TFLOPs.

 

Now if they’re on 7nm and not 12nm I would expect that performance  at the supposed TDP

[Taf the Ghost]

25 minutes ago, Coaxialgamer said:

We might see architectural improvements , we might not . I was just reflecting that it would make sense given what we already know . It all depends on how Nvidia plans to exploit AMD's un-competitiveness for the time being . Although the upside is that it could leave nvidia open to a zen-like blow i, just like how intel was thoroughly unprepared on the cpu front.

 

that being said ; if this chip is actually what we get , it is actually fairly small on a now-mature process ( given that 12nm is just an improved version of their existing tech , just like GF's 12nm ). We don't know what clocks the chip will reach at max , but both size and power have plenty of wiggle room . We could easily see a traditional titan or 1180ti class chip pushing 600mm² with a 250-300w TDP like usual . The extra cores would improve performance significantly .

There will be a G*-102 part because of the need for Quadro cards, and they can sell enough of them to gamers to make up the development costs anyway. However, it strikes me the reason Turing is more than a respin of Pascal is that just adding more cuda cores doesn't produce linear improvements. They have to design the larger dies to deal with the new set of bottlenecks.

 

As for AMD, they're rushing to 7nm, roughly before everyone else. They're going to update & shrink Vega (which is what Navi will mostly be) with GDDR6 and replace the RX500 line with roughly similar performance to the current Nvidia offerings, at a smaller die size. This is important as we might actually see a fairly short generation for Nvidia. They could have their first 7nm part out next year pretty easily, which might just be the 11xx Generation Titan (Titan T?). 2019 is going to be a busy tech year as everyone drops important new nodes.

[laminutederire]

44 minutes ago, Coaxialgamer said:

We might see architectural improvements , we might not . I was just reflecting that it would make sense given what we already know . It all depends on how Nvidia plans to exploit AMD's un-competitiveness for the time being . Although the upside is that it could leave nvidia open to a zen-like blow i, just like how intel was thoroughly unprepared on the cpu front.

 

that being said ; if this chip is actually what we get , it is actually fairly small on a now-mature process ( given that 12nm is just an improved version of their existing tech , just like GF's 12nm ). We don't know what clocks the chip will reach at max , but both size and power have plenty of wiggle room . We could easily see a traditional titan or 1180ti class chip pushing 600mm² with a 250-300w TDP like usual . The extra cores would improve performance significantly .

I am not sure it's that intentional of them. They'd be more than happy to bury AMD and have a free reign until 2020 at least when Intel will come into play again. If they don't it's probably that they can't.

[Dionyz]

July??? They are not going to release it anytime soon. You have to understand business. If they are selling out of their current cards then why would they release a new better product? It would be retarded for Nvidia to release a product anytime soon.

[BasicallyAMod]

11 minutes ago, Dionyz said:

July??? They are not going to release it anytime soon. You have to understand business. If they are selling out of their current cards then why would they release a new better product? It would be retarded for Nvidia to release a product anytime soon.

because then they can sell out of both

[AresKrieger]

2 hours ago, Humbug said:

080ti is pretty much in line with Nvidia's usual pace of improvement correct?

Eh it depends, the 1070~ 980ti and the 780ti~980 or at least was but the 980 is so close to a 970 that I think its a bit low tbh, hard to say prior to launch but I would have guessed 1170~1080ti

[Coaxialgamer]

12 minutes ago, Razor01 said:

 

The architecture is different its not a pascal refresh.  This is something we do know, nV stated Volta's architecture will be in the cards that come after Volta.

I had written a really fleshed out answer , and accidentally pressed backspace . This sucks , but i have to write everything back again .... If we assume that the leak IS correct , here's what we can indeed assume with relative confidence.

 

 

You certainly have a point by saying that new chips will be using volta . But you do have to ask yourself : what is volta ? By looking at various whitepapers and the like , it can be boiled down to 3 things , for GV100 which is the only implementation of volta currently known:

1) Tensor cores

2) New SM layout ( 64 FP32 cores per SM , split as 4*16 FP32 . Each quarter SM also has 8 FP64 cores , 16 INT32 and 2 tensor cores , and this is done to further increase utilization , at the expense of transistor count )

3)improved caching 

 

The first 2 points are the ones to focus on . It's safe to say that tensor cores will be absent from consumer volta cards . Now , about the SM layout , let's look at the improvements vs the chip it replaces to get an idea of the improvements : GP100 has SM's with 64 FP32 cores , split as 2*32FP32 . Each section also gets 32 INT32 cores as well as 16 FP64 cores . The improvements brought by volta are significant . However , it's also irrelevant.

 

See , pascal had 2 implementations , one for server and one for workstation/consumer . The one we got for geforce and most quadros wasn't as listed above . Instead , we got an SM layout with 128 FP32 cores and borked FP64 capability due to the lack of dedicated FP64 compute cores. In essence identical to the existing Maxwell SM .

Despite being based on pascal , we got almost none of the attributes of the bigger brother. That's not necessarily a good or bad thing , but whatever .

 

What's my point with this ? It's likely what we get as consumer volta cards will be extremely similar to existing consumer pascal cards , and here's why :

- Tensor cores will likely be cut , as mentioned

- FP64 capability will remain poor due to lack of demand in the consumer space ( 1/32 most likely )

- Splitting SM's and/or making them smaller is done to increase utilization of said SM by decreasing the required size of a wavefront for optimal use. Poor SM/CU utilisation is usually observed in compute workloads ; as graphics workloads are usually embarrassingly parallel and run fine on the the existing setup . Given the use of these cards for lighter compute applications and graphics work , the SM core count will likely remain fixed to 128 FP32 cores to reduce transistor count vs splitting them into separate units.

-This is further shown by the similar relative die size of the chip , given what we know about the new process. If we did see the optimized SM layout , it is unlikely for so many cores would fit in that space , especially considering how much stuff nvidia has already thrown out in maxwell and pascal.

 

In the end , while technically being based on volta , gtx 1180 might be extremely similar to pascal in both performance and features.

 

Tl,dr: It's unlikely many of volta's features made it over to the geforce line . As such it's likely volta gefore ( or turing , whatever they try to call it ) will end up extremely similar to pascal , turning it effectively into a 12nm refresh with more cores

 

 

[Coaxialgamer]

45 minutes ago, laminutederire said:

I am not sure it's that intentional of them. They'd be more than happy to bury AMD and have a free reign until 2020 at least when Intel will come into play again. If they don't it's probably that they can't.

You know how it is . If they try pushing their market share even more , they'd likely start having problems with monopoly market share and the like . AMD is basically non-existent in the gaming space at the moment , and is still below nvidia in the professional space . they can effectively be ignored for the moment .

Increasing R&D to squash AMD even further likely wouldn't give them a significant advantage in terms of cash , so why bother.

 

Nvidia can afford to be complacent , and i think that's what they'll do for the foreseeable future . I mean , the gtx 1080 is nearly 2 years old at this point . If that's not a sign nvidia is holding back , i don't know what is...

 

[Razor01]

15 minutes ago, Coaxialgamer said:

I had written a really fleshed out answer , and accidentally pressed backspace . This sucks , but i have to write everything back again .... If we assume that the leak IS correct , here's what we can indeed assume with relative confidence.

 

 

You certainly have a point by saying that new chips will be using volta . But you do have to ask yourself : what is volta ? By looking at various whitepapers and the like , it can be boiled down to 3 things , for GV100 which is the only implementation of volta currently known:

1) Tensor cores

2) New SM layout ( 64 FP32 cores per SM , split as 4*16 FP32 . Each quarter SM also has 8 FP64 cores , 16 INT32 and 2 tensor cores , and this is done to further increase utilization , at the expense of transistor count )

3)improved caching 

 

The first 2 points are the ones to focus on . It's safe to say that tensor cores will be absent from consumer volta cards . Now , about the SM layout , let's look at the improvements vs the chip it replaces to get an idea of the improvements : GP100 has SM's with 64 FP32 cores , split as 2*32FP32 . Each section also gets 32 INT32 cores as well as 16 FP64 cores . The improvements brought by volta are significant . However , it's also irrelevant.

 

See , pascal had 2 implementations , one for server and one for workstation/consumer . The one we got for geforce and most quadros wasn't as listed above . Instead , we got an SM layout with 128 FP32 cores and borked FP64 capability due to the lack of dedicated FP64 compute cores. In essence identical to the existing Maxwell SM .

Despite being based on pascal , we got almost none of the attributes of the bigger brother. That's not necessarily a good or bad thing , but whatever .

 

What's my point with this ? It's likely what we get as consumer volta cards will be extremely similar to existing consumer pascal cards , and here's why :

- Tensor cores will likely be cut , as mentioned

- FP64 capability will remain poor due to lack of demand in the consumer space ( 1/32 most likely )

- Splitting SM's and/or making them smaller is done to increase utilization of said SM by decreasing the required size of a wavefront for optimal use. Poor SM/CU utilisation is usually observed in compute workloads ; as graphics workloads are usually embarrassingly parallel and run fine on the the existing setup . Given the use of these cards for lighter compute applications and graphics work , the SM core count will likely remain fixed to 128 FP32 cores to reduce transistor count vs splitting them into separate units.

-This is further shown by the similar relative die size of the chip , given what we know about the new process. If we did see the optimized SM layout , it is unlikely for so many cores would fit in that space , especially considering how much stuff nvidia has already thrown out in maxwell and pascal.

 

In the end , while technically being based on volta , gtx 1180 might be extremely similar to pascal in both performance and features.

 

Tl,dr: It's unlikely many of volta's features made it over to the geforce line . As such it's likely volta gefore ( or turing , whatever they try to call it ) will end up extremely similar to pascal , turning it effectively into a 12nm refresh with more cores

 

 

The volta architecture is not just its tensor and DP cores as you stated.  Every single one of nV's architectures since g80 have improved their throughput in various units and pathways.  I expect the same with Volta.  Volta has shown its compute throughput (without DP and Tensor core usage) is anywhere from 50% to 100%, factoring in the units, even at 50% that is higher throughput than what pascal gives.  (different tests with mining algos).  Even in some DX12 games we have seen it scaling more than what its units are providing.

 

Volta's SM's != to Pascal's SM's, the independent ability for processing threads means the sm's have changed quite a bit.  From a compute stand point we do see the performance differences there as mentioned above

 

I don't know at this point if they will completely cut out the tensor cores either, tensor cores might be linked with the DP units.  So what ever we see that cut down to, we might see the same ratio of tensor units as well.

 

Similar die size means similar amount of transistors being used but how they were used we don't know that, I expect a similar SM count chip with Volta SM's to use more transistors and thus be bigger than Pascal if its the same/similar process, at this point we don't even know if its on 12nm TSMC ;).  To me WTFtech is just extrapolating from a Maxwell to Pascal to next gen based on V100, which at this point I don't think its wise to do because we don't know the node its coming out on.  Things we can see is the performance of V100 in specific tasks, those synthetics come in quite handy in this regard.

[Coaxialgamer]

1 minute ago, Razor01 said:

The volta architecture is not just its tensor and DP cores as you stated.  Every single one of nV's architectures since g80 have improved their throughput in various units and pathways.  I expect the same with Volta.  Volta has shown its compute throughput (without DP and Tensor core usage) is anywhere from 50% to 100%, factoring in the units, even at 50% that is higher throughput than what pascal gives.  (different tests with mining algos).  Even in some DX12 games we have seen it scaling more than what its units are providing.

 

Volta's SM's != to Pascals SM's, the independent ability for processing threads means the sm's have changed quite a bit.  From a compute stand point we do see the performance differences there as mentioned above

 

I don't know at this point if they will completely cut out the tensor cores either, tensor cores might be linked with the DP units.  So what ever we see that cut down to, we might see the same ratio of tensor units as well.

 

Similar die size means similar amount of transistors being used but how they were used we don't know that, I expect a similar SM count chip with Volta SM's to use more transistors and thus be bigger than Pascal if its the same/similar process, at this point we don't even know if its on 12nm TSMC ;).  To me WTFtech is just extrapolating from a Maxwell to Pascal to next gen.

Well , my point isn't that volta isn't a significant improvement over pascal  : it clearly is , as you've just stated. My point is that the the only current implementation of volta is gv100 , with all it's bells and whistles. What i'm saying is that what we consumers actually get will likely be considerably different , to the point that that the consumer arch and gv100 arch will be similar by name only .