Upcoming Geforce GTX Volta cards Use GDDR5X not HBM2

[Fluffmeister]

Volta is supposed to increase performance per watt even further over Pascal, AMD are really falling behind at this point. I don't think they even fully match Maxwell on this metric.

 

[Vya Domus]

Ya think? Look at Nvidia's GP104 which has +30-35% over the 980ti (GM200) which is right next to it.

https://goo.gl/images/PW4sWK

Pascal Refresh is probably going to be a similar move to Kepler Refresh, might bring 10-20% rebrands with somewhat faster GDDR5X and replaces the existing GP10x parts entirely, better binning/process and they don't even need to increase shader counts. That's gonna be the same die size. The 1070 becomes the 1180 and done.

For Volta, they can just add a couple clusters on there and they're already done, they got plenty of room there. Heck, they could double the die size and keep refreshing Pascal 16nm until 2020 if they want to.

And after this Pascal refresh ? What will they do ? There wont be any room left. Jensen himself basically said the same on stage when V100 was announced.

Nvidia is in fact the ONLY chip architect that is still giving us performance increases and they are also the only one doing it while staying at roughly the same TDP. Even mighty Intel hasn't matched that feat for the last decade or so.

Can't compare power/performance characteristics of a CPU to a GPU. They react differently to what you change and what you gain. You can allocate 30% more die space for control/cache/wider SIMD instructions on CPU and you will never see a 30% performance boost. Meanwhile , on a GPU for 30% more shaders you get exactly 30% more performance at the same clocks. On a CPU it is much more difficult to gain performance without destroying power consumption. Take a look at Skylake-X

 

[Vayra86]

And after this Pascal refresh ? What will they do ?

By then its 2020 and we have 7nm fo sho and AMD is probably not even competing any longer.

Also, I added the final solution as well to my previous post, way ahead of you mate

Also there is this:
http://research.nvidia.com/publication/2017-06_MCM-GPU:-Multi-Chip-Module-GPUs

on a GPU for 30% more shaders you get exactly 30% more performance at the same clocks. On a CPU it is much more difficult to gain performance without destroying power consumption. Take a look at Skylake-X

Well, for your answer on both of these, I'd say: look at AMD. Nvidia doesn't put 30% more shaders, they put 30% more performance with similar shader counts and a lower TDP. AMD cannot seem to achieve that. And on the Intel side, AMD is destroying the perf/watt that Intel has been carefully building on for over 10 years, with one CPU release, while Intel can only present us with clock bumps for higher TDP, basically no progression.

 

[Vya Domus]

By then its 2020 and we have 7nm fo sho and AMD is probably not even competing any longer.

Also, I added the final solution as well to my previous post, way ahead of you mate

Also there is this:
http://research.nvidia.com/publication/2017-06_MCM-GPU:-Multi-Chip-Module-GPUs



Well, for your answer on both of these, I'd say: look at AMD. Nvidia doesn't put 30% more shaders, they put 30% more performance with similar shader counts and a lower TDP. AMD cannot seem to achieve that. And on the Intel side, AMD is destroying the perf/watt that Intel has been carefully building on for over 10 years, with one CPU release.

And I am telling you there is so much performance you can gain when you do that.

7nm ? Sure. At what cost ? You fail to factor in how expensive this die shrink practice is becoming with each step.

More dies on the same package ? Guess what that costs next to nothing to do , AMD can and will jump into that. As a matter of fact they brought that up way before Nvidia. So I fail to see how they can't be competitive.

You are praising Nvidia way to much. They have done their fair share of engineering achievements , but there are limits. As a matter of fact , you should be praising TSMC

 

[Fluffmeister]

Nvidia are doomed!!!!1

 

[Vayra86]

And I am telling you there is so much performance you can gain when you do that.

7nm ? Sure. At what cost ? You fail to factor in how expensive this die shrink practice is becoming with each step.

More dies on the same package ? Guess what that costs next to nothing to do , AMD can and will jump into that. As a matter of fact they brought that up way before Nvidia.

Well more than 50% of my previous posts was an 'educated joke/stab' at the oddities of semiconductor progress. Realistically though.... they're probably all viable ways to go forward.

7nm costly, yes, but only because EUV isn't ready yet. That is why I reckon 2020 is a sensible guesstimate on that, it also gives 14/16nm some time to get marketed. Regardless, we can look ahead more than one generation but these are all guesses at this point. We started this discussion with 'the wall is going to hit Nvidia soon' but as you can see there are tons of ways they can still move forward with the tech they have available *today*, while its direct competitor struggles to not burn the GPU alive and that other chip maker can't even produce a gaming capable APU in the first place.

Bottom line, Nvidia is probably in the most comfortable position of all three. Oh yeah, they have an inhouse APU and CPU design too, and its selling to some of the most lucrative branches in the world: cars.

 

[Vya Domus]

Well more than 50% of my previous posts was an 'educated joke/stab' at the oddities of semiconductor progress. Realistically though.... they're probably all viable ways to go forward.

7nm costly, yes, but only because EUV isn't ready yet. That is why I reckon 2020 is a sensible guesstimate on that, it also gives 14/16nm some time to get marketed. Regardless, we can look ahead more than one generation but these are all guesses at this point. We started this discussion with 'the wall is going to hit Nvidia soon' but as you can see there are tons of ways they can still move forward with the tech they have available *today*, while its direct competitor struggles to not burn the GPU alive and that other chip maker can't even produce a gaming capable APU in the first place.

Bottom line, Nvidia is probably in the most comfortable position of all three. Oh yeah, they have an inhouse APU and CPU design too, and its selling to some of the most lucrative branches in the world: cars.

Who said they aren't in a comfortable position. I am saying it's not going to be easy for them to hold onto it , at least with regards to some markets.

CPUs ? You mean ARM ? That ain't Nvidia I'm afraid. Those APUs for cars and embedded are great and they can make a lot of money with them. For anything else though , not so much. Meanwhile AMD with x86 and way faster APUs as a result , can pretty much garner every other market.

You seem to think they have a genie somewhere that will grant them immunity and unrivaled success in all aspects. Cool , we'll see. I am telling you they are more scared than ever , hence the mad push for data centers/embedded.

 

[Vayra86]

CPUs ? You mean ARM ? That ain't Nvidia I'm afraid.

'fraid it is.
https://en.wikipedia.org/wiki/Drive_PX-series

Denver cores are an Nvidia design (and a good one at that, too). And marriage of all the different parts is also their own design. Nvidia is also good on an inhouse fast interconnect called NvLink. They are rapidly building an ecosystem where they can mix and match inhouse parts (or licensed ARM if needed) to come up with custom SoCs for specific purposes. They're leading deep learning performance as well. I don't see a genie somewhere, I see tangible moves and they are all still extremely close to the core business. Almost everything they do is multi purpose. They have way too much riding on this, and no reason to slow down, plus they have headroom on every physical aspect, be it power or size.

Immunity, no, but very well insulated.

 

[Vya Domus]

'fraid it is.
https://en.wikipedia.org/wiki/Drive_PX-series

Denver cores are an Nvidia design (and a good one at that, too). And marriage of all the different parts is also their own design. Nvidia is also good on an inhouse fast interconnect called NvLink. They are rapidly building an ecosystem where they can mix and match inhouse parts (or licensed ARM if needed) to come up with custom SoCs for specific purposes. They're leading deep learning performance as well. I don't see a genie somewhere, I see tangible moves and they are all still extremely close to the core business. Almost everything they do is multi purpose. They have way too much riding on this, and no reason to slow down, plus they have headroom on every physical aspect, be it power or size.

Immunity, no, but very well insulated.

Yes it's their implementation but that's still ARM at core , a weak RISC based architecture designed in the late 80s. Good enough for embedded but not brilliant ( no matter how much headroom you have , you can't just put 100 cores on there) , considering that demands for performance in this industry have raised significantly over the years.

Extremely close to core businesses ? That of making GPUs ? Yes they are. But they aren't close to core target markets anymore.

I am just going leave you with this : V100 is a whopping 800mm^2 on 12nm , Intel's largest Skylake-SP die is 700mm^2 on 14nm. Intel however has it's own fabs and 60 billion in revenue. Nvidia doesn't and that GPU is probably the most expensive chip to manufacture in large quantities on the planet currently. That says a lot about how much they are investing in just one market , it can pay off big time but at the same time it can end up a disaster , meanwhile the focus on other fronts diminishes.

Anyway , I am glad to have such discussions on here , better than the "this sucks cuz it red/green/blue" crap you find in other places .

 

[Vario]

Anyway , I am glad to have such discussions on here , better than the "this sucks cuz it red/green/blue" crap you find in other places .

Most members here are all old timers that have had both brands over the last decade or two and so there isn't as much fanboyism.