GK104's successor?

[Anarchist420]

I thought it was interesting when I heard that the Unreal 4 engine was going to use more compute functions. That probably means that GTX680's won't be able to compete.

Anyway, predict whether you think the full hardware scheduler is done away with for good in nvidia's consumer GPUs, what you think the ratio of FP64 to FP32 performance will be, if you think RGBA16 blending will remain half speed, if you think the read only data cache for the CUDA cores will be increased, if you think the ROPs will be decoupled from the ROPs, and stuff like that.

I'm personally going to guess that full hardware scheduling will be back and that the ratio of FP64 to FP32 performance will be 1/6, I think that more ROPs will be added rather than ones that do RGBA16 in a single clock cycle, and I think that the ROPs will still be attached to the memory controller. I'm guessing the 48KB read only data cache for each CUDA cluster will be increased to 64KB since that would be necessary to maximize FP64 performance. I think nv kind of dropped the ball with Kepler, so I may wait until it's successor to upgrade.

I know that most here know more than I do, so I'd love to hear some predicitons

 

[boxleitnerb]

Wouldn't all this put "GK114" too close to GK110 and tank energy efficiency?

 

[ShintaiDK]

Compute can be SP too, just like PhysX. No problem for GK104 there. You must assume that Unreal 4 engine will use DP. And thats very unlikely. HD78xx would face the same problems as GK104. And making an engine that basicly only works on HD79xx and hopes on the shear generousity from GPU manufactors to implement higher DP performance in all products? Really?

Either Unreal 4 wont use any compute. Or it will use SP.

 

[Anarchist420]

Doesn't the full hardware scheduler make compute performance easier to obtain?

 

[Silverforce11]

Compute can be SP too, just like PhysX. No problem for GK104 there. You must assume that Unreal 4 engine will use DP. And thats very unlikely. HD78xx would face the same problems as GK104. And making an engine that basicly only works on HD79xx and hopes on the shear generousity from GPU manufactors to implement higher DP performance in all products? Really?

Either Unreal 4 wont use any compute. Or it will use SP.

You do realize the 78xx is nowhere near as crippled in DP as gk104??

But you are correct, compute is a very general terms, but bitcoin mining is all SP, why is it Kepler still struggle there, some one knows?

 

[Homeles]

Compute can be SP too, just like PhysX. No problem for GK104 there. You must assume that Unreal 4 engine will use DP. And thats very unlikely. HD78xx would face the same problems as GK104. And making an engine that basicly only works on HD79xx and hopes on the shear generousity from GPU manufactors to implement higher DP performance in all products? Really?

Either Unreal 4 wont use any compute. Or it will use SP.

Pitcairn has 1/16th FP64 performance. GK104 is 1/24.

 

[Pr0d1gy]

I was going to post my own thread but I will toss my hat into this one. Does anyone know if nVidia plans to make a new GPU lineup for Haswell's release in the Spring?

 

[Homeles]

Almost undoubtedly, yes (although not specifically made for Haswell). There's been no official announcement for GeForce products specifically, but it'd be absolutely foolish of them not to have a new line of cards next year.

 

[ShintaiDK]

You do realize the 78xx is nowhere near as crippled in DP as gk104??

But you are correct, compute is a very general terms, but bitcoin mining is all SP, why is it Kepler still struggle there, some one knows?

Pitcairn has 1/16th FP64 performance. GK104 is 1/24.

They perform roughly the same.

This one uses SP.

For the Bitcoin question. A huge chunk relates to SHA-256. This operation can be implemented as a single hardware instruction on AMD GPUs (BIT_ALIGN_INT), but requires three separate hardware instructions to be emulated on Nvidia GPUs (2 shifts + 1 add). This alone gives AMD a 1.7x performance advantage (~1900 instructions instead of ~3250 to execute the SHA-256 compression function).

 

[Homeles]

They perform roughly the same.

As far as theorteticals go, Pitcairn's roughly ~25% faster (160 DP FLOPS vs. ~128). With software, your mileage will vary, of course. Still, any meaningful performance hit from hypothetical DP compute would be negated by the fact that the 680 is faster at everything else.

 

[Anarchist420]

This one uses SP.

What the hell is nearest neighbor still being used for?

 

[f1sherman]

Performance comparison between Tesla (GTX285), Fermi (GTX480) and Kepler (GTX680)

 

[Lonbjerg]

Funny how a few OpenCL benches makes people think GK104 will suck for gaming compute

 

[dangerman1337]

Epic used GTX 680 for UE4 demonstrations so I don't think it can't use its features well IIRC.

 

[RussianSensation]

Epic used GTX 680 for UE4 demonstrations so I don't think it can't use its features well IIRC.

This. Considering EPIC is already working closely with NV and GTX680 ran all the UE4 demos so far, if anything Kepler will work faster than 7900 series of cards. Either way, Fortnight (the 1st UE4 engine game) doesn't look that revolutionary in terms of graphics aspects of UE4.

By the time UE4 game engines are starting to come to market in similar volumes to UE3 games (2014-2017), we'll have moved on from this generation a long time ago. It's no different than when 7800GTX 256mb was capable of playing UE3 and 6800U struggled. Sure, 7800GTX 256mb could play those newer games a little bit better but really it's a moot point since we'll be playing those next generation games on GTX700/800/HD8000/9000 series, etc.

 

[Lonbjerg]

This. Considering EPIC is already working closely with NV and GTX680 ran all the UE4 demos so far, if anything Kepler will work faster than 7900 series of cards. Either way, Fortnight (the 1st UE4 engine game) doesn't look that revolutionary in terms of graphics aspects of UE4.

By the time UE4 game engines are starting to come to market in similar volumes to UE3 games (2014-2017), we'll have moved on from this generation a long time ago. It's no different than when 7800GTX 256mb was capable of playing UE3 and 6800U struggled. Sure, 7800GTX 256mb could play those newer games a little bit better but really it's a moot point since we'll be playing those next generation games on GTX700/800/HD8000/9000 series, etc.

How about an UE3 vs UE4 shot?

UE3:

UE4:

Not that I like that type of graphic though...but it is "better"...

 

[dangerman1337]

This. Considering EPIC is already working closely with NV and GTX680 ran all the UE4 demos so far, if anything Kepler will work faster than 7900 series of cards. Either way, Fortnight (the 1st UE4 engine game) doesn't look that revolutionary in terms of graphics aspects of UE4.

Well Fortnight was originally on the UE3 IIRC (though I do hope they still have a UT on UE4 in the works but I dobut it).

EDIT: It's also on a team with 15 developers on it so it looks very good for what it is.

 

[Pottuvoi]

I really doubt full scheduling will come back anymore, the idea is to have as simple as possible cores and as many as possible.

Reducing Power consumption is the only way forward and scheduling wastes a lot of power.
If they do not already have things like 1KB L0 caches and such it is guaranteed that Maxwell will have them as well. (if data is found from L0 instead of 32KB L1 it costs 1/32 of the power.)

Nvidia has been quite open about their future platforms in compute side and explained the reasons why they are going that way. (AMD is going for similar direction.)

 

[RussianSensation]

Not that I like that type of graphic though...but it is "better"...

The 2nd one does look much nicer. What other UE3 games are scheduled for 2013?

 

[dangerman1337]

The 2nd one does look much nicer. What other UE3 games are scheduled for 2013?

Here's a list: http://en.wikipedia.org/wiki/List_of_Unreal_Engine_games#Unreal_Engine_3