Nvidia Pascal Announcement

[Ext3h]

Yes, that's a great improvement indeed. Because that means mixing dispatch and draw calls is no longer likely to cause catastrophic outcomes.

sebbbi - is it just one compute shader and one graphics shader per CU, or can each wave effectively be doing its own thing?

10 waves at most, regardless of where they come from. The GPU will try to fit as many as possible.

The heuristics on wave level scheduling must be insane. How does AMD decide on the number of slots to allocate to graphics or compute?

My guess (no details are known): At least fixed priorities, best-fit based on various metrics, possibly some of them compiler-guided, possibly even priority aging.

 

[silent_guy]

The heuristics on wave level scheduling must be insane. How does AMD decide on the number of slots to allocate to graphics or compute?

And how do you decide whether or not a particular load will trash the cache? It's probably the kind of thing that explains performance increases with new driver release: somebody manually tuned a dial.

 

[Genotypical]

Possibly, but all of this is coming from my original question that pertained to a single pass rendering for multiple displays; whether that is both left+right lens, 3 displays,etc.

I have gone through the previous VRWorks and Maxwell presentation and none of them showed that capability, their focus was on a single window/display.
Now with Pascal they are showing the ability to do this for both multiple displays (so you can wrap 3 monitors and have the 3d environment corrected) and VR.
I appreciate I could be wrong with regards to what the presentation showed at 55mins and 1hr 4mins, but I have not seen any information from NVIDIA before Pascal to suggest otherwise.
Anyone with NVIDIA Maxwell-VR papers saying otherwise?

Edit:
In some way it could be deemed as serious as that would mean 30 minutes of that presentation was meaningless and should be criticised for misrepresenting the technology implemented.

Cheers

If they hadn't developed the feature in VRworks at the time then they would not say maxwell had it. In fact Pascal might not even have it yet on release if its down to software that's not fully ready. Never know with this kind of presentation. A difference in what was said between pascal and maxwell would not exclude software as the main source of the capability.

Here you go: http://arstechnica.com/gadgets/2016/04/amd-polaris-will-be-a-mainstream-gpu/

Take it whatever way you want, some may argue he was only talking about price, while others including myself is that AMD identified the initial Pascal to be a high-end consumer focused product.

Likely he was talking about the 1080, which is high end priced. 1070 is a product that varies based on yields etc so thats harder to pin down on price and performance. I expect polaris to be competitive with the 1070 at least, and cheaper. I think people are thinking polaris clocks will be low. Nvidia managed 1600mhz stock with 16nm, so would AMD really end up that low (seems people assume around 1Gz or even 800Mhz)?

 

[CSI PC]

If they hadn't developed the feature in VRworks at the time then they would not say maxwell had it. In fact Pascal might not even have it yet on release if its down to software that's not fully ready. Never know with this kind of presentation. A difference in what was said between pascal and maxwell would not exclude software as the main source of the capability.



Likely he was talking about the 1080, which is high end priced. 1070 is a product that varies based on yields etc so thats harder to pin down on price and performance. I expect polaris to be competitive with the 1070 at least, and cheaper. I think people are thinking polaris clocks will be low. Nvidia managed 1600mhz stock with 16nm, so would AMD really end up that low (seems people assume around 1Gz or even 800Mhz)?

In one of the links I provided NVIDIA specifically mention Pascal architecture with the points I am raising.

Sorry but everyone knows 1070 and 1080 are releasing very close to each other and before the GP106 that is meant to be in next 2-5 months...
Therefore if Roy was thinking initial release Pascal with the 1070 not high end, he should not say Pascal is only a high-end part and not competing against Polaris...
Again, this should be taken to the AMD Polaris thread if one wants to discuss why they feel the Ars Technica interview and their statements about Polaris being mainstream is wrong.....

 

[PeanutButterOnPickles]

EVGA PrecisionX OC
GPU Clock: 2114 MHz
Memory Clock: 5508 MHz
GPU Temp: 67c

Wow! I can't wait to see these GPUs water cooled.

 

[firstminion]

In one of the links I provided NVIDIA specifically mention Pascal architecture with the points I am raising.

Sorry but everyone knows 1070 and 1080 are releasing very close to each other and before the GP106 that is meant to be in next 2-5 months...
Therefore if Roy was thinking initial release Pascal with the 1070 not high end, he should not say Pascal is only a high-end part and not competing against Polaris...
Again, this should be taken to the AMD Polaris thread if one wants to discuss why they feel the Ars Technica interview and their statements about Polaris being mainstream is wrong.....

Correct me, but at the time Nvidia only talked about GP100 and it would be a really expensive consumer card. 1080 and 1070 are GP104 if anandtech is correct.

 

[Infinisearch]

http://www.pcgamer.com/id-software-shows-doom-running-on-gtx-1080-using-vulkan/

200fps at 1080p ultra.

edit - actually paying attention to the fps counter its seems to fluctuate from 120 upto 200

 

[lanek]

the new cooler and better overclocking (probably better vrm's, more power connectors etc.)

If i read well anandtech, that just concern reference card with reference pcb and availability ( and looking at the term founders, thats the only concern) .. theres not more power connectors, they are the reference card who will be the only card available the 27tth may. at this price.... AT first i was think it was pre order thing, but it seems not.

so basically the 27th may: Only reference founders in shop at there 699$ price.

 

[lanek]

They may also present it in VR form that will skew it a fair bit.
With the Mech VR demo the 1080 seriously outperforms the TitanX.....

Cheers

Damn i have not seen the "relative VR gaming performance metrics on the left" when watching the stream, i was think " relative gaming performance. " ..

 

[Razor1]

If i read well anandtech, that just concern reference card with reference pcb and availability ( and looking at the term founders, thats the only concern) .. theres not more power connectors, they are the reference card who will be the only card available the 27tth may. at this price.... AT first i was think it was pre order thing, but it seems not.

so basically the 27th may: Only reference founders in shop at there 699$ price.

ah

 

[CSI PC]

Damn i have not seen the "relative VR gaming performance metrics on the left" when watching the stream, i was think " relative gaming performance. " ..

Yeah there are two very similar charts
The key is the positioning of the 1080, with the relative gaming performance it is somewhere around 4.5, the 980 around 2.6, and the TitanX around 3.6.
Basically it is around 75%-80% faster in Rise Tomb Raider and 65%-70% faster in Witcher 3 compared to the 980, so faster than 980 in SLI.
Worth remembering this will be I think reference to reference card comparison, but it looks like it OCs as well as Maxwell 2.

Cheers

 

[CSI PC]

If i read well anandtech, that just concern reference card with reference pcb and availability ( and looking at the term founders, thats the only concern) .. theres not more power connectors, they are the reference card who will be the only card available the 27tth may. at this price.... AT first i was think it was pre order thing, but it seems not.

so basically the 27th may: Only reference founders in shop at there 699$ price.

Yeah that is what I gathered reading the NVIDIA News brief:

Availability and Pricing
The NVIDIA GeForce GTX 1080 "Founders Edition" will be available on May 27 for $699. It will be available from ASUS, Colorful, EVGA, Gainward, Galaxy, Gigabyte, Innovision 3D, MSI, NVIDIA. Palit, PNY and Zotac. Custom boards from partners will vary by region and pricing is expected to start at $599.

The GeForce GTX 1080 will also be sold in fully configured systems from leading U.S.-based system builders, including AVADirect, Cyberpower, Digital Storm, Falcon Northwest, Geekbox, IBUYPOWER, Maingear, Origin PC, Puget Systems, V3 Gaming and Velocity Micro, as well as system integrators outside North America.

The NVIDIA GeForce GTX 1070 "Founders Edition" will be available on June 10 for $449. Custom boards from partners are expected to start at $379.

I guess the custom AIB will be anywhere from 4-6 weeks later.
Cheers

 

[Razor1]

it sounds like GDDR5X might be causing the stretched out time line?

 

[lanek]

it sounds like GDDR5X might be causing the stretched out time line?

Its effectively what sound to be the problem as Micron is their only source for GDDr5x. And Micron have allways said their mass production for it have not started yet.

peoples who was point to it before was right, but no mass production, dont mean no production at all.

Sorry i have edit a bit the last lines as the industry question was not so interesting .

 

[CSI PC]

Well I checked and it seems the custom AIB for the 980 were just under 4 weeks later than the reference design.
If it goes over that then maybe supply issues, but I do not think that is an issue because I checked the status of the 10Gb/s GDDR5X awhile back and it was in production status.
Even the 12Gb/s went into sampling status back in early March, so not long until that can start to be ramped up for production (that will have supply issues but fingers crossed it will be used on the custom AIBs 1080 when it does become available)
Cheers

 

[lanek]

Well I checked and it seems the custom AIB for the 980 were just under 4 weeks later than the reference design.
If it goes over that then maybe supply issues, but I do not think that is an issue because I checked the status of the 10Gb/s GDDR5X awhile back and it was in production status.
Even the 12Gb/s went into sampling status back in early March, so not long until that can start to be ramped up for production (that will have supply issues but fingers crossed it will be used on the custom AIBs 1080 when it does become available)
Cheers

Nobody tell you that they are not in production.. But looking that they only use 10GB/s ( the 1080 with GDDR5x have a lower bandwith than the 980TI ), can make think that it was due to the production .. But it can mean too ( for me ) , that they keep a bit of "overclock" margin for non reference gpu's on the memory side.
( at contrario of AMD gpu's, most non reference gpu's maxwell was sold with higher speed on memory and a large difference in some case, it was even a big part of their marketing )

 

[Kaotik]

http://www.pcgamer.com/id-software-shows-doom-running-on-gtx-1080-using-vulkan/

200fps at 1080p ultra.

edit - actually paying attention to the fps counter its seems to fluctuate from 120 upto 200

Pointelss without some other card as reference point

 

[Razor1]

they did say other cards had much less frames (not in those exact words), and actually with Pascal they found bugs when the game was going so fast.

 

[LiXiangyang]

I hope Nvidia will design a two-8-pin water-cooled dual-chip GP104 card to push the computing density futher, should be more attractive to me than the GP100.

 

[Genotypical]

Pointelss without some other card as reference point

this is the problem with that doom vid. But some comparisons could be possible. 390x can do 60 fps at 1440 (59 minimum). 980 does around 35fps at 4k. Should be around 120-140 fps at 1080p. Though not sure the test done for these numbers is as demanding, but gives some idea. Unfortunately the beta was capped at 60 fps, which I guess is why they are using doom so there is no direct comparison (plus the vulkan). Under vulkan the 390x at least should do even better, probably the 980 as well.