NVIDIA Pascal Thread

[maddie]

Why? You can get 4096bit HBM2 to communicate with a GPU.

All that's important are lanes of communications.

When you look at GPU uarchs, they have blocks communicating with cache & each other. Within each block, there's TMUs, ROPS, and cores. Thus they often cut a few blocks out to make a lower end SKU.

A multi-chip approach wouldn't work without the interposer giving it the lanes of communications that it needs.

We'll see it with Navi, late 2018 or whenever 10/7nm comes around early on.

As someone who has speculated in this a fair amount, my biggest concern is the large energy cost in moving data across larger distances and how to mitigate that issue.

If we assume that the best games going forward will use DX12, we might see, as an intermediate step the use of explicit multi-adapter techniques for the multiple die. Maybe 1 bank of HBM2 per die with no texture, etc, duplication, instead of all the die needing access to all the memory. Worse case a DX11 game will see each small die sufficient to give good frame rates. This would also allow great memory scaling.

Small die in this case is relative as interposers allow the total GPU area to exceed the past single die practical limit of 600mm^2.

 

[HurleyBird]

As opposed to laying out multiple dies side by side on an interposer, die stacking might also be a possibility.

And/or instead of identical dies, there might be specialized ones arranged in a way where inter-die communication is minimized.

Of course, all of this sounds like something that could easily happen eventually, but it's difficult to imagine it in the next couple of years (at least for the likes of Nvidia and AMD, Intel is a different matter).

 

[nvgpu]

http://videocardz.com/59154/nvidia-geforce-gtx-1080-pictured-again

 

[antihelten]

http://videocardz.com/59154/nvidia-geforce-gtx-1080-pictured-again

While it's nice to see pics like this (since it indicates that release is getting close), I have to say that I'm not a fan of the visuals of this design. Oh well, hopefully it performs well.

 

[Silverforce11]

I think it looks awesome, very premium worthy.

 

[antihelten]

I think it looks awesome, very premium worthy.

To me it pretty much looks like they took the old cooler and just threw a bunch of angles and edges at it, because that's what the cool kids want nowadays. It's the same kind of thinking that leads to PC cases like this (the Nvidia cooler isn't quite that bad though).

Oh well, different strokes for different people I suppose (besides I don't use cases with windows so it really doesn't affect me at all).

 

[Silverforce11]

Their reference cooler is good enough, making it look better will be minimal for R&D and maximal for returns.

AMD's reference is junk. They shouldn't launch any at all unless they really improved it.

 

[Sweepr]

Beautiful, but I hate the name. Geforce X80 sounds better.

 

[antihelten]

Their reference cooler is good enough, making it look better will be minimal for R&D and maximal for returns.

AMD's reference is junk. They shouldn't launch any at all unless they really improved it.

Nvidia's cooler is good enough for TDPs up to ~200W, but beyond that and it starts to get a bit iffy (nowhere near AMD reference bad of course). Of course GP104 should be below 200W so the performance should be fine.

I'm not talking about the performance though, I'm talking purely about looks, and what can I say, I'm not a fan (no pun intended).

 

[ShintaiDK]

Beautiful, but I hate the name. Geforce X80 sounds better.

I agree. But marketing is marketing.

 

[Adored]

Ugly as hell for me, and ridiculous naming on top.

 

[nvgpu]

http://videocardz.com/59163/nvidia-pascal-gp106-in-autumn

The same source also claims that GP106 cards will not require additional power connectors. So it might fit into same power envelope as GTX 750 series or recently announced 75W version of GTX 950

GTX 970/980 levels of performance without a power connector? That's insane performance per watt improvement there.

 

[Krteq]

GTX 970/980 levels of performance...

Where is this info from?

 

[ShintaiDK]

Ugly as hell for me, and ridiculous naming on top.

It not like you was ever in the market for one. And you can always get a 3rd party cooler.

Not that I see it matters in any way, unless you run a glass window case.

 

[kraatus77]

Where is this info from?

From his behind obviously. even with 2x p/w, it will require atleast 80 watts to deliver 970's stock performance. and that means no additional power connector will limit it's oc ability. in any case we'll see soon.

 

[Adored]

It not like you was ever in the market for one. And you can always get a 3rd party cooler.

Not that I see it matters in any way, unless you run a glass window case.

I fully intend to buy one but it'll depend on pricing too.

 

[ShintaiDK]

From his behind obviously. even with 2x p/w, it will require atleast 80 watts to deliver 970's stock performance. and that means no additional power connector will limit it's oc ability. in any case we'll see soon.

Not to mention half the bandwidth, unless its GDDR5X equipped.

Its no different than the fantasy dreams with Polaris.

 

[kraatus77]

Not to mention half the bandwidth, unless its GDDR5X equipped.

Its no different than the fantasy dreams with Polaris.

Actually it is since 290/x is already faster than 970 in dx12 or any situation where its running optimized games, since gcn>polaris is a bigger difference than maxwell>pascal, it's possible that 40cu polaris chip can beat 970 at 90-110w range. but green eyes won't see it. in any case we'll see it in 2-3 months.

 

[nvgpu]

http://international.download.nvidi...tional/pdfs/GeForce-GTX-750-Ti-Whitepaper.pdf

Because of GM107’s remarkable architectural efficiency, at 1080p resolution a GeForce GTX 750 Ti will frequently match the
performance of our flagship GPU from four years ago, the GeForce GTX 480, but with only a 60W TDP, consumes a fourth of the power.

There's still doubters here that think a 16nm FF+ part with more transistors being able to be packed in won't match or exceed an older part in performance.

 

[Krteq]

Nvidia’s mid-range Pascal GP106 may not debut at Computex

 

[Adored]

http://international.download.nvidi...tional/pdfs/GeForce-GTX-750-Ti-Whitepaper.pdf



There's still doubters here that think a 16nm FF+ part with more transistors being able to be packed in won't match or exceed an older part in performance.

Comparing any GPU in perf/Watt to the 480 is gonna look good...

 

[kraatus77]

http://international.download.nvidi...tional/pdfs/GeForce-GTX-750-Ti-Whitepaper.pdf



There's still doubters here that think a 16nm FF+ part with more transistors being able to be packed in won't match or exceed an older part in performance.

Nobody is doubting anything. but to say we'll get 970's performance without any power connector is nonsense. it will require same amount of bandwidth + more than 2x p/w.

that example you gave is irrelevant because 480 wasn't that efficient at first place. 970 is already using color compression and it's already limited by bandwidth. any more reduction in bandwidth means less performance regardless whatever efficiency those shaders provide. at best i can see it delivering 15-20% less performance without additional power.

 

[Glo.]

Nobody is doubting anything. but to say we'll get 970's performance without any power connector is nonsense. it will require same amount of bandwidth + more than 2x p/w.

that example you gave is irrelevant because 480 wasn't that efficient at first place. 970 is already using color compression and it's already limited by bandwidth. any more reduction in bandwidth means less performance regardless whatever efficiency those shaders provide. at best i can see it delivering 15-20% less performance without additional power.

You forget IPC improvements over Maxwell, and... GPC configurations. 1280 Pascal CUDA core GPU clocked at 1480 MHz with GDDR5X can achieve performance of GTX 970.

The question is: would it be able to maintain thermal envelope without power connector. That last bit is most questionable.

 

[R0H1T]

http://international.download.nvidi...tional/pdfs/GeForce-GTX-750-Ti-Whitepaper.pdf



There's still doubters here that think a 16nm FF+ part with more transistors being able to be packed in won't match or exceed an older part in performance.

Would you be willing to eat that (proverbial) hat of yours when we don't get GTX 970/980 level of performance on launch day? I doubt Nvidia, or AMD for that mater, will be able to pull such a coup without GDDR5x & even then 980 level of performance (for ~75W) is a pipedream sans HBM2

You forget IPC improvements over Maxwell, and... GPC configurations. 1280 Pascal CUDA core GPU clocked at 1480 MHz with GDDR5X can achieve performance of GTX 970.

The question is: would it be able to maintain thermal envelope without power connector. That last bit is most questionable.

What about compute, you don't think lack of compute, as compared to GCN4, will have a major impact on DX12 titles? This assuming (gaming) Pascal isn't all that good at compute, of course we don't know that for certain.

 

[kraatus77]

You forget IPC improvements over Maxwell, and... GPC configurations. 1280 Pascal CUDA core GPU clocked at 1480 MHz with GDDR5X can achieve performance of GTX 970.

The question is: would it be able to maintain thermal envelope without power connector. That last bit is most questionable.

Hh also i forgot one thing, drivers which can also be used to nerf 970 and that memory configuration will take the blame. LOL


im not saying there won't be any p/w improvements. but you need to look at bandwidth too, there's a limit to how one can make it efficient. since maxwell already has color compression and there's a very little chance of 1070 using gddr5x i can't see it delivering that kind of performance in 75w envelope. specially with less bandwidth than what 970 has.