Polaris 10 and 11 confirmed to be GDDR5 based

[jpiniero]

that would be unprecedented.

That's the reality of how expensive this node is.

 

[Adored]

Nvidia aren't stupid enough to take AMD on in per/area. They got lucky with the 680 early on, but that won't happen again on 14nm. Polaris is clearly not another Tahiti. And remember, AMD will be 9% smaller anyway.

For GP104 I backed a 384-bit bus around 300-350mm2 but that looks to be too much now that the latest leak suggests a 256-bit bus. I guess it's possible that with better compression even a 256-bit bus won't be too bad, but bandwidth must be starting to at least put a top limit on performance. GDDR5X is not ready for these cards...

I do feel that if Nvidia doesn't beat the 980 Ti by 25%, they'll struggle to find takers like they have been getting on 28nm (hence why I believed the 384-bit bus was likely). I can't see 980 buyers swapping for 10% over a 980 Ti regardless of power consumption tbh.

 

[Silverforce11]

That's what GP100 is for.

They have mid-range Teslas as well.

GF104: 332 mm^2
GK104: 294 mm^2
GM204: 398 mm^2

I expect GP104 to fit somewhere into this range. If it was substantially smaller than 300 mm^2, that would be unprecedented.

Polaris 10 is a Pitcairn-sized chip; its Nvidia counterparts would be the 6-series, not the 4-series.

Exactly. Basically AMD's GP104 competitor is Vega 11.

Polaris 11 and 10 is to fight for notebook wins and offer great bang for buck to entry & mainstream gamers.

 

[Qwertilot]

Also, if NV can ship a '980ti + 50%' sort of chip, they'll be able to charge almost whatever they want even if it is only formally mid range!
(And sell a lot of them regardless. Like the 980 but more so.).

The eventual GP104 will surely be round that sort of level? The Titan is going to be around a 2x increase, and not much in between, so there really will be a crazy performance gap otherwise.

The only real question is going to be about whether its actually rationally possible in a technical sense.

One good thing for AMD is that delaying their GP104 equivalent will give Vega11 HBM2 which could quite plausibly give them a performance edge until Volta rolls round.

 

[Silverforce11]

Also, if NV can ship a '980ti + 50%' sort of chip, they'll be able to charge almost whatever they want even if it is only formally mid range!
(And sell a lot of them regardless. Like the 980 but more so.).

The eventual GP104 will surely be round that sort of level?

If it's ~300 - 350 mm2, no. Pascal is very likely to be Maxwell with better mix mode compute throughput.

GP104 at those sizes, will have performance around Titan X, minus any uarch gains. If I had to guess, it would be ~Titan X +20%, due to some uarch and some clock speed.

GP100 will sacrifice some perf/mm2 and perf/w for compute, in gaming performance, expect ~Titan X +60-70%. However, such a big chip with that many TFlops and 1/3 FP64, it would be one heck of a beast as a Tesla SKU.

 

[Silverforce11]

How GP104 will stack up vs Polaris 10.

I still only expect Polaris 10 to be ~390X performance in most scenarios, it's major advantage, that was showcased in Hitman demo, are related to scenes that would cause a major FPS dip.

This is where some of the new GCN features will kick in, lifting performance in bottleneck scenarios.

Overall, GP104 will easily be faster than Polaris 10, however, P10 will tank less in demanding scenes.

 

[PPB]

Lovely how we haven't heard jack about GP104's parametric specs, neither any of the other Pascal dies, yet people are already predicting it's die size, how it will perform overall and even already claiming it will be more performant that the bigger Polaris chip.

The ability to give form to what probably at this point is still hot air as NOTHING has been divulged by Nvidia besides the usual "it will be better than last generation GPUs! Just wait and see!" PR slide material is really astounding on this subforum.

And so as a reminder, AMD made a 439mm2 die that competed head on with Nvidia's 561mm2 die, only to bring a 398mm2 die to compete with AMD's one year later with profound architectural compromises at stake (gimped Xbar, no AC capabilites, totally crippled FP64). AMD, even tho improving GCN altogether with Polaris and that comes with a higher transistor cost, is betting on lower shader count to give performance between Hawaii and Fiji. Nvidia has to start getting back all the stuff they took out with Kepler and Maxwell to improve perf/watt, in order to regain SP perf and specially DP FP. If we have to make bets, mine is Nvidia this round will sacrifice gaming perf/mm2 in order to make up for the sacrifices made with Maxwell, unless they want to compete in the HPC market with only GP100 dies and place the burden on the die that will probably yield less but have the biggest margins, while AMD will probably keep similar values as Hawaii with the corrections applied with the new process shrinkage in die size. Fiji was a totally unbalanced design and it's bottlenecks usually put it nearer than Hawaii that AMD would hope, so I can't really extract meaningful information from that experience and even AMD would refrain from making the same mistakes seen in Fiji's layout again.

 

[Silverforce11]

@PPB

I like how you make estimates too, without any info.

See how that works?

It's just educated guesses at this point, from everyone. I have my reasons to think GP104 won't be a small chip like Polaris 10.

 

[Vesku]

@PPB

I like how you make estimates too, without any info.

See how that works?

It's just educated guesses at this point, from everyone. I have my reasons to think GP104 won't be a small chip like Polaris 10.

Yeah, I'm going by Nvidia needing 1080 to be at least close to 980ti performance.

 

[dark zero]

If this is true, there goes the only chance of AMD to catching up nVIDIA.

 

[PPB]

@PPB

I like how you make estimates too, without any info.

See how that works?

It's just educated guesses at this point, from everyone. I have my reasons to think GP104 won't be a small chip like Polaris 10.

The difference between my guesses and yours is that mine are based on previos data and the only information given by both AIBs. No one gave you die size data about Pascal, yet you make bold claims about how it will be bigger than P10.

We know AMD is redesigning most of the logical blocks inside their GPU? Yep. They even made a shiny slide explaining it.
We know those redesigns are aimed at more throughput? Yep. Raja talked about it
Do we know what "more throughput" means regarding transistor cost? Yep. This we can infer applying common sense and looking a bit in xPU design decision history regarding throughput and transistor cost.

On the other hand:

We know Nvidia is claiming higher FP DP performance for Pascal? Yep. There is a slide floating around claiming such.
We know what that implies in transistor cost? Yep. Like the last of the above points, common sense and examples regarding what transistor cost lies in upping your SP/DP FP ratio in xPUs.


You have no basis to say how big GP104 but what Nvidia has done on previous nodes. And guess what? Previous nodes had a far more forgiving price per wafer per parametric specs (both logic and memory cell's size) for both companies at the start of said nodes, which allowed totally different die size's spectrum for a new product stack on a new node. The people that can't grasp how this new product stack introduced roughly at next mid year will completely go against past experience because of the innate cost of 14/16nm FF wafers compared to the other nodes' starting prices and electrical behavior at certain parameters of die size, clock and voltage targets.

I talk about the little information provided, you talk about extrapolations of past experiences with a total disregard of actual context for this new node. You aren't far away from the people already dismissing 14nm LPP and hyping 16nm FF+ just because Apple's experience with 16nm FF and 14nm LPE.

If that level of whishy-washyness floats your boat, by all means. But dont try to compare my predictions with yours. As mine was an attempt to bring a counter-point with what we actually do know about this new product stack to your predictions which yet have to have some info to be backed up.

PS: Vesku's post is a clear example of what I'm talking about. We neither know if GP104 will become GTX 1080, and neither that Nv's desires will align with this new node's transistor cost considering the design decisions WE DO know for Pascal.

 

[jpiniero]

Lovely how we haven't heard jack about GP104's parametric specs, neither any of the other Pascal dies, yet people are already predicting it's die size, how it will perform overall and even already claiming it will be more performant that the bigger Polaris chip.

It is a guess, yes. But it makes sense that nVidia would want full GP104 to be enough faster than 980 Ti to make it worthwhile for e-penis people and people with Big Kepler while cheap enough given how expensive this node is that they can offer a worthwhile upgrade to 970 users at similar power consumption with a "small" price increase.

I'm not seeing the desire as much from AMD to be #1, especially when they seem to be focusing on getting mobile OEM wins.

And so as a reminder, AMD made a 439mm2 die that competed head on with Nvidia's 561mm2 die

And failed miserably, even with Hawaii having more bandwidth and 1/2 DP.

, only to bring a 398mm2 die to compete with AMD's one year later with profound architectural compromises at stake (gimped Xbar, no AC capabilites, totally crippled FP64).

And sold a ton of cards doing so.

AMD, even tho improving GCN altogether with Polaris and that comes with a higher transistor cost, is betting on lower shader count to give performance between Hawaii and Fiji.

Which is perfectly fine given that it should be competitive with the cut GP104.

Nvidia has to start getting back all the stuff they took out with Kepler and Maxwell to improve perf/watt, in order to regain SP perf and specially DP FP.

No they don't.

they want to compete in the HPC market with only GP100 dies and place the burden on the die that will probably yield less but have the biggest margins, while AMD will probably keep similar values as Hawaii with the corrections applied with the new process shrinkage in die size.

This is fine, nVidia has HPC contracts who will take full and cut dies. Yields at TSMC should be Not Awful even for a ~450 mm2 die either given how long the node has been in production. Super expensive though.

Fiji was a totally unbalanced design and it's bottlenecks usually put it nearer than Hawaii that AMD would hope, so I can't really extract meaningful information from that experience and even AMD would refrain from making the same mistakes seen in Fiji's layout again.

Fiji was a mistake. They should not have bothered.

 

[Vesku]

You have no basis to say how big GP104 but what Nvidia has done on previous nodes. And guess what? Previous nodes had a far more forgiving price per wafer per parametric specs (both logic and memory cell's size) for both companies at the start of said nodes, which allowed totally different die size's spectrum for a new product stack on a new node. The people that can't grasp how this new product stack introduced roughly at next mid year will completely go against past experience because of the innate cost of 14/16nm FF wafers compared to the other nodes' starting prices and electrical behavior at certain parameters of die size, clock and voltage targets.

We can draw on historical Nvidia business to say they are unlikely to want to deliver a 1080 that performs only about equal or less than a 980, certainly not on purpose. How would they even sell that given Maxwell is quite reasonable in power draw? I think they'll need closer to 300mm2 than 200mm2 to reach that performance. A bit like 580 to 680. If that 232mm2 is accurate for AMD I think it would be an impressive feat of engineering if it performs like Fury X/980 Ti rather than 10-20% below.

Raja is talking up price/perf and "affordable VR GPU". Nvidia's JHH has mainly focused his time promoting Pascal on industries that pay more than consumers for their GPUs. Best guess from this is that AMD will have their own GTX 970 moment with an affordable Polaris 10. Nvidia likely faster top 2016 SKU but they won't be talking up price/perf and they may take a month or three later than Polaris 10 to reach decent shipping volume.

Unless Nvidia is going to do something a bit unexpected and only launch a sub 260mm2 1070/60 for 2016, then I could see AMD 232mm2 having a shot at a tie or win.

 

[Silverforce11]

@PPB

See, that's great, back up your guesses with logic. Always good to see.

I'll stick with my own, that history is a good prediction for the future.

You may have missed it, but all those points you raised about Polaris with GCN improvements, I've already mentioned several times in this and other threads. Likewise for Pascal's focus on compute and FP64 (which is absent in mid-range chips like GP104, it's still purely a gaming chip).

The only guess that isn't backed by more solid evidence, is my die size estimate (based on historic precedent). Of which, I think it will be much bigger than Polaris 10's 232mm2.

 

[Silverforce11]

We can draw on historical Nvidia business to say they are unlikely to want to deliver a 1080 that performs only about equal or less than a 980. If their 2016 GDDR5 chip is similar size to AMD's, if 232mm2 is accurate, I don't think they will reach that target, therefore I'm guessing closer to 300 than 200. A bit like 580 to 680. Again, if that 232mm2 is accurate for AMD I think it would be an impressive feat of engineering if it performs like Fury X/980 Ti rather than 10-20% below.

Comparing NV's x04 chips from several generations, it has always been a huge leap in performance compared to previous gen big-chips. This means GP104 has to be significantly faster than GM200/Titan X.

It needs to be big to do it.

Why are we suddenly expecting them to go backwards?

Let's face it, Polaris 11 and 10 are low-end and mainstream. They aren't in the same class, it's more a GP107 and 106 competitor.

This is why AMD has Vega 11 and 10. Vega 11 is the real GP104 competitor and Vega 10 will bring the fight to GP100.

 

[jpiniero]

Why are we suddenly expecting them to go backwards?

Because this node is expensive.

This is why AMD has Vega 11 and 10. Vega 11 is the real GP104 competitor and Vega 10 will bring the fight to GP100.

Vega 10 is the GP100 competitor, yes. Vega 11 is the competitor to the theoretical GP102.

 

[Silverforce11]

Because this node is expensive.

Vega 10 is the GP100 competitor, yes. Vega 11 is the competitor to the theoretical GP102.

It's about margins. Unless the node has terrible yields, it doesn't matter as GPU prices have been going up (which I am sure we've all noticed). NV also sells x04 chips as Teslas, focused on FP32 and perf/w factors (they call it deep learning category, or whatever new fancy marketing term).

GP102 is about as theoretical as unicorns are.

 

[jpiniero]

(they call it deep learning category, or whatever new fancy marketing term).

Deep learning is a big deal these days. I think you can get away with FP16 on it too.

 

[Vesku]

It's about margins. Unless the node has terrible yields, it doesn't matter as GPU prices have been going up (which I am sure we've all noticed). NV also sells x04 chips as Teslas, focused on FP32 and perf/w factors (they call it deep learning category, or whatever new fancy marketing term).

GP102 is about as theoretical as unicorns are.

Given all 2016 FinFet GPUs appear to be GDDR5 I don't see Nvidia dropping a 350-450mm2 behemoth (for new node with uncertain yields at large die size) 1080 on the market. Unless they managed to pull off some 512-bit 8000MHz implementation or will launch in November-December with limited availability and GDDR5X.

 

[jpiniero]

Given all 2016 FinFet GPUs appear to be GDDR5 I don't see Nvidia dropping a 350-450mm2 behemoth (for new node with uncertain yields at large die size) 1080 on the market. Unless they managed to pull off some 512-bit 8000MHz implementation or will launch in November-December with limited availability and GDDR5X.

GP100 is HBM2. HBM2 is available right now, not in volume of course but that should be enough to fulfill the HPC contracts and perhaps a super expensive Titan.

 

[jpiniero]

Let's face it, Polaris 11 and 10 are low-end and mainstream. They aren't in the same class, it's more a GP107 and 106 competitor.

Oh and Polaris 10 is High End, Polaris 11 is mainstream. Below that are 28 nm rebrands (Weston and Banks)

High DP ~ GP100 ~ Vega 10
Extreme ~ GP102? ~ Vega 11
High End ~ GP104 ~ Polaris 10
Mainstream ~ GP106 ~ Polaris 11
Low End ~ GP107 ~ Weston?
Poor OEMs ~ GP108? + 28 nm rebrands? ~ Banks?

 

[Silverforce11]

Oh and Polaris 10 is High End, Polaris 11 is mainstream.

I doubt it will priced as such, not after all AMD's commotion about bringing great perf per dollar with Polaris.

They can really give a good bang to the market if they price it where it belongs, entry and mainstream.

 

[gamervivek]

According to some random german guy in a forum...

According to the open source driver that AMD released from which he has extracted the ip table for the new chips. I wasn't going to wait for him to post it to beyond3d.

https://forum.beyond3d.com/posts/1877668/


However, AMD's John Bridgman says there are quite a few changes in the hardware that are not reflected in the ip blocks(gfx one that we are concerned with) and so polaris is going to be more than just Tonga/Fiji cores with more ROPs.

I was initially surprised how few software-visible changes there were in Polaris but AFAICS the original slides seem to be about right.

https://semiaccurate.com/forums/showpost.php?p=258568&postcount=524

The high price of Baffin XT on zauba(60% of Fiji XT) makes more sense then.

 

[3DVagabond]

NV doesn't operate on the same basis as AMD.

They have Teslas and Quadros to make the bulk of their profits. These markets expect big performance gains to justify upgrades of major infrastructure.

You know they just released a 24GB GM200 Quadro?
http://www.techpowerup.com/221113/nvidia-unveils-the-quadro-m6000-24gb-graphics-card.html

Seems strange to release a new flagship Quadro if 16nm is close to ready.

 

[kagui]

You know they just released a 24GB GM200 Quadro?
http://www.techpowerup.com/221113/nvidia-unveils-the-quadro-m6000-24gb-graphics-card.html

Seems strange to release a new flagship Quadro if 16nm is close to ready.

no if you think the next release is about perf/watt