(GURU3D RUMOR) AMD Polaris 10 GPU To Offer Near 980 Ti Performance For 299 USD?

[LurchFrinky]

Currently there is 58.02% (76 votes) for "The rumor is too optimistic. I expect less performance in that price range"

What performance do you guys expect from the top P10 SKU ? 390x, 390, 380x, 380 or even 370 ? ()

The poll is asking for the performance @$300, to which I would guess around 390X.
However, I think the top P10 SKU can reach 980ti performance with GDDR5X, but the price will have to climb higher - maybe $400?
I don't know if this top SKU will be released in the initial wave.

 

[maddie]

Going by the current rumors:

1080 = high end market
1070 = mainstream market
1060 = mid-range

Polaris will compete with the 1060

Vega will compete with 1080 next year but NVIDIA will have a new card as well.

Early next year? Which card is this? GP100 or GP102?

How many think that info about Vega becomes a lot more available soon after GP104 cards launch?

 

[Game_dev]

Early next year? Which card is this? GP100 or GP102?

How many think that info about Vega becomes a lot more available soon after GP104 cards launch?

They are already making Tesla's. So it would be some variation of that maybe slightly improved after this much time. Might be called titan

 

[maddie]

They are already making Tesla's. So it would be some variation of that maybe slightly improved after this much time. Might be called titan

What I failed to clarify is if it's early next year for Nvidia because I think Vega will be very early in the year.

 

[Silverforce11]

Going by the current rumors:

1080 = high end market
1070 = mainstream market
1060 = mid-range

Polaris will compete with the 1060

Vega will compete with 1080 next year but NVIDIA will have a new card as well.

Sure, it's priced that way.

But it's still a mid-range chip.

Btw, you got it mixed around buddy.

The order is mainstream goes after mid-range, ie. mainstream is slower, example = 960, mid-range = 970.

Vega is AMD's mid-range and high-end chip. Equivalent to GP104 and GP100.

 

[Silverforce11]

Case 1: AMD tesselation performance HD5000 series

Case2: Nvidia shader and compute switch penalty [async shaders]

Exercise: Explain how software solved the respective deficiencies.

This one is pretty easy.

1. AMD implements x16/x32 tessellation override in drivers. So x64 tessellation usage cannot gimp it.

2. Bribe developers to not use async compute, and to use less compute shader effects in their games. Replace it with tessellation based effects, like GodRays, HFTS (the new shadow tracing in The Division using... tessellation).

You can do a lot in software to reduce the weakness of your uarch, especially since DX11 is driver reliant, AMD & NV can intercept the GPU calls and replace any shader they deem unoptimized with their own equivalent variant.

DX12, I believe they lose this ability. So if devs use an unoptimized shader, there's nothing they can do. Thus for DX12, #2 bribe/working closely with devs is more important else the game can be quite broken. Example: Quantum Break on NV GPUs. NV should have sponsored that title.

 

[Game_dev]

What I failed to clarify is if it's early next year for Nvidia because I think Vega will be very early in the year.

I imagine they will release it just after Vega like the 780ti. Spoil another launch again. It depends on hbm availability and whether they can keep up with the Tesla demand.

 

[Game_dev]

Sure, it's priced that way.

But it's still a mid-range chip.

Btw, you got it mixed around buddy.

The order is mainstream goes after mid-range, ie. mainstream is slower, example = 960, mid-range = 970.

Vega is AMD's mid-range and high-end chip. Equivalent to GP104 and GP100.

Semantics on mainstream ;-p but how I listed it looks like how it will pan out. AMD is abandoning the high end to focus on consoles. In the long run that's a smart move.

 

[Silverforce11]

Semantics on mainstream ;-p but how I listed it looks like how it will pan out. AMD is abandoning the high end to focus on consoles. In the long run that's a smart move.

They aren't abandoning the high-end. You are ignoring Vega.

For the mid-range and high-end Vega, AMD at best, can be 3 months behind, at worse, 6 months behind. According to their own roadmap.

AMD is targeting low-end with Polaris 11, mainstream with Polaris 10. NV is targeting mid-range with GP104 and will price the 1080 at high-end $$.

AMD's strategy hopes to get the high volume segment on PC and get Polaris 11 and 10 into notebooks (like the Apple & HP wins). NV aims to cash in with selling GP104 for high $$.

Really, the idea that the high-end is being abandoned is silly when they have already published public info on upcoming Vega GPUs with a big leap in performance and perf/w compared to Polaris. It's just the staggered launch we have seen before over and over. Some SKUs have to go first and AMD/NV choose different market segments to prioritize target.

 

[tential]

Semantics on mainstream ;-p but how I listed it looks like how it will pan out. AMD is abandoning the high end to focus on consoles. In the long run that's a smart move.

They didn't abandon the high end just like Nvidia didn't abandon the low end. They just both prioritized different markets. I don't think it's a smart move at all. I think Nvidia will capture the high end and mindshare the low end afterwards

 

[Game_dev]

I'm not ignoring Vega, but it won't be here until 2017. By then it may not be a high end chip compared to whatever NVIDIA has to offer. If they truly wanted to compete they would have a high end chip right now.

 

[swilli89]

They aren't abandoning the high-end. You are ignoring Vega.

For the mid-range and high-end Vega, AMD at best, can be 3 months behind, at worse, 6 months behind. According to their own roadmap.

AMD is targeting low-end with Polaris 11, mainstream with Polaris 10. NV is targeting mid-range with GP104 and will price the 1080 at high-end $$.

AMD's strategy hopes to get the high volume segment on PC and get Polaris 11 and 10 into notebooks (like the Apple & HP wins). NV aims to cash in with selling GP104 for high $$.

Really, the idea that the high-end is being abandoned is silly when they have already published public info on upcoming Vega GPUs with a big leap in performance and perf/w compared to Polaris. It's just the staggered launch we have seen before over and over. Some SKUs have to go first and AMD/NV choose different market segments to prioritize target.

Still, I get the feeling AMD overstepped in its risk mitigation for 14nm finfet. There is no way they would have designed their first round of chips topping out at 230mm if they had known yields would be good.

Why in the world shoot themselves in the foot like this? Maybe they just really don't want to use GDDR5X, which, now that I think about it gives just about the only plausible reason they wouldn't have readied a higher end chip.

Something like a 320mm ~3200 GCN4 SPs. Why in the world does AMD ALWAYS design a higher performing/mm die then release a smaller chip that merely competes? They already spent the resources on the architecture, why not just spend a little extra and dominate from low end to upper mid-range from the get-go? They force themselves to do this bullshit value proposition when its pretty obvious gamers are about the worst value shoppers out there; they want the 1337 framez and bad@ss box art.

A SKU between Polaris 10 and Vega could have costed maybe $25 million to develop and netted them hundreds of millions if they executed well at the $500-$600 price point. Sigh.. Obviously AMD has people a lot smarter than I but I have to wonder why they continue to let nvidia bring larger dies to market first ESPECIALLY when AMD has such a close relationship with Global foundries and knew on the front end that GF would be leveraging Samsung's process tech. Once they release Vega nvidia will have the still much larger GP100(102) ready to go.

Just frustrating watch AMD design better chips but lets the competition just brute force it via die size every generation.

 

[Game_dev]

They didn't abandon the high end just like Nvidia didn't abandon the low end. They just both prioritized different markets. I don't think it's a smart move at all. I think Nvidia will capture the high end and mindshare the low end afterwards

NVIDIA is still selling boatloads of 750ti's, they can easily carry the low end with the new low power 950.

 

[renderstate]

Replace it with tessellation based effects, like GodRays, HFTS (the new shadow tracing in The Division using... tessellation).
.

HTFS uses HW conservative rasterization, which is unsupported by AMD, not tessellation.

Also there is nothing wrong with using tessellation to reasonable levels. But of course, tessellation is considered a gift from the devil or something on this forum.

If Polaris has fast(er) tessellation and conservative rasterization I imagine it will all be forgotten and/or forgiven

 

[Silverforce11]

I'm not ignoring Vega, but it won't be here until 2017. By then it may not be a high end chip compared to whatever NVIDIA has to offer. If they truly wanted to compete they would have a high end chip right now.

What NV has to offer beyond GP104 is GP100 with a LOT of FP64 cuda cores that are doing nothing in games. They are a waste of die space and TDP as a GTX.

AMD's FP16, 32 & 64 is unified by design. Hawaii actually has half FP64 ratio that GP100 has, without wasting SP dedicated to FP64.

Vega with HBM2 is going to beat GP100 easily in gaming performance. NV will need a dedicated gaming chip, aka GP102 to actually compete with it.

Just look and use your logic:

Vega is set to improve performance and perf/w over Polaris.

Do you think GP100 is going to improve perf/w over GP104 with so many useless FP64 CC?

They are hitting the 300W TDP limit for a professional Tesla product already in the P100. It's being pushed to the max. It has a high chance of being worse than the original 680 vs Titan comparison, or the 980 vs Titan X.

So it's quite clear to me, unless NV has GP102, without those useless FP64 CC, they are in fact abandoning the enthusiast segment because Vega 10 is going to blow it away.

 

[Silverforce11]

HTFS uses conservative rasterization, not tessellation.

My bad, was it VXAO then? It tanked performance in TR massively.

 

[SimianR]

What I'm not getting is the absolute need for HBM on VEGA. I suppose it's more complicated with the memory integrated onto the chip itself, but I'm surprised they aren't going to launch a larger chip than polaris 10 with GDDR5 or GDDR5X (when it becomes available). I think it's a mistake to completely abandon the high end for the remainder of 2016.

 

[renderstate]

My bad, was it VXAO then? It tanked performance in TR massively.

Fast voxelization also requires conservative rasterization. Tessellation has nothing to do with it.

 

[Silverforce11]

Still, I get the feeling AMD overstepped in its risk mitigation for 14nm finfet. There is no way they would have designed their first round of chips topping out at 230mm if they had known yields would be good.

Why in the world shoot themselves in the foot like this? Maybe they just really don't want to use GDDR5X, which, now that I think about it gives just about the only plausible reason they wouldn't have readied a higher end chip.

It's not just yields. It's volume, production capability, wafer assignments.

What is AMD making on 14nm FF now?

You can bet it's APUs for Sony's PS4K, Xbox Next and the NX.

What else? Well, a LOT of Polaris 11 and 10 for Apple's upcoming refresh.

It's a better use of wafers to make smaller chips, as you can get more chips and more functional chips out of them to meet demand. AMD simply cannot launch a big Vega at the same time as these productions, IMO.

Compare this to NV, we see the same. Most of the wafers will go to GP100 for expensive Teslas, the rest goes to GP104 and GP106. Why don't they launch GP106 at the same time? Why don't they launch GTX GP100 soon? They can't. TSMC 16nm is only scheduled to get up to ~20% THIS YEAR. They are infact behind, because LAST YEAR they claimed to hit ~30% this year, but their recent conference, their CEO revised the figure down.

We all know TSMC wafers are used by many companies, not just NV.

 

[Silverforce11]

Fast voxelization also requires conservative rasterization. Tessellation has nothing to do with it.

It was one of them, you should recheck it. IIRC, NV's slides showed tessellated ray tracing to get approximate shadow casting.

 

[Game_dev]

And why wouldn't NVIDIA already have a GP102 ready to go by next year? 18 months from now they could even have Volta or whatever ready.

 

[JDG1980]

NVIDIA is still selling boatloads of 750ti's, they can easily carry the low end with the new low power 950.

There's no way that the GTX 950 will be competitive with 14nm FinFET products.

 

[maddie]

My bad, was it VXAO then? It tanked performance in TR massively.

Recommended a member [India] contact you by PM about used stock 290X. Hope you don't mind.

 

[renderstate]

It was one of them, you should recheck it. IIRC, NV's slides showed tessellated ray tracing to get approximate shadow casting.

Perhaps you should recheck it

While you can use tessellation in both cases the key feature that allows ray traced shadows (HTFS) and voxelization is conservative rasterization. Keep in mind that you can do CS in SW with a geometry shader, but AMD is not very good at geometry shaders. I bet Polaris will support CS in HW since it's part of the new DX12 features.

I can only imagine how happy people would be if AMD had introduced ray traced shadows in real-time in complex scenes. But once again, since in this case it's an NVIDIA innovation, well, it must be bad/evil.. worse, tessellated! Ahh.. Pavlov could have a written an essay based on certain automatic responses.

 

[Silverforce11]

While you can use tessellation in both cases the key feature that allows ray traced shadows (HTFS) and voxelization is conservative rasterization.

Right, see, I knew I saw tessellation being used for ray tracing shadow in one of NV's new GameWorks features.

So let's return back to the original post before you get all grand standing again..

IHVs can indeed bypass hardware weaknesses with software approaches.