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

[Silverforce11]

http://www.tweaktown.com/news/51959/polaris-exciting-gpu-looking/index.html

An editorial from TT.

Polaris 10 is the Hawaii/390X killer at a much reduced TDP. Taking 390X+ performance down to mainstream "cheap" levels for the masses.

In essence, it's equivalent to GP106 with some tricks up it's sleeves. In complex scenes, it may punch well above it's weight, but in general, it's unlikely to be a Fury X or 980Ti slayer.

They obviously can't launch all the 14nm FF chips immediately, so attacking the low-end and mainstream with amazing perf/w is a sound strategy. Those notebook wins are a must and volume is good at around the $200 mark +/- $50.

In many ways, this is turning out not a surprise at all. Most of us saw this coming with 14/16nm, low-end, mid-range first, then later as volumes & yields improve further, we get big chips (Vega & GP100) out to play.

 

[Silverforce11]

HD7870 vs. 6970 comparison isn't valid:

Fury X is also very gimped with a poor front end setup. Tonga's front-end for 4K shaders is a recipe for fail. Hence, Fury X is what, 15% above 390X at 1080p? ~20% at 1440p.

Pretty weak.

I don't think Polaris 10 will beat Fury X overall, but really, 390X + 10% is awfully close at 1080/1440p.

 

[Pinstripe]

This should primarily attract the GTX 960/R9 380 users to upgrade. If AMD prices these cards well ($200-250), they should have a winner. Until Nvidia gets their GP106 out the doors, all they can do is slashing GTX 970/980 prices. Or, offer the GTX 1060 Ti at super low price too.

We shall find out blabla.

 

[RussianSensation]

Fury X is also very gimped with a poor front end setup. Tonga's front-end for 4K shaders is a recipe for fail. Hence, Fury X is what, 15% above 390X at 1080p? ~20% at 1440p.

Pretty weak.

I don't think Polaris 10 will beat Fury X overall, but really, 390X + 10% is awfully close at 1080/1440p.

I'd argue that people who bought a Fury X aren't even gaming at 1080 60Hz. Therefore, the comparison to Fury X at 1080p is more theoretical for Fury X users. Anyway, even if we go with that, look at the title of the thread:

10% over 390X = Fury, ways off 980Ti @ 1080p
http://www.techpowerup.com/reviews/ASUS/GTX_950/23.html

 

[Aristotelian]

But for me (to RS and others) the question remains - yes, it's good to release power efficient cards that have perf/watt. But I doubt those people buying video cards for 300 USD or so are measuring perf/watt. They are trying to measure perf/dollar. But if Nvidia does release 1080, 1070, 1060ti - with the 1080 being faster than a 980Ti, the 1070 being faster than a 970, and the 1060ti being perhaps slightly faster than a 970, where would Polaris 10 fit in, especially if Nvidia offers the 1080 for say, 599 or 649 USD (MSRP)? Perhaps I'm not understanding the dynamic here.

 

[maddie]

HD7870 vs. 6970 comparison isn't valid:

1) 6970 used far inferior IPC architecture with VLIW-4. 6970 was not a compute architecture. Fury X has the fastest Async compute horsepower out of any consumer GPU today. In DX12 games, Fury X starts to pull away from 290X and beats the 980Ti. Those 4000+ shaders are better utilized than under DX11. A lot more DX12 games will come out, further allowing Fury X to extend its lead over the 290X.

2) 6970 was not a 280W GPU with an AIO CLC (effectively the Fury X on air would likely be pushing close to 300W)

3) 6970 only had a die size or 389mm2, Fury X has 596mm2. Asking 232mm2 Polaris 10 to pull off a similar feat is much harder without a 30-40% GPU clock uplift over Fury X.

4) 6970 didn't have the most advanced per transistor/die space memory controller, Fury X does. HBM1 Fiji memory controller is more efficient than any AMD memory controller in the entire stack. It's why the even more efficient Vega uses HBM2, not GDDR5X.

5) 6970 didn't have almost 2X the memory bandwidth of 7870, Fury X appears to have that over 256-bit Polaris 10 GDDR5.

6) 6970 had lower clocks than 7870 too, while leaks for Polaris 10 show 850-1000mhz. Bad sign. Not 1 leak so far indicating Polaris 10 could have 1400-1500mhz clocks.

7) AMD decoupled the ROPs from the memory controllers in GCN which allowed a huge throughput in real world ROP performance. With the same # of ROPs, 7970 had > 50% higher pixel Fillrate.
http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/26

I don't expect Polaris 10's ROPs to be 50% faster than Fury X's.

Using 7870 vs. 6970 actually shows how much the cards are stacked against Polaris 10 to match the Fury X at 1440-4K. GCN was the biggest leap in a decade for AMD, surely far more than the leap GCN 4.0 will bring over Fury X.

And yet, we have this from Koduri directly during an interview.

Maybe he's wrong.

"
VB: Is that with a generation coming in 2016?
Koduri: Yes. We have two versions of these FinFET GPUs. Both are extremely power efficient. This is Polaris 10 and that’s Polaris 11. In terms of what we’ve done at the high level, it’s our most revolutionary jump in performance so far. We’ve redesigned many blocks in our cores. We’ve redesigned the main processor, a new geometry processor, a completely new fourth-generation Graphics Core Next with a very high increase in performance. We have new multimedia cores, a new display engine."

 

[3DVagabond]

HD7870 vs. 6970 comparison isn't valid:

1) 6970 used far inferior IPC architecture with VLIW-4. 6970 was not a compute architecture. Fury X has the fastest Async compute horsepower out of any consumer GPU today. In DX12 games, Fury X starts to pull away from 290X and beats the 980Ti. Those 4000+ shaders are better utilized than under DX11. A lot more DX12 games will come out, further allowing Fury X to extend its lead over the 290X.

2) 6970 was not a 280W GPU with an AIO CLC (effectively the Fury X on air would likely be pushing close to 300W)

3) 6970 only had a die size or 389mm2, Fury X has 596mm2. Asking 232mm2 Polaris 10 to pull off a similar feat is much harder without a 30-40% GPU clock uplift over Fury X.

4) 6970 didn't have the most advanced per transistor/die space memory controller, Fury X does. HBM1 Fiji memory controller is more efficient than any AMD memory controller in the entire stack. It's why the even more efficient Vega uses HBM2, not GDDR5X.

I see Maddie beat me to it.

5) 6970 didn't have almost 2X the memory bandwidth of 7870, Fury X appears to have that over 256-bit Polaris 10 GDDR5.

6) 6970 had lower clocks than 7870 too, while leaks for Polaris 10 show 850-1000mhz. Bad sign. Not 1 leak so far indicating Polaris 10 could have 1400-1500mhz clocks.

7) AMD decoupled the ROPs from the memory controllers in GCN which allowed a huge throughput in real world ROP performance. With the same # of ROPs, 7970 had > 50% higher pixel Fillrate.
http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/26

I don't expect Polaris 10's ROPs to be 50% faster than Fury X's.

Using 7870 vs. 6970 actually shows how much the cards are stacked against Polaris 10 to match the Fury X at 1440-4K. GCN was the biggest leap in a decade for AMD, surely far more than the leap GCN 4.0 will bring over Fury X.

Not disagreeing with any of that. One thing to point out though, AMD claims Polaris is the single biggest jump in performance ever for them. It could be the 28nm to 14ff or new uarch, or a combination. Most likely the latter.

I see Maddie beat me to it.

 

[maddie]

http://www.tweaktown.com/news/51959/polaris-exciting-gpu-looking/index.html

An editorial from TT.

Polaris 10 is the Hawaii/390X killer at a much reduced TDP. Taking 390X+ performance down to mainstream "cheap" levels for the masses.

In essence, it's equivalent to GP106 with some tricks up it's sleeves. In complex scenes, it may punch well above it's weight, but in general, it's unlikely to be a Fury X or 980Ti slayer.

They obviously can't launch all the 14nm FF chips immediately, so attacking the low-end and mainstream with amazing perf/w is a sound strategy. Those notebook wins are a must and volume is good at around the $200 mark +/- $50.

In many ways, this is turning out not a surprise at all. Most of us saw this coming with 14/16nm, low-end, mid-range first, then later as volumes & yields improve further, we get big chips (Vega & GP100) out to play.

I do wish we would start to see a range of performance estimates.

Surely we don't expect there to be a single Polaris 10 model. With Finfet's power consumption/clock range so much better than previous, the opportunity for many models is obvious. GDDR5X only adds to the performance variations possible.

I expect from a 390 to a FuryX in performance for the various models

 

[coercitiv]

10% over 390X = Fury, ways off 980Ti @ 1080p

One thing I don't understand: if we look at the relative die size between Polaris 10 and Hawaii on one side, GP104 and GM200 on the other side, we have a ratio of approx. 1:2 happening on a node jump with comparable characteristics. (200mm2 vs 400mm2 and 300mm2 vs 600mm2 respectively)

If we expect Polaris 10 to hardly beat (or match) Hawaii , why do we expect GP104 to be 30%+ faster than GM200?

 

[airfathaaaaa]

I do wish we would start to see a range of performance estimates.

Surely we don't expect there to be a single Polaris 10 model. With Finfet's power consumption/clock range so much better than previous, the opportunity for many models is obvious. GDDR5X only adds to the performance variations possible.

I expect from a 390 to a FuryX in performance for the various models

so they are basicly gonna fill every segment possible but doing so (if those cards have any good OC room) some cards will be obsolete

 

[Silverforce11]

One thing I don't understand: if we look at the relative die size between Polaris 10 and Hawaii on one side, GP104 and GM200 on the other side, we have a ratio of approx. 1:2 happening on a node jump with comparable characteristics. (200mm2 vs 400mm2 and 300mm2 vs 600mm2 respectively)

If we expect Polaris 10 to hardly beat (or match) Hawaii , why do we expect GP104 to be 30%+ faster than GM200?

Because Pascal will clock higher at a higher TDP.

AMD is aiming for perf/w so I expect it's out of the box to be lower clocked.

That's one way to look at it. If that's the case, there should be some nice OC potential like in the 7800 and 7900 series.

 

[Timorous]

HD7870 vs. 6970 comparison isn't valid:

1) 6970 used far inferior IPC architecture with VLIW-4. 6970 was not a compute architecture. Fury X has the fastest Async compute horsepower out of any consumer GPU today. In DX12 games, Fury X starts to pull away from 290X and beats the 980Ti. Those 4000+ shaders are better utilized than under DX11. A lot more DX12 games will come out, further allowing Fury X to extend its lead over the 290X.

2) 6970 was not a 280W GPU with an AIO CLC (effectively the Fury X on air would likely be pushing close to 300W)

3) 6970 only had a die size or 389mm2, Fury X has 596mm2. Asking 232mm2 Polaris 10 to pull off a similar feat is much harder without a 30-40% GPU clock uplift over Fury X.

4) 6970 didn't have the most advanced per transistor/die space memory controller, Fury X does. HBM1 Fiji memory controller is more efficient than any AMD memory controller in the entire stack. It's why the even more efficient Vega uses HBM2, not GDDR5X.

5) 6970 didn't have almost 2X the memory bandwidth of 7870, Fury X appears to have that over 256-bit Polaris 10 GDDR5.

6) 6970 had lower clocks than 7870 too, while leaks for Polaris 10 show 850-1000mhz. Bad sign. Not 1 leak so far indicating Polaris 10 could have 1400-1500mhz clocks.

7) AMD decoupled the ROPs from the memory controllers in GCN which allowed a huge throughput in real world ROP performance. With the same # of ROPs, 7970 had > 50% higher pixel Fillrate.
http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/26

I don't expect Polaris 10's ROPs to be 50% faster than Fury X's.

Using 7870 vs. 6970 actually shows how much the cards are stacked against Polaris 10 to match the Fury X at 1440-4K. GCN was the biggest leap in a decade for AMD, surely far more than the leap GCN 4.0 will bring over Fury X.

Why did you think I was relating the 6970 to Fury X? The 6970 is much closer to the 390X comparatively and the 7870 beat the 6970 by around 15-20%. If P10 can emulate the 7870 by being 15-20% faster than the 390X it puts it right in the ballpark of Fury X and 980Ti performance at 1440p.

AMD have done it before and I don't see why they can't do it again.

EDIT: Re-reading my previous post I see I did not make it clear that I thought the 6970 was similar to the 390X and that the GTX580 was similar to the Fury X and 980Ti in terms of comparisons vs 7870 and P10.

 

[maddie]

This should primarily attract the GTX 960/R9 380 users to upgrade. If AMD prices these cards well ($200-250), they should have a winner. Until Nvidia gets their GP106 out the doors, all they can do is slashing GTX 970/980 prices. Or, offer the GTX 1060 Ti at super low price too.

We shall find out blabla.

This is part of a previous post of mine. Apologies for posting again, but it's relevant to your speculation.

3] Using a die yield calculator, a die of 320mm^2 and defect of 0.2/cm^2 [claimed by Samsung and applied to TSMC], we get 55% full die/wafer.

4] I can't see them getting enough defective die to fill 1060Ti orders as at best they can harvest 45% of all produced.

Nvidia will have to cut perfectly good 320mm^2 Gp104 dies to compete with a 232mm^2 top Polaris model.

How will that affect pricing power?

 

[Aristotelian]

Because Pascal will clock higher at a higher TDP.

AMD is aiming for perf/w so I expect it's out of the box to be lower clocked.

That's one way to look at it. If that's the case, there should be some nice OC potential like in the 7800 and 7900 series.

And this is part of the question I posted further above. Performance per watt - all else being equal, is a valuable metric. For laptops, consoles, whatever - especially so. But if AMD plan to replace its current lineup with cards as powerful as the current lineup is, at a better TDP threshold, and at a lower price (and on this last one I don't see how it's possible given the wafer cost increase), how could they compete against the new Nvidia offerings that are meant to be faster than their current lineup at similar TDP thresholds (and perhaps more expensive than AMD's offerings then).

 

[maddie]

Because Pascal will clock higher at a higher TDP.

AMD is aiming for perf/w so I expect it's out of the box to be lower clocked.

That's one way to look at it. If that's the case, there should be some nice OC potential like in the 7800 and 7900 series.

AMD can play the higher clock/power game also.

If they do have good power control as they have hinted, then we really will have a wide range of performance in the various models. Some of which will be high perf/W and others will be high performance outright. Choose your poison. If they use GDDR5X in the top performance models, then these models are protected from being cannibalized by lower end models being overclocked.

During the CES 2016 Polaris 11 demo, AMD claimed that there were many power optimizations still to be implemented. This is in addition to what was witnessed in the demo.

 

[coercitiv]

Because Pascal will clock higher at a higher TDP.

AMD is aiming for perf/w so I expect it's out of the box to be lower clocked.

Lower clocked than Hawaii?! We have a significant node jump advantage in front of us, and you think AMD felt the need to improve on that by further limiting operating clocks on Polaris 10? I find that very hard to believe.

 

[Silverforce11]

Lower clocked than Hawaii?! We have a significant node jump advantage in front of us, and you think AMD felt the need to improve on that by further limiting operating clocks on Polaris 10? I find that very hard to believe.

No, lower than Pascal. P100 has 1480mhz boost clocks, very high for a Tesla card. A GTX will likely clock 100mhz higher than that. That's how they can get some nice performance gains. GP104 is a true mid-range chip with the TDP to match. Polaris 10 is mainstream. It's not designed for high TDP as that does not belong in that segment and the goal of perf/w for notebooks.

And this is part of the question I posted further above. Performance per watt - all else being equal, is a valuable metric. For laptops, consoles, whatever - especially so. But if AMD plan to replace its current lineup with cards as powerful as the current lineup is, at a better TDP threshold, and at a lower price (and on this last one I don't see how it's possible given the wafer cost increase), how could they compete against the new Nvidia offerings that are meant to be faster than their current lineup at similar TDP thresholds (and perhaps more expensive than AMD's offerings then).

They will compete with Vega. You guys are acting like Polaris is the only thing AMD has. On their Roadmap, Vega falls into Q4 2016 and Q1 2017.

AMD can play the higher clock/power game also.

If they do have good power control as they have hinted, then we really will have a wide range of performance in the various models. Some of which will be high perf/W and others will be high performance outright. Choose your poison. If they use GDDR5X in the top performance models, then these models are protected from being cannibalized by lower end models being overclocked.

During the CES 2016 Polaris 11 demo, AMD claimed that there were many power optimizations still to be implemented. This is in addition to what was witnessed in the demo.

Sure, but unless they go with an GDDR5X SKU, it may simply be bandwidth starved so that clocking it much higher won't yield much gains. We can't rule out a 5X option though, so we'll see.

 

[ShintaiDK]

If we expect Polaris 10 to hardly beat (or match) Hawaii , why do we expect GP104 to be 30%+ faster than GM200?

Unless Polaris uses GDDR5X, that alone will give it much less bandwidth, most likely around 192GB/sec or close to half of Hawaii. While GP104 with GDDR5X will have 50GB/sec faster memory than GTX980TI with a total of 384GB/sec.

 

[coercitiv]

No, lower than Pascal.

I ask why is it that people think Polaris 10 will hardly outmatch Hawaii considering density/power improvements on the new node, and your answer is because Pascal will clock 1500Mhz+?

 

[coercitiv]

Unless Polaris uses GDDR5X, that alone will give it much less bandwidth, most likely around 192GB/sec or close to half of Hawaii. While GP104 with GDDR5X will have 50GB/sec faster memory than GTX980TI with a total of 384GB/sec.

So this entire line of reasoning is based on Polaris not using GDDR5X?

 

[mkmitch]

I will look closely at both of the companies new releases to see if it makes sense to upgrade from a gtx 770 4mb. Not sure 1920x1080 gaming I should.It will be interesting to see the numbers. Not concerned about price only performance.

 

[ShintaiDK]

So this entire line of reasoning is based on Polaris not using GDDR5X?

GDDR5X sure would help to begin with. But we also have to see clocks etc. GP100 clocks up to 40% higher than GM200 as stock. So we can assume in favour of GP104 vs GM200 that its somewhat similar.

 

[guskline]

We kind of know that on May 6-7 we will see much more about Pascal during the Battlefield 5 event.

What about Polaris?

End of the month in Taipei ??

 

[beginner99]

I ask why is it that people think Polaris 10 will hardly outmatch Hawaii considering density/power improvements on the new node, and your answer is because Pascal will clock 1500Mhz+?

So this entire line of reasoning is based on Polaris not using GDDR5X?

Yes, there isn't a magic bullet to get same performance with a fraction of the bandwidth. Hawaii lacks color compression from Tonga which is further improved for Polaris plus maybe some additional tweaks and Hawaii wasn't really bandwidth starved. That makes it somewhat reasonable to believe Polaris 10 can reach Hawaii performance with half the bandwidth but going even further seems unlikely. Would be a huge improvement and kind of question the need for HBM in this generation.

If AMD came up with such a huge improvement, HBM would be useless as they could get very far with a 384 or 512 bit GDDR5X bus.

The surprise could be that Polaris is GDDR5x only and hence the whole memory argument is void as then it would not be bandwidth starved. But as far as we know GDDR5x is not ready yet, eg. for a June launch.

 

[coercitiv]

But as far as we know GDDR5X is not ready yet, eg. for a June launch.

That's not what I read on the Nvidia threads, GDDR5X is supposedly ready for their imminent May/June launch.