[Benchlife] R9 480 (Polaris 10 >100w), R9

[dzoni2k2]

R7 470
Capacity - 4GB
Memory type - GDDR5
Performance level - R9 380
TDP - 60W no pcie connector required
Price - $149.

R9 480
Capacity - 8GB
Type - GDDR5
Performance level - R9 390
TDP - under 100W single 6 pin
Price - $249.

Make it happen AMD. Balls in your court now.

You are probably very close price wise.

 

[swilli89]

R7 470
Capacity - 4GB
Memory type - GDDR5
Performance level - R9 380
TDP - 60W no pcie connector required
Price - $149.

R9 480
Capacity - 8GB
Type - GDDR5
Performance level - R9 390
TDP - under 100W single 6 pin
Price - $249.

Make it happen AMD. Balls in your court now.

You're spot on. POSSIBLY $299 for a top end SKU Polaris 10, 480X and a downclocked/chopped down Polaris 10, 480 at $249.

nvidia's 1080 may have a 10% advantage yet people will still pay twice as much for no impact on playability. In DX12 I suspect the 480X will trump it at a lower cost and better power efficiency.

Looks like Radeon 4870 vs 280 all over again, will consumers purchase more logically this time around? 90% of the performance for half the cost?

 

[FatherMurphy]

http://videocardz.com/59206/amds-official-gpu-roadmap-for-2016-2018

I'll leave this here. It's not clear where WhyCry found this alleged roadmap for AMD.

 

[antihelten]

You're spot on. POSSIBLY $299 for a top end SKU Polaris 10, 480X and a downclocked/chopped down Polaris 10, 480 at $249.

nvidia's 1080 may have a 10% advantage yet people will still pay twice as much for no impact on playability. In DX12 I suspect the 480X will trump it at a lower cost and better power efficiency.

I agree that mohit is probably very close to the real number (at least as far as performance and price goes), but I think a third top GPU from AMD would probably be $350 and not $300, as the later is to close to the $250 SKU.

Also people are currently speculation that a GTX 1080 will be somewhere around a stock 980 Ti up to a factory overclocked 980 Ti. If that is the case then it would be 35-60% faster than the R9 480 mohit listed, and as such if it only had a 10% advantage over a theoretical 480X then said 480X would be 20-45% faster than the 480, and then it definitely wouldn't sell for $300, with the 480 at $250. $350 is probably realistic at the low end (20% higher performance),but all the way up to $450 isn't unrealistic if the performance is at the higher end (45% faster). That might still be cheaper than Nvidia, but not half the price.

http://videocardz.com/59206/amds-official-gpu-roadmap-for-2016-2018

I'll leave this here. It's not clear where WhyCry found this alleged roadmap for AMD.

He probably found it here

 

[Glo.]

Unless R7 470 has 1536 GCN cores, it will be impossible for that chip to achieve R9 380 performance level.

1280 GCN cores clocked at 1150 MHz brings us to 2.9 TFLOPs of compute power.

R9 380 has 3.5 TFLOPs and 256 Bit memory bus.

 

[Kenmitch]

1280 GCN cores clocked at 1150 MHz brings us to 2.9 TFLOPs of compute power.

Is that based on current cores or the unreleased and untested ones in Polaris?

 

[swilli89]

I agree that mohit is probably very close to the real number (at least as far as performance and price goes), but I think a third top GPU from AMD would probably be $350 and not $300, as the later is to close to the $250 SKU.

Also people are currently speculation that a GTX 1080 will be somewhere around a stock 980 Ti up to a factory overclocked 980 Ti. If that is the case then it would be 35-60% faster than the R9 480 mohit listed, and as such if it only had a 10% advantage over a theoretical 480X then said 480X would be 20-45% faster than the 480, and then it definitely wouldn't sell for $300, with the 480 at $250. $350 is probably realistic at the low end (20% higher performance),but all the way up to $450 isn't unrealistic if the performance is at the higher end (45% faster). That might still be cheaper than Nvidia, but not half the price.

AMD has already hinted at downwards pricing pressure for the Polaris release this summer "VR capable performance at new price points". R9 390 is already $350 and is the recommended minimum for VR according to VR companies. A new price point would not be $350 as it is currently, but beating the 390 by 30% and pricing it at $299 would be a big time improvement in perf/$.

We have no idea how:

• These new 4.0 GCN cores perform over 1.2 GCN cores
• The clockspeed that final 14nm silicon will be running at
• Memory performance changes
• AMD's late but welcome per core boosting (finally!)
• and many other things

I think we are in for a nice surprise when it comes to the performance of AMD's little die that could, Polaris 10. Nvidia did a great job of optimizing for 2014 games with Maxwell, now its AMD's turn to optimize for 2016 games.

 

[antihelten]

Unless R7 470 has 1536 GCN cores, it will be impossible for that chip to achieve R9 380 performance level.

1280 GCN cores clocked at 1150 MHz brings us to 2.9 TFLOPs of compute power.

R9 380 has 3.5 TFLOPs and 256 Bit memory bus.

If Polaris 11 has 1280 cores, then it would need an increase in performance per core (from either IPC or clockrate) of 40% to reach 1792 cores in performance (i.e. R9 380).

The recent PS4 neo rumors has a the new GPU with a 14% higher clock (911 MHz vs. 800 MHz), so if Polaris 11 achieves a similar increase over the 380 (which would put it at 1100 MHz), then it would only need an additional ~20% increase in IPC.

Not entirely impossible imho.

A new price point would not be $350 as it is currently, but beating the 390 by 30% and pricing it at $299 would be a big time improvement in perf/$.

Sure that would be a big improvement, but if AMD sells a GPU with 130% of the performance of the 390 for $300, then if would be close to impossible to also sell a GPU with 100% the performance of the 390 for $250. Those two price points are simply too close for such a performance gap.

 

[beginner99]

R7 470
Capacity - 4GB
Memory type - GDDR5
Performance level - R9 380
TDP - 60W no pcie connector required
Price - $149.

R9 480
Capacity - 8GB
Type - GDDR5
Performance level - R9 390
TDP - under 100W single 6 pin
Price - $249.

Make it happen AMD. Balls in your court now.

Never. Either performance is worse or price is higher. And rightly form AMDs view. That what you call a 480 will be $300 at least.

 

[swilli89]

If Polaris 11 has 1280 cores, then it would need an increase in performance per core (from either IPC or clockrate) of 40% to reach 1792 cores in performance (i.e. R9 380).

The recent PS4 neo rumors has a the new GPU with a 14% higher clock (911 MHz vs. 800 MHz), so if Polaris 11 achieves a similar increase over the 380 (which would put it at 1100 MHz), then it would only need an additional ~20% increase in IPC.

Not entirely impossible imho.



Sure that would be a big improvement, but if AMD sells a GPU with 130% of the performance of the 390 for $300, then if would be close to impossible to also sell a GPU with 100% the performance of the 390 for $250. Those two price points are simply too close for such a performance gap.

Yeah I don't get how people are stating so matter of fact that if it doesnt have x amount of SPs then it won't have y performance. We have NO IDEA how this will all shake out at this point. All we know is we saw a demo of a ~2300 core Polaris 10 part possibly standing up to a 4096 core Fury chip, all with half the technical bandwidth. If anything, I'm expecting a large per CU increase in real performance.

 

[Glo.]

Is that based on current cores or the unreleased and untested ones in Polaris?

Doesn't matter. Method is the same.

Overall if we are having fun and playing with bets:

Polaris 11: 1536 GCN cores, 128 Bit, GDDR5X memory, 1150 MHz core clock, under 60W of TDP.
Polaris 10: 3072 GCN cores, 256 Bit memory bus, GDDR5X, 1200 MHz core clock, 125W of TDP.

 

[swilli89]

Doesn't matter. Method is the same.

Overall if we are having fun and playing with bets:

Polaris 11: 1536 GCN cores, 128 Bit, GDDR5X memory, 1150 MHz core clock, under 60W of TDP.
Polaris 10: 3072 GCN cores, 256 Bit memory bus, GDDR5X, 1200 MHz core clock, 125W of TDP.

Despite the fact that the terraflops number is useless by itself anyways. Fury X has like 20% higher number than 980 Ti.

Ok if we are doing projections:

Polaris 11: 1536 GCN cores, 128 Bit, 4-8GB GDDR5, 1200 MHz core clock w/ additional boost, under 60W of TDP, $199 ----- R9 390 Performance
Polaris 10: 3072 GCN cores, 256 Bit, 8GB GDDR5, 1200 MHz core clock w/additional boost, 125W of TDP, $299 -------------- Fury X Performance

Yes bold predictions given current landscape but I believe AMD is hungry, they're confident and so they will come out aggressive. I think most here have forgotten what a true leap in nodes can do for price/performance.

 

[Glo.]

For gaming in DX11, maybe. For DX12 gaming - not at all. Whats more, Vulkan will use OpenCL in future, because it is one of parts of HSA 2.0 nature of applications. If you want to see gaming performance of the GPUs you will have to look at compute power.

For you it may be useless. But not for me . I work with compute performance, and game on the computer . That is why I am interested in this .

 

[swilli89]

For gaming in DX11, maybe. For DX12 gaming - not at all. Whats more, Vulkan will use OpenCL in future, because it is one of parts of HSA 2.0 nature of applications. If you want to see gaming performance of the GPUs you will have to look at compute power.

For you it may be useless. But not for me . I work with compute performance, and game on the computer . That is why I am interested in this .

Yes then I can understand the importance of that number, and it does make sense that it would matter in DX 12. I can't recall a time I've been this excited for a new release in recent memory. So many pieces in play, new way of coding and optimizing games, new node process (double jump), new consoles, new generation of GPU architecture!

 

[RussianSensation]

R7 470
Capacity - 4GB
Memory type - GDDR5
Performance level - R9 380
TDP - 60W no pcie connector required
Price - $149.

R9 480
Capacity - 8GB
Type - GDDR5
Performance level - R9 390
TDP - under 100W single 6 pin
Price - $249.

Make it happen AMD. Balls in your court now.

Something about AMD's next gen cards is not adding up. I'll explain.

If we go with the idea that Polaris 10 is a 2304-2560 shader card and it improves GCN 1.0/1.1 perf/watt 2.5X, and that it ONLY matches a 390, then things start to not make any sense for Vega.

We have a rumor that Vega is a 4096 shader part if Polaris 10 is a 2304 one. Recall that Pitcairn and Tahiti had similar GPU clocks for the most part.

HD7870 = 1Ghz
HD7970 = 0.925Ghz
HD7970 = 1.05Ghz

That means if Polaris 10 ~ 390 = 62%

Now assuming perfect scaling for Vega and similar clocks to Polaris 10, 4096 / 2304 = 77.8% increase in performance. Following along, that means Vega = 390/Polaris 10 = 62% x 1.778 = 110%***

That means Vega would only be about 78-80% faster than an R9 290X/390.

Ok still following me? Didn't we earlier state that Polaris 10 has 2.5X perf/watt over R9 290X/390 but now with a 4096 Vega we only ended up with a card that's 78-80% faster than a 290X.

However, AMD has Vega with even higher perf/watt over GCN 1.0/1.1 than Polaris 10 has........we have a logical inconsistency.

It also means that Vega 10 would only be 110 / 80% = 38% faster than Fury X. That's not fast enough.

Working backwards instead:

290X = 61% x At least 2.5X increase in performance/watt for Vega 10 = 152%

152% / 1.778 = Polaris 10 = 85%

That's why it's not adding up. If OTOH, we start making extra assumptions for Vega such as having 128 ROPs or much higher clock speeds than Polaris 10, the performance delta between a 2304 Polaris 10 and 4096 Vega would grow even more. This would become too unrealistic if Polaris 10 is a mid-range card. It would then straight up make Polaris 10 a low-end card for the next gen AMD stack.

The only way any of this makes sense is if Polaris 10 is much faster than a 390, Vega is clocked way way higher than Polaris 10, Polaris 10 isn't mid-range but more low end priced even lower than $249, AMD lied about 2.5X perf/watt estimates, AMD is moving away from 250-275W flagships, or Vega is a much larger chip than 4096 shaders / 96 ROPs.

***Something else. If NV improves perf/watt just 1.6X over 980, 980 = 67% x 1.6 = 107% (29% faster than the 980Ti), then our hypothetical Vega 10 is only as fast as a flagship GP104 = almost an inconceivable scenario for a 1TB HBM2 flagship Vega. This is another reason, the 2304 Polaris 10 = 390 and 4096 Vega 10 estimate make no sense.

 

[swilli89]

Something about AMD's next gen cards is not adding up. I'll explain.

If we go with the idea that Polaris 10 is a 2304-2560 shader card and it improves GCN 1.0/1.1 perf/watt 2.5X, and that it ONLY matches a 390, then things start to not make any sense for Vega.

We have a rumor that Vega is a 4096 shader part if Polaris 10 is a 2304 one. Recall that Pitcairn and Tahiti had similar GPU clocks for the most part.

HD7870 = 1Ghz
HD7970 = 0.925Ghz
HD7970 = 1.05Ghz

That means if Polaris 10 ~ 390 = 62%

Now assuming perfect scaling for Vega and similar clocks to Polaris 10, 4096 / 2304 = 77.8% increase in performance. Following along, that means Vega = 390/Polaris 10 62% x 1.778 = 110%***

That means Vega would only be about 78-80% faster than an R9 290X/390.

Ok still following me? Didn't we earlier state that Polaris 10 has 2.5X perf/watt over R9 290X/390 but now with a 4096 Vega we only ended up with a card that's 78-80% faster than a 290X.

However, AMD has Vega with even higher perf/watt over GCN 1.0/1.1 than Polaris 10 has........we have a logical inconsistency.

It also means that Vega 10 would only be 110 / 80% = 38% faster than Fury X. That's not fast enough.

Working backwards instead:

290X = 61% x At least 2.5X increase in performance for Vega 10 = 152%

152% / 1.778 = Polaris 10 = 85%

That's why it's not adding up. If OTOH, we start making extra assumptions for Vega like it having 128 ROPs or much higher clock speeds than Polaris 10, the performance delta between a 2304 Polaris 10 and 4096 Vega would grow even more. This would become too unrealistic if Polaris 10 is a mid-range card. It would then straight up make Polaris 10 a low-end card for the next gen AMD stack.

The only way any of this makes sense if is Polaris 10 is much faster than a 390, Vega is clocked way way higher than Polaris 10, Polaris 10 isn't mid-range but more low end priced even lower than $249, AMD lied about 2.5X perf/watt estimates, or Vega is a much larger chip than 4096 shaders / 96 ROPs.

***Something else. If NV improves perf/watt just 1.6X over 980, 980 = 67% x 1.6 = 107% (29% faster than the 980Ti), then our hypothetical Vega 10 is only as fast as a flagship GP104 = almost an inconceivable scenario for a 1TB HBM2 flagship Vega. This is another reason, the 2304 Polaris 10 = 390 and 4096 Vega 10 estimate make no sense.

There is absolutely no way the Polaris 10 only matches the 390. Its going to be near Fury X/980 Ti IMO.

 

[Glo.]

Performance per watt does not always mean raw performance, RS.

Lower power consumption, at the same performance level and you end up with higher PPW.

 

[Glo.]

There is absolutely no way the Polaris 10 only matches the 390. Its going to be near Fury X/980 Ti IMO.

How it will achieve this, can you tell me?

 

[R0H1T]

There is absolutely no way the Polaris 10 only matches the 390. Its going to be near Fury X/980 Ti IMO.

It really depends on the clocks & GCN4, forgot about bandwidth so GDDR5x as well, otherwise you're looking at a 1080p monster that turns into a meek cat at 4k or even 1440p.

 

[Kenmitch]

The only way any of this makes sense if is Polaris 10 is much faster than a 390, Vega is clocked way way higher than Polaris 10, Polaris 10 isn't mid-range but more low end priced even lower than $249, AMD lied about 2.5X perf/watt estimates, or Vega is a much larger chip than 4096 shaders / 96 ROPs.

You'll need to find all the speculative posts/rumors then divide that by the hype factor for your basis of your own speculated performance. Once you have that figure you'll need to way it against the fud and divide it by the concrete information we have so far....Once you have that figure you should be able to figure it out easily.

Notice I said you....I'm just going to wait it out for official launch and benches. Saving my brain power currently for work and gaming with some posting here on the side.

 

[USER8000]

The HD7970 at launch was around 40% faster than an HD6970 at 2560X1440 according to Techpowerup.

 

[Headfoot]

How it will achieve this, can you tell me?

If it ends up that fast, it would be because it will be faster on a per-shader and clock equalized basis, despite people acting like it wont be. For the record, Im not endorsing any particular end performance figure with this post. All I'm saying is you can't compare a GCN3 shader to a GCN4 shader and act like they will be the same speed.

 

[RussianSensation]

Performance per watt does not always mean raw performance, RS.

Lower power consumption, at the same performance level and you end up with higher PPW.

I addressed this in my post. If we assume Polaris 10 has 2.5X perf/watt, this is equivalent to stating that you get 275W of 390 performance in a 110W card.

Vega is shown as having more than 2.5X perf/watt over 390 gen cards because Vega itself has higher perf/watt than Polaris 10. That means every 1 Watt over Polaris has to retain the same 2.5X perf/watt to make logical sense because you just assumed that 1W Polaris card is equivalent to 2.5W 390 Hawaii gen when you did the exact same calculation in deriving Polaris 10's perf/watt.

That means Vega 10 would end up 2.27X faster than 110W Polaris 10 (250W/ 110W). This cannot be true if Polaris is also a mid-range card. That means a mid-range Polaris 10 has to be faster than 390 OR AMD's slides are made-up OR Polaris 10 is NOT a mid-range card of next generation, but just a low end sub-$249 segment. Then Vega 11 is the mid-range and Vega 10 is the flagship.

Notice I said you....I'm just going to wait it out for official launch and benches. Saving my brain power currently for work and gaming with some posting here on the side.

:thumbsup::thumbsup: Awesome post. Ya, practically, you are 100% correct.

I just cannot reconcile how you can have a 110W Polaris 10 card ~ 390 and then end up with a 250W Vega 10 card. It leaves a gigantic gap in performance between the market segments. Either that or Polaris 10 is going to be priced much more competitively, resulting in AMD cutting prices on Fury/Fury X to go against GP104 while AMD waits for Vega 11 to be true replacements for 390/390X and Vega 10 to replace Fiji.

 

[tential]

R7 470
Capacity - 4GB
Memory type - GDDR5
Performance level - R9 380
TDP - 60W no pcie connector required
Price - $149.

R9 480
Capacity - 8GB
Type - GDDR5
Performance level - R9 390
TDP - under 100W single 6 pin
Price - $249.

Make it happen AMD. Balls in your court now.

Profits, nonexistent.....

 

[Glo.]

If it ends up that fast, it would be because it will be faster on a per-shader and clock equalized basis, despite people acting like it wont be. For the record, Im not endorsing any particular end performance figure with this post. All I'm saying is you can't compare a GCN3 shader to a GCN4 shader and act like they will be the same speed.

I am not, but you forget, that because it will be 256 Bit GPU(most likely) it will have... 32 ROPs. Unless AMD will increase the ratio of ROP to memory controller.

Something just does not work for me here.

There is also a possibility that that 2304 GCN core GPU that we saw was GPU from... PS4K APU. Then it would have sense.