AMD Polaris Thread: Radeon RX 480, RX 470 & RX 460 launching June 29th

[Glo.]

https://www.reddit.com/r/Amd/commen...une_29th_nda_is_for_the_480x490_and/?sort=old

480X doesn't exist. There's 3 cards in the Polaris lineup, 470 480 and 490, NDA for 480 expires on 29th. No spy plots, no conspiracy shit, nothing of nothing. Stop creating stories, please.

He says that those 3 SKUs have 24 CU's, 36, and... 48 CU's. Which would have sense. 67DF:C7 is 6x6 Unit setup in the die. So other are very logical: 6x4 and 6x8.

3072 GCN cores. VERY interesting if true. 299$?

 

[Sweepr]

for the last time, show me within a review same delta between performance and tflops of 980/1070. if you can't than it's proven there's a ipc regression.

As you wish:

Geforce GTX 980 = ~1225 MHz (between 1200 and 1250 MHz) x 2 x 2048 = 5.02 TFLOPS

Geforce GTX 1070 = Starts slightly above 1800 MHz, averaging at ~1775 MHz x 2 x 1920 = 6.82 TFLOPS

6.82/5.02 = 35.8 % more TFLOPS for 39% / 57% more performance (DX11 / DX12 @ 1440P).

 

[Glo.]

Sweeper, if you ask about Teraflops compute power, answer yourself why is GTX 1080 scoring lower on Luxmark 3.1 OpenCL renderer that GTX 980 TI, despite having higher compute power?

Exactly the same thing we see in Computerbase.de review of GTX 1080. In Luxmark GTX 1080 is slower than previous generation of GPU.

P.S. Stop this nonsense. This is RX 480 thread.

 

[.vodka]

https://www.reddit.com/r/Amd/commen...une_29th_nda_is_for_the_480x490_and/?sort=old


He says that those 3 SKUs have 24 CU's, 36, and... 48 CU's. Which would have sense. 67DF:C7 is 6x6 Unit setup in the die. So other are very logical: 6x4 and 6x8.

3072 GCN cores. VERY interesting if true. 299$?

So Polaris 10 actually has 3072SPs if we go by this? AMD talked about the $100 to $300 range, if the $300 card is actually a 3072SP part.. It makes sense to disable that much of the chip to build up stock for the mainstream part for a hard launch and keep the full dies for reveal later... considering GF's reputation for screwing up, this makes sense.


If the 232mm^2 chip size ends up being correct... 1070, 315mm^2, you're in trouble with that alone, stock vs stock. If Polaris also overclocks better than 28nm GCN, well.. for $300 such a card is a no brainer vs scam edition $430-450 1070. $380 is another thing and is a much closer contest. The 1080 could be in trouble in that case too (not the >2GHz ones).


Hell, 2560sp as the full chip is what one would expect, but 3072sp? If this ends up being true, this is another repeat of the 4850 and 4870 and nV will have to slash prices almost in half like they did with 260 and 280. Nice!

 

[kraatus77]

As you wish:

Geforce GTX 980 = ~1225 MHz (between 1200 and 1250 MHz) x 2 x 2048 = 5.02 TFLOPS

Geforce GTX 1070 = Starts slightly above 1800 MHz, averaging at ~1775 MHz x 2 x 1920 = 6.82 TFLOPS

6.82/5.02 = 35.8 % more TFLOPS for 39% / 57% more performance (DX11 / DX12 @ 1440P).

For the first time, good job :thumbsup: have a 1 internet gold coin.

 

[Glo.]

So Polaris 10 actually has 3072SPs if we go by this? AMD talked about the $100 to $300 range, if the $300 card is actually a 3072SP part.. It makes sense to disable that much of the chip to build up stock for the mainstream part for a hard launch and keep the full dies for reveal later... considering GF's reputation for screwing up, this makes sense.


If the 232mm^2 chip size ends up being correct... 1070, 315mm^2, you're in trouble with that alone, stock vs stock. If Polaris also overclocks better than 28nm GCN, well.. for $300 such a card is a no brainer vs scam edition $430-450 1070. $380 is another thing and is a much closer contest. The 1080 could be in trouble in that case too (not the >2GHz ones).


Hell, 2560sp as the full chip is what one would expect, but 3072sp? If this ends up being true, this is another repeat of the 4850 and 4870 and nV will have to slash prices almost in half like they did with 260 and 280. Nice!

If I count this correctly, 48 CU P10 with 1.35 GHz, and GDDR5X memory would be within 10% of GTX 1080. For half the price, and with much lower power consumption...

Amazing possibility, have to say. That would tell why Nvidia rushed so hard with GTX 1080 and GTX 1070.

 

[happy medium]

Amazing possibility, have to say

If this is true, for 300$ , even with 44cu's, I'm in, if its as fast as a gtx1070. Unless the 1070 has a price cut.

 

[littleg]

So Polaris 10 actually has 3072SPs if we go by this? AMD talked about the $100 to $300 range, if the $300 card is actually a 3072SP part.. It makes sense to disable that much of the chip to build up stock for the mainstream part for a hard launch and keep the full dies for reveal later... considering GF's reputation for screwing up, this makes sense.


If the 232mm^2 chip size ends up being correct... 1070, 315mm^2, you're in trouble with that alone, stock vs stock. If Polaris also overclocks better than 28nm GCN, well.. for $300 such a card is a no brainer vs scam edition $430-450 1070. $380 is another thing and is a much closer contest. The 1080 could be in trouble in that case too (not the >2GHz ones).


Hell, 2560sp as the full chip is what one would expect, but 3072sp? If this ends up being true, this is another repeat of the 4850 and 4870 and nV will have to slash prices almost in half like they did with 260 and 280. Nice!

We know there must be something above the 480 as Lisa Su specifically said Polaris up to $300. Kind of makes me wonder if NV already know of it which is why they rushed the launch of the 1080/70 up.

 

[garagisti]

980= 1200mhz x 2 x 2048 = 4.9 tflops ( actual clock)

1070= 1800mhz x 2 x 1920 = 6.9 tflops (actual clock)

6.9/4.9= 40.8 % more tflops for 40.8% more performance, where are you getting that 25.3% difference lol.

and tpu's 1070was running at 1873mhz, so it's actually lower in ipc.

there's no ipc increase for pascal. period.

To measure increase over 980 clocks, 980 speeds must be used as base. So if you use 1200 as base, the 600 mhz increase is 50% increase in clocks.

 

[.vodka]

If the 3dmark leaks hold true and the 36CU (2304sp) RX480 is at 980/Fury level, 44CU (2816sp) is 22% extra hardware, 48CU (3072sp) is 33% extra. Nevermind all the extra hardware that's also enabled, AMD also disables TMUs and other units when cutting CUs.

At the same 1266MHz clocks in the leak, and assuming somewhat linear scaling, 44CU/+22% puts P10 at 1070/TX tier, 48CU/+33% with a 100MHz bump is probably at 1080 territory.

If so, this also explains why Fiji was so lackluster and an afterthought of a GPU just to test HBM and have something to match stock 980Ti, all the crucial R&D went to Polaris (and Vega) and it could very well pay off.

We know there must be something above the 480 as Lisa Su specifically said Polaris up to $300. Kind of makes me wonder if NV already know of it which is why they rushed the launch of the 1080/70 up.

Should P10 overclock well and the full variants speculated here fill the $300 bracket, we're in for a treat and a reset of the GPU market pricing that's long overdue. Hell, the $200-230 480 already does that in some ways, although not at the full effect this 3072sp part at the stated max $300 price tag could.

As you say, this also explains the 1070 and 1080 coming so early out of the gate, with a $100 early adopter tax, and with so little supply that you see it out of stock for weeks and it sells out within minutes.

Maybe Raja saying AMD was ahead of the transition to 14nm in the mainstream segment back then in January wasn't smoke and mirrors after all.


RV770 did have 800sp in it when the reviews came out. Who knows, maybe this ends up being true.

 

[Mopetar]

https://www.reddit.com/r/Amd/commen...une_29th_nda_is_for_the_480x490_and/?sort=old



He says that those 3 SKUs have 24 CU's, 36, and... 48 CU's. Which would have sense. 67DF:C7 is 6x6 Unit setup in the die. So other are very logical: 6x4 and 6x8.

3072 GCN cores. VERY interesting if true. 299$?

That seems very unlikely, but if it were true it would be deserving of the 490 moniker. At the same time, that makes the 480 a massively cut die, the likes of which we haven't seen since . . . well the 1070 actually, but that's pretty unexpected for a ~230 mm^2 chip to yield that poorly, even if I believe that Global Foundries has had issues with their new node.

Also, it doesn't make sense for other reasons. All of the rumors have pointed towards Polaris 10 being around ~230 mm^2. If they managed to fit 3072 SPs in that area, it would be around 2.3 times as dense (I'm basing density off of SP/mm^2 as we don't have a transistor count for Polaris 10 yet) as their previous generation chips, when Nvidia has well under that with the 1070/1080.

As exciting as it would be to see a 3072 core GCN part right now, I don't think we're going to get one. I'd be happy for a 2560 SP chip that can hit 1.4 GHz clocks, which should be close to a 25% performance gain over the 480 and worth an extra $100 (or $70 assuming the $230 8 GB 480 is true) over the cost of a 480.

 

[BlitzWulf]

We know there must be something above the 480 as Lisa Su specifically said Polaris up to $300. Kind of makes me wonder if NV already know of it which is why they rushed the launch of the 1080/70 up.

I also wonder if Nvidia had some idea of Fat P10's performance and GP104 MSRP's werent some sort of pre-emptive price cut to save face and avoid another 280 situation.

I honestly thought as much when I first saw the relation of MSRP to reference/FE price.

 

[garagisti]

https://www.reddit.com/r/Amd/commen...une_29th_nda_is_for_the_480x490_and/?sort=old



He says that those 3 SKUs have 24 CU's, 36, and... 48 CU's. Which would have sense. 67DF:C7 is 6x6 Unit setup in the die. So other are very logical: 6x4 and 6x8.

3072 GCN cores. VERY interesting if true. 299$?

I may just cream my pants with excitement, if that price of $299 comes true for 3072 GCN cores for a 490 (about 33% more cores over 480).

 

[Mopetar]

The GP104 pricing has nothing to do with P10 performance. It's there because Nvidia knows its effectively a paper launch with very limited supply. This is just Nvidia deciding that they'll take the extra money off the top instead of people who are buying and reselling cards.

 

[garagisti]

If the 3dmark leaks hold true and the 36CU (2304sp) RX480 is at 980/Fury level, 44CU (2816sp) is 22% extra hardware, 48CU (3072sp) is 33% extra. Nevermind all the extra hardware that's also enabled, AMD also disables TMUs and other units when cutting CUs.

At the same 1266MHz clocks in the leak, and assuming somewhat linear scaling, 44CU/+22% puts P10 at 1070/TX tier, 48CU/+33% with a 100MHz bump is probably at 1080 territory.

If so, this also explains why Fiji was so lackluster and an afterthought of a GPU just to test HBM and have something to match stock 980Ti, all the crucial R&D went to Polaris (and Vega) and it could very well pay off.



Should P10 overclock well and the full variants speculated here fill the $300 bracket, we're in for a treat and a reset of the GPU market pricing that's long overdue. Hell, the $200-230 480 already does that in some ways, although not at the full effect this 3072sp part at the stated max $300 price tag could.

As you say, this also explains the 1070 and 1080 coming so early out of the gate, with a $100 early adopter tax, and with so little supply that you see it out of stock for weeks and it sells out within minutes.

Maybe Raja saying AMD was ahead of the transition to 14nm in the mainstream segment back then in January wasn't smoke and mirrors after all.


RV770 did have 800sp in it when the reviews came out. Who knows, maybe this ends up being true.

I'm sure some here would tell you that paying as much as $600 for what is but a midrange card of new arch is worth it It's not like it is only recently since the 680, that the midrange cards have been priced aggressively by one company. I genuinely think Nvidia will continue to offer such immense "value" to its buyers in next 2-3 years or however long it takes for purchases of such priced hardware to slow down.

This doesn't explain why Fiji wasn't all we hoped for, but heck, you live you learn. However that thing can still pull its weight in DX12. Will see how it works in upcoming DX12 titles in a year's time or so.

 

[garagisti]

That seems very unlikely, but if it were true it would be deserving of the 490 moniker. At the same time, that makes the 480 a massively cut die, the likes of which we haven't seen since . . . well the 1070 actually, but that's pretty unexpected for a ~230 mm^2 chip to yield that poorly, even if I believe that Global Foundries has had issues with their new node.

Also, it doesn't make sense for other reasons. All of the rumors have pointed towards Polaris 10 being around ~230 mm^2. If they managed to fit 3072 SPs in that area, it would be around 2.3 times as dense (I'm basing density off of SP/mm^2 as we don't have a transistor count for Polaris 10 yet) as their previous generation chips, when Nvidia has well under that with the 1070/1080.

As exciting as it would be to see a 3072 core GCN part right now, I don't think we're going to get one. I'd be happy for a 2560 SP chip that can hit 1.4 GHz clocks, which should be close to a 25% performance gain over the 480 and worth an extra $100 (or $70 assuming the $230 8 GB 480 is true) over the cost of a 480.

I doubt it has more to do with yields, than to provide a cut chip with solid performance for a more accessible price to the market.

 

[swilli89]

I may just cream my pants with excitement, if that price of $299 comes true for 3072 GCN cores for a 490 (about 33% more cores over 480).

Yep, very interesting development. I hope we could see this chip sooner rather than later.

 

[sirmo]

I doubt it has more to do with yields, than to provide a cut chip with solid performance for a more accessible price to the market.

I agree, if they are doing a massive cut, it's to ensure high yields since as it looks right now rx480 should sell in large numbers.

 

[JDG1980]

https://www.reddit.com/r/Amd/commen...une_29th_nda_is_for_the_480x490_and/?sort=old



He says that those 3 SKUs have 24 CU's, 36, and... 48 CU's. Which would have sense. 67DF:C7 is 6x6 Unit setup in the die. So other are very logical: 6x4 and 6x8.

3072 GCN cores. VERY interesting if true. 299$?

That sounds too good to be true. I'm taking it with a grain of salt for now. But if it is accurate... Nvidia could be in for some serious trouble.

AtlasRush seems to be saying that 470/480/490 would all be Polaris 10 SKUs. Further down the thread, he indicates that Polaris 11 would be the 460/450/440 (with "R" prefix instead of "RX"). If all this is true, it seems to indicate that GloFo 14LPP yields are fairly bad at this time. Not only would there be third-tier salvage parts for even the small Cape-Verde-sized Polaris 11 die, but the third-tier salvage for Polaris 10 (RX 470) would be cut down by a full 50% (1536 of 3072 shaders enabled). AtlasRush was also careful to say "NDA for 480 expires on 29th." Thus even if this is accurate, we may not see the full 490 until later. And if yields are really bad, it may take some time for availability to become widespread.

 

[boozzer]

We know there must be something above the 480 as Lisa Su specifically said Polaris up to $300. Kind of makes me wonder if NV already know of it which is why they rushed the launch of the 1080/70 up.

glo's post clearly found a source that stated 470 480 490. if there is a 300$ gpu it will be the 490.

could be 200 250 300$.

That sounds too good to be true. I'm taking it with a grain of salt for now. But if it is accurate... Nvidia could be in for some serious trouble.

AtlasRush seems to be saying that 470/480/490 would all be Polaris 10 SKUs. Further down the thread, he indicates that Polaris 11 would be the 460/450/440 (with "R" prefix instead of "RX"). If all this is true, it seems to indicate that GloFo 14LPP yields are fairly bad at this time. Not only would there be third-tier salvage parts for even the small Cape-Verde-sized Polaris 11 die, but the third-tier salvage for Polaris 10 (RX 470) would be cut down by a full 50% (1536 of 3072 shaders enabled). AtlasRush was also careful to say "NDA for 480 expires on 29th." Thus even if this is accurate, we may not see the full 490 until later. And if yields are really bad, it may take some time for availability to become widespread.

nah, it reads like AMD is trying to corner even single market segment under 300$

every single one of them.

 

[crisium]

Don't get your hopes up. 2304SP is all that's confirmed, and AMD always barely cuts their chips at all unlike Nvidia. 2304 could still be full, and if its cut then 2560 is the most likely.

 

[littleg]

That sounds too good to be true. I'm taking it with a grain of salt for now. But if it is accurate... Nvidia could be in for some serious trouble.

AtlasRush seems to be saying that 470/480/490 would all be Polaris 10 SKUs. Further down the thread, he indicates that Polaris 11 would be the 460/450/440 (with "R" prefix instead of "RX"). If all this is true, it seems to indicate that GloFo 14LPP yields are fairly bad at this time. Not only would there be third-tier salvage parts for even the small Cape-Verde-sized Polaris 11 die, but the third-tier salvage for Polaris 10 (RX 470) would be cut down by a full 50% (1536 of 3072 shaders enabled). AtlasRush was also careful to say "NDA for 480 expires on 29th." Thus even if this is accurate, we may not see the full 490 until later. And if yields are really bad, it may take some time for availability to become widespread.

Same guy also says yields are good enough.

 

[moonbogg]

Imagine if the $230 8GB version was a mean ass overclocker? Probably too good to be true, but imagine if that was the case. Holy crap that would be a mind blowing awesome value. Approaching 980ti performance for $230 with 2GB more vram. I want this to happen.

 

[Glo.]

Pricing is from what we know:
RX 480 4 GB - 199$
RX 480 8 GB - 229$
? - <300$

 

[boozzer]

Imagine if the $230 8GB version was a mean ass overclocker? Probably too good to be true, but imagine if that was the case. Holy crap that would be a mind blowing awesome value. Approaching 980ti performance for $230 with 2GB more vram. I want this to happen.

no hype bro, no hype. I am just reading this, and putting down hypes.

I want benchmarks before any super hype.