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

[Head1985]

are these specs real? 1536 SPs, 48ROPs and so on!?
https://www.techpowerup.com/gpudb/2861/radeon-rx-470

or did he get the specs from a 470M or something?

1536SP is very low.rx 480 will be like 40-50% faster than rx 470

 

[maddie]

are these specs real? 1536 SPs, 48ROPs and so on!?
https://www.techpowerup.com/gpudb/2861/radeon-rx-470

or did he get the specs from a 470M or something?

This was the amazing sentence.

"We recommend the AMD Radeon RX 470 for gaming with highest details at resolutions up to, and including, 5760x1080."

 

[MrTeal]

are these specs real? 1536 SPs, 48ROPs and so on!?
https://www.techpowerup.com/gpudb/2861/radeon-rx-470

or did he get the specs from a 470M or something?

I'd say there's almost no way that's correct. It would make the 470 only 67% of a full die. Hawaii Pro was 91% of a full die, Tonga and Tahiti Pro were 87.5%, and even Pitcairn Pro was 80% of a full die. That'd be one of the lowest cut dies ever made.

 

[4K_shmoorK]

Edit

 

[crisium]

are these specs real? 1536 SPs, 48ROPs and so on!?
https://www.techpowerup.com/gpudb/2861/radeon-rx-470

or did he get the specs from a 470M or something?

Whoa. That changes everything if that is accurate. If a 1536SP 470 can very nearly be 290 performance (AMD slides), then there is a lot more hope for 2304SP.

I think we would have heard in a news story of this was more than guesswork though.

 

[Head1985]

I'd say there's almost no way that's correct. It would make the 470 only 67% of a full die. Hawaii Pro was 91% of a full die, Tonga and Tahiti Pro were 87.5%, and even Pitcairn Pro was 80% of a full die. That'd be one of the lowest cut dies ever made.

Its not like 7950 vs 7970 and 290 vs 290x

RX470 vs rx480 are differend tier.

 

[Sweepr]

I'd say there's almost no way that's correct. It would make the 470 only 67% of a full die. Hawaii Pro was 91% of a full die, Tonga and Tahiti Pro were 87.5%, and even Pitcairn Pro was 80% of a full die. That'd be one of the lowest cut dies ever made.

Looks like pure guesswork, just like their Radeon RX 460 specs. Unless yields are very subpar, but sounds unlikely.

 

[MrTeal]

Its not like 7950 vs 7970 and 290 vs 290x

RX470 vs rx480 its differend tier.

P10 is a similar size and performance class to Pitcairn. Cape Verde Pro was also 80% cut, Barts (68x0) Pro was 87% of a full die, and Juniper (57x0) Pro was 90% of a full die.

A die being cut down 2/3rd is almost unheard of. Even GF100 which was a disaster yield wise from the get go only had the super cut down GTX 465 being 69% of a full die, and that was on a chip where they couldn't even produce a fully enabled part.

 

[Head1985]

P10 is a similar size and performance class to Pitcairn. Cape Verde Pro was also 80% cut, Barts (68x0) Pro was 87% of a full die, and Juniper (57x0) Pro was 90% of a full die.

A die being cut down 2/3rd is almost unheard of. Even GF100 which was a disaster yield wise from the get go only had the super cut down GTX 465 being 69% of a full die, and that was on a chip where they couldn't even produce a fully enabled part.

So you think rx 470 will have same performance as rx 480 or what?If its cut by 10% it will have same performance.
It has rx 470 name for reason.

They can ADD 2048SP version later as R 480

 

[MrTeal]

So you think rx 470 will have same performance as rx 480 or what?If its cut by 10% it will have same performance.
It has rx 470 name for reason.

They can ADD 2048SP version later as R 480

Of course not, but I also don't think they'll cut that much out. It would be way out of line with historical trends, and if the 470 is actually $149 as reported it would mean that you're getting 67% of the performance of the 480 for 75% of the price. I'm not aware of any recent cut down part that offered worse perf/$ than the full part.

 

[Bacon1]

270x ~2,7TFlops
470 -> ~5TFlops? (2048*1200)

Meh Seems GNC4 generation is not really improving efficiency at all.

You are using the 480 stats of >5TFlops not 470.

 

[Stuka87]

The GP104 competitor is a Vega card that is probably late from the decision to use HBM2. You likely have to wait because of this choice that Nvidia did not make. If it's GDDR5/X then who knows why it is delayed.

Vega isn't delayed. It was always intended to release in early 2017. However rumors have stated that it was being pushed up to October. Only time will tell if those were true.

 

[Coalscraper]

Unless i'm mistaken, i think that '+' means plus from the baseline of the GT730 price:

http://i.imgur.com/RamNi51.jpg

Which is around ~70 euro on amazon.de. That would make:

460 = 70 + 79 = 149€ (167$)
470 = 70 + 149 = 219€ (245)
480 = 70 + 209 = 279€ (313$)
480(8gb) = 70 + 249 = 319€(358$)

So if this is true, on the same amazon.de, for that price you can buy:

460 ~= GTX950
470 ~= R9 380X
480 ~= GTX970
480(8gb) ~= R9 390

Would be pretty sad if these prices are close to what we will get in Euorope.

Nope. If you look on the upgrade price for other Cards you will see that prices are quite high there:
MSI 1080 + 879€. You can get that card for 719$ here and for and for 787€ here. In this list upgrading to this card is for 879€.
This might be due to the unavailability of that card but let's look on the r9 fury.
Upgrade is 449€, best price in Germany is 398,95 here.

I don't want to fire the hype train even more and tell that the cards are even cheaper than 200€ but I don't think we'll end up with these prices. 260€ for a AIB 8GB 480 card should be in range...

 

[zinfamous]

This was the amazing sentence.

"We recommend the AMD Radeon RX 470 for gaming with highest details at resolutions up to, and including, 5760x1080."

Man, I want a display like that.

 

[topmounter]

Man, I want a display like that.

With a display like that, who needs VR goggles

 

[Head1985]

Of course not, but I also don't think they'll cut that much out. It would be way out of line with historical trends, and if the 470 is actually $149 as reported it would mean that you're getting 67% of the performance of the 480 for 75% of the price. I'm not aware of any recent cut down part that offered worse perf/$ than the full part.

rx 480 will not be 50% faster.Maybe in some games, but average will be 40%.
Btw 380x have better/Same perf/price like 380 and its full SKU vs cutdown.

 

[VulgarDisplay]

4 months? Is this true?

One set of benchmarks is using 16.1 and the other is using 16.2. amd driver numbers are always year.month.revision/hot fix.

If the cards are performing like this on unsupported drivers I think it's a good sign. I would expect some increases from drivers alone based on what we've seen.

Also. There is a very interesting thread on Reddit that shows that the steam vr scores are basically pointless to determine performance. People are running the test and getting a score and running it again later at the same settings and getting a completely different score.

For the life of me I cant get the link to work on my phone but it's in the amd subreddit.

 

[MrTeal]

rx 480 will not be 50% faster.Maybe in some games, but average will be 40%.
Btw 380x have better perf/price than 380 and its full SKU vs cutdown.

True, the performance delta will likely be less than 50%, though that could also depend on if there's a substantial clock speed difference in addition to having a full third of the SP's fused off.

Your chart seems to show the 380 having 16% higher perf/$ than the 380x. It also shows the 290 at 16.5% higher than the 290X, the 970 as 41% higher than the 980, and the Fury as 13% higher than the Fury X. Are you sure you're not linking the wrong one?

 

[3DVagabond]

For the price & (OpenCL) performance not even close, I guess you were hinting at something from Nvidia but till the time they continue to gimp compute on consumer cards, I doubt Apple will even consider them from hereon.

I'd say $700 mid range cards likely prices them out of a lot of OEM. Outside of gaming rigs, anyway.

 

[xthetenth]

One set of benchmarks is using 16.1 and the other is using 16.2. amd driver numbers are always year.month.revision/hot fix.

If the cards are performing like this on unsupported drivers I think it's a good sign. I would expect some increases from drivers alone based on what we've seen.

Also. There is a very interesting thread on Reddit that shows that the steam vr scores are basically pointless to determine performance. People are running the test and getting a score and running it again later at the same settings and getting a completely different score.

For the life of me I cant get the link to work on my phone but it's in the amd subreddit.

My 290 suddenly becomes able to do VR when I disable most of my monitor space. It's a pretty bad benchmark for proper comparisons.

 

[Headfoot]

Man, I want a display like that.

It's just 3x1080p. I run it

 

[Armsdealer]

I'd say there's almost no way that's correct. It would make the 470 only 67% of a full die. Hawaii Pro was 91% of a full die, Tonga and Tahiti Pro were 87.5%, and even Pitcairn Pro was 80% of a full die. That'd be one of the lowest cut dies ever made.

and why are you assuming 28nm defects are having the same impact at 14nm ff?

It could very well be a manifestation of the new process and architecture. We really don't know. Although it would certainly explain the reduction of segmentation that appears to be occurring with the 'x' gpus having disappeared.

 

[Headfoot]

I gotta agree with MrTeal. 67% cut die does not fit with any historical norm. Even more so there should be less need to cut on a small die. Not sold on 67% size at all. I bet it's less cut, closer to 80+%, but lower clockspeed.

 

[MrTeal]

and why are you assuming 28nm defects are having the same impact at 14nm ff?

It could very well be a manifestation of the new process and architecture. We really don't know. Although it would certainly explain the reduction of segmentation that appears to be occurring with the 'x' gpus having disappeared.

I'm not. Tahiti Pro however was 87.5% on a brand new node. 5850 was 90% of the 5870 on the brand new 40nm node. 4850 wasn't cut down in shaders at all vs the 4870 when 55nm was new, and both the 3870/3850 pair and 2900 XT and 2900 Pro had the same core config.

It's not impossible that AMD's releasing such a gimped die, it's just unlikely given the historical precedent especially since it would mean the cut part would have worse perf/$ which really never happens.

 

[crisium]

I'm not sold on the idea either, because it's likely guesswork, but if P10 is 2304 P11 is 1024 that does leave a lot of room for a very cut card. AMD have never had that large of a gap (2048 vs 1280, 2816 vs 2048, for example). Especially if P11 only has 32 ROPs and P10 is 64, that's a titanic gap.