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

[sirmo]

Now all weakness covered by AMD, except perf/mm!!!

heh, perf/mm is such a trivial thing too, like it makes no difference to the consumer.. and there is no meaningful way to even compare the two, as there are things that have to be on the GPU no matter how small the GPU is so a larger GPU will always have an edge there.

here is something that actually makes difference to the consumer:

How come an AMD's rx480 $199 reference card delivers 6 VRM phases to the GPU and supposedly superior "Craftsmanship" FE edition 1070 $449, Nvidia completely skimps on VRMs and only delivers 4 VRM phases, leaving two completely unpopulated.

source: https://www.youtube.com/watch?v=X6dCz4Wt1Ek

 

[Maverick177]

People care about brand > Price . Performance > Power consumption. Die size only makes sense for people like us aka tech forums user and / or chip makers.

If an OC RX480 can reach within 5% of Fury X level, I will part ways with my Fury Nitro.

 

[Armsdealer]

People care about brand > Price . Performance > Power consumption. Die size only makes sense for people like us aka tech forums user and / or chip maker.

People don't matter as much as Dell, Apple, HP, Lenovo, Microsoft Surface , and they care a lot about perf / watt and price. the stock isn't going up because of what "people" think. It's going up because AMD's going to get a lot of share back in the places they lost it because their perf / watt is now ahead of NVidia and they're offering good prices. IF we really cared about discussing what AMD cares about as opposed to what enthusiasts care about, the most important product is 35W P11.

I think AMD has solved the perf / watt issue to a large degree, and now they can also attack their brand perception issue amongst enthusiast dGPU buyers.

 

[Ma_Deuce]

If this card can pull GTX980TI performance at 1.5Ghz then it is really new 4870.It will be within 20% of 1070 even after OC, because 1070 cant OC.

If it can OC to 1.5Ghz and actually maintain that speed gaming, color me impressed. NV seems to be having a really difficult time with that and I wonder if AMD is going to have the same shrinking pains.

I have been holding off on replacing my 780 for a while now. I think 480 @ 1.4Ghz or so would be a solid upgrade and for a great price.

 

[Silverforce11]

Some shops are listing retail spec for custom RX 480s...

This one is ridiculous if true.

1.6ghz and 10Gbps GDDR5..

I don't believe it though, since GDDR5 doesn't like running above 9Gbps.

It would only make sense if its the special sauce $299 model with GDDR5X..

 

[Maverick177]

People don't matter as much as Dell, Apple, HP, Lenovo, Microsoft Surface , and they care a lot about perf / watt and price. the stock isn't going up because of what "people" think. It's going up because AMD's going to get a lot of share back in the places they lost it because their perf / watt is now ahead of NVidia and they're offering good prices. IF we really cared about discussing what AMD cares about as opposed to what enthusiasts care about, the most important product is 35W P11.

I think AMD has solved the perf / watt issue to a large degree, and now they can also attack their brand perception issue amongst enthusiast dGPU buyers.

In case of Nvidia, Gaymerz matter alot, look at their financial statement.

If AMD wants to beat Nvidia in terms of perf/w, POLARIS HAS TO BEAT PASCAL WHEN BOTH OC IN TERMS OF PERF/W.

Polaris beating pascal perf/w at stock simply won't cut it anymore. As of now, Pascal is twice times perf/w of Fiji.

AMD claimed UP TO 2.8x perf/w from the previous generation ( not sure which ), so in order to beat Pascal when OCed in terms of Perf/w, Polaris maximum power consumption cannot go higher than 90W when OC ( 10% less perf than Fury X ). I expect Polaris will scare well with voltage both perf gains and power consumption. Probably why AIBs have 6 + 8 pins card lying around.
http://www.techpowerup.com/reviews/ASUS/GTX_1080_STRIX/22.html

 

[iiiankiii]

heh, perf/mm is such a trivial thing too, like it makes no difference to the consumer.. and there is no meaningful way to even compare the two, as there are things that have to be on the GPU no matter how small the GPU is so a larger GPU will always have an edge there.

here is something that actually makes difference to the consumer:

How come an AMD's rx480 $199 reference card delivers 6 VRM phases to the GPU and supposedly superior "Craftsmanship" FE edition 1070 $449, Nvidia completely skimps on VRMs and only delivers 4 VRM phases, leaving two completely unpopulated.

source: https://www.youtube.com/watch?v=X6dCz4Wt1Ek

perf/mm is important if you care about absolute performance. It helps gauge how big pascal and vega would do. To me, that is pretty important because I know there is limit to how big a die size can be from a technical/business standpoint. You can only jam so many cores into a die.

Yes, I will grab a RX480 to hold me over until big pascal/vega. The RX480 is an excellently positioned card. But, discussing perf/mm isn't a knock on the RX480. It's a knock on AMD and its ability to compete at the HIGHEST level. In that regards, discussing the perf/mm of these small die helps us guesstimate how well the bigger die will perform. That's all.

 

[Xpage]

Huh?

GTX1070 is only 37% faster than R9 390X using a GPU-dependent resolution.

When was the last time you used SLI/CF, if ever? There is SO much non-sense being spouted about SLI/CF online, it's nuts. Not all of us spend $60-120 on launch day AAA titles. Did it occur to you that some people buy games 12-24 months from launch as GOTY edition and/or for 1/3 to 1/4th the price? By that time SLI/CF work in 80-90% of PC games.

What's the fastest card setup in this review?
http://www.anandtech.com/show/9390/the-amd-radeon-r9-fury-x-review/13

What about this review?
http://www.guru3d.com/articles_pages/geforce_gtx_1080_2_way_sli_review,13.html

Don't make blanket statements about SLI/CF if you've never used them for years.



The cheapest GTX1070 for "sale" is $419. Let's just face the truth and admit that $379.99 GTX1070 should not be used as a valid example for GTX 1070's market price. I will admit that it's probably unfair to use $199-249 RX480 prices since we don't know their retail market pricing after June 29th but at least be genuine about GTX 1070's market pricing and lack of inventory.

http://www.nowinstock.net/computers/videocards/nvidia/gtx1070/

$419.99 / 1.37 (R9 390X's speed) = RX 480 can cost $305 and have directly proportional price/performance.

Your analysis completely ignores 5 major factors (at minimum):

1. GTX1070 is looking like a nice paper launch. Chances are AMD will deliver 10-20X the amount of RX 480 cards worldwide when it launches. If there are no GTX1070s to buy, it makes no differences how much better it is if it's not for sale.

2. You are only looking at % charts and completely ignoring that for 1080p 60hz and below gaming (>85% of PC gamers), in terms of frames per second, RX 480 is close to perfect for the mainstream/performance gamer. % charts are useless if a $200-250 card provides 40-60 fps levels of performance that gamers find satisfactory. If this weren't the case, almost no one would have purchased GTX750/750Ti/950/960. In fact, RX480 will provide GTX970/980/390X level of performance. That means using your own argument, a $550 GTX980 or $430 R9 390X was worth buying in 2014-2015 but a card with this level of performance for $200-250 is garbage now for 1080p 60Hz High settings? :sneaky:

3. You are 100% ignoring CPU bottlenecks. Nice try. The target market for RX 480 $200-300 GPU landscape tends to have i3/i5/FX or stock i7s. All of these are too slow to extract maximum performance out of the GTX1070.

http://www.purepc.pl/pamieci_ram/te...pamieci_ram_wybrac_do_intel_skylake?page=0,12

http://www.techspot.com/article/1171-ddr4-4000-mhz-performance/page3.html

Nearly every professional review online is using i7 6700K @ 4.5Ghz or 5820K-5960X OC (or similar). The vast majority of Mainstream/Performance PC gamers do not own such fast CPUs. For many of them, GTX1070 would be CPU bottlenecked straight up.

4. History has proven already that it's going to be better to purchase a $250 RX 480, then sell it for $100 in 2 years and buy another $250-300 card for the 2018-2020 period. This is better than buying a $420-450 GTX1070 and holding onto to it for 4 years.

Considering you bought a $550 GTX980 and used it for the last 2 years, calling a $200-250 card with similar level of performance and better feature set (HDR, DX12, 4K video acceleration, 8GB of VRAM, lower power usage) meh sounds like you are straight up trolling.

5. Given GTX1070's higher launch price worldwide, the taxes that most of us have to pay on top of the already inflated price and the strong USD imply that it's highly likely that the disparity between ALL-IN price (retail USD MSRP converted to local currency + import tariffs + local VAT) will be greater worldwide than it will be in the U.S. This implies the $379.99 USD vs. $199-249 RX 480 USD is actually a best case scenario for NV. Looking at prices of GTX1070 in Europe, Russia, Canada, the RX 480 will be an even better value.

You have now resorted to using the most useless engineering metric of all time to discredit P10 -- perf/mm2.

487mm2 GTX275 vs. 282mm2 HD4890 -- you bought the 275
520mm2 GTX570 vs. 389mm2 HD6950/6970 -- you bought the 570
561mm2 GTX780 vs. 438mm2 R9 290/290X -- you bought the 780

To summarize, every new generation, you pick the most favourable metric in which NV leads and hype it up to the moon but looking at the past history of your GPU ownership, you clearly couldn't care less about perf/watt or perf/mm2.

If you are going to use the perf/mm2 metric and expect to be taken seriously, then you'd trash all over the mid-range Pascal GP104 as one overpriced turd and tell everyone to not buy it. Why aren't you telling everyone how a $599-699 GTX1080 is a $249 GTX560Ti line? All of a sudden your perf/mm2 metric doesn't sound so good when 95% of PC gamers cannot afford this product. Instead you manipulate this metric to imply how much faster the full GP104 is, while 100% ignoring:

1) RX 480 will use less power in games than GTX1070/1080 will
2) GTX1080 uses much faster and more efficient GDDR5X over RX 480's regular GDDR5
3) Even if GP104 1080 was 9mm2, do you think we care? It costs almost $1,000 Canadian over here after taxes. You can keep hyping up perf/mm2 and perf/watt all you want but in the real world, price/performance smashes both of those metrics for mainstream/performance consumers when comparing a $200-250 videocard with a $400-450 one.

On top of all of this, you keep ignoring the MOST important barrier to entry for mainstream/performance segment -- PRICE. Are you seriously telling everyone here with a straight face that even if GTX1070 had superior or similar price/performance scaling to a $199-249 RX480 that someone would just go from that to a $419-449 GTX1070? :sneaky:

Your posts are seriously taking a nosedive as of late and the hypocrisy from you is out of this world since you never ever recommended R9 280X/380X/290 over the GTX950/960 but the price difference was smaller and the price/performance advantage of those AMD cards was greater than GTX1070 will have over RX 480. If perf/mm2 was a huge metric you make it out to be, then you'd have never purchased or recommended the GTX275/280/285, GTX570/580, GTX780/780Ti, etc. but you did....

*slow clap*

great post. I don't game much, sold my 6950 back in the day during the mining craze, got good $ for it. For $199 less than I paid for the 6950, I'll be getting Polaris, I'd like to wait for Vega but I doubt the price will be less than $299. Thus for occasional gamers this is a great deal vs paying 2x the price for a tiny bit of performance increase.

 

[itsmydamnation]

If AMD wants to beat Nvidia in terms of perf/w, POLARIS HAS TO BEAT PASCAL WHEN BOTH OC IN TERMS OF PERF/W.

ROLF sure mate, pro tip anyone OC'ing doesn't give a crap about power, all they care about is clock vs fan noise. This invention of what people care about is really quite sad.....

 

[SimianR]

I'm wondering if because AMD was hit so hard for being the hot, power hungry, "inefficient" brand with the R9 200/300 series that they left a lot of performance on the table and decided to go with lower base clocks this time around. Launch with lower clocks, lower power, lower temps, but still decent performance at an affordable price, and also an overclocking monster? Might be win/win.

 

[MrTeal]

Team Fortress 2.

Plus will it also work for DX10 and 11 games? Would love to see a boost in Overwatch performance.

Not sure if you're joking. Is that some kind of heavily modded TF2? I just spun it up at 4k 8xAA 16xAF on my system and it was topping out the afterburner FPS overlay at 289FPS on a bot training match. I was still getting mid 130s while still mining Ethereum while playing.

 

[Silverforce11]

perf/mm is important if you care about absolute performance. It helps gauge how big pascal and vega would do.

It doesn't because Vega is actually GCN 5, a new architecture and with HBM2.

You cannot draw comparisons from Polaris to Vega, well, it won't be accurate, it's similar to drawing comparisons from Pascal to Volta.

AMD's JPMorgan interview is quite clear, they said never in their history have them rolled out subsequent architectures as fast as they expect to following on from Polaris, with Vega architecture coming out much more quickly than usual.

 

[sirmo]

perf/mm is important if you care about absolute performance. It helps gauge how big pascal and vega would do. To me, that is pretty important because I know there is limit to how big a die size can be from a technical/business standpoint. You can only jam so many cores into a die.

Polaris and Vega are different architectures.. they could have entirely different densities depending on even the libraries they are using in Cadence.

Also unless we actually get to see the die shots from someone like Chipworks we can't even begin to compare the perf per mm^2.

GPU is not just the CUs. There are caches there, memory controllers, command processors ACEs.. decoders, all sorts of "uncore" circuitry that wouldn't necessarily scale linearly for a larger chip.

Some parts of the chip might deal with different voltages and might be scaled up to handle those voltages.

Further completely skewing any sort of perf/mm^2 comparison you might have.

It's completely idiotic to compare perf/mm^2 of two completely architectures, targeting different performance segments, with different features, built on a different process (like in case of Polaris and Pascal), and expect any meaningful performance conclusion, without even knowing what the die looks like.

We also don't know how much of the chip has redundant circuits to improve yields.

This isn't lumber, these are electronic circuits.

 

[iiiankiii]

Polaris and Vega are different architectures.. they could have entirely different densities depending on even the libraries they are using in Cadence.

Also unless we actually get to see the die shots from someone like Chipworks we can't even begin to compare the perf per mm^2.

GPU is not just the CUs. There are caches there, memory controllers, command processors ACEs.. decoders, all sorts of "uncore" circuitry that wouldn't necessarily scale linearly for a larger chip.

Some parts of the chip might deal with different voltages and might be scaled up to handle those voltages.

Further completely skewing any sort of perf/mm^2 comparison you might have.

It's completely idiotic to compare perf/mm^2 of two completely architectures, targeting different performance segments, with different features, built on a different process, and expect any meaningful performance conclusion.

What? Wait. Vega will be entirely new arch? Well, if that's the case. You're right. It doesn't make sense to compare Polaris with Vega. But, big pascal will basically be a larger die of small pascal, right? If they're the same arch, you can definitely draw some conclusion as to how well its bigger counterpart will perform. But, since Vega isn't just a scaled up Polaris,(but an entirely new and different arch?), Vega is definitely an unknown even if we know its die size.

To me, being an entirely new arch, makes Vega an much, much exciting piece of technology to look forward to than big pascal. Exciting times ahead for AMD!

 

[maddie]

It doesn't because Vega is actually GCN 5, a new architecture and with HBM2.

You cannot draw comparisons from Polaris to Vega, well, it won't be accurate, it's similar to drawing comparisons from Pascal to Volta.

AMD's JPMorgan interview is quite clear, they said never in their history have them rolled out subsequent architectures as fast as they expect to following on from Polaris, with Vega architecture coming out much more quickly than usual.

Correct, but we can estimate the rumoured GDDR5X GP102. 150% Gp104 in shaders should max out power.

 

[sirmo]

What? Wait. Vega will be entirely new arch? Well, if that's the case. You're right. It doesn't make sense to compare Polaris with Vega. But, big pascal will basically be a larger die of small pascal, right? If they're the same arch, you can definitely draw some conclusion as to how well its bigger counterpart will perform. But, since Vega isn't just a scaled up Polaris,(but an entirely new and different arch?), Vega is definitely an unknown even if we know its die size.

To me, being an entirely new arch, makes Vega an much, much exciting piece of technology to look forward to than big pascal. Exciting times ahead for AMD!

Did you read anything else I wrote? Even if it was the identical architecture you still can't use perf/mm2 as a meaningful metric.

 

[monstercameron]

http://wccftech.com/amd-rx-480-1500mhz-overclocking-tool-voltage-2/

I think its safe to board the [REDACTED] hype train!!!!!!!!!!

 

[.vodka]

Seems the new node + GCN4 has given AMD a chance to work on GCN's clock speed wall on 28nm. Also nice to see them learning from their past failures. Voltage control on day one or so is a nice surprise.


Rumored 1266MHz performance +18-26% (1266->1500-1600) is nearby FuryX/980Ti = 1070 stock performance. Should the $200-240 RX 480s overclock like this (the reference PCB in that picture looks solid)... well.. a single 8GB RX480 is potentially able to make the 1070 redundant, or force a price cut. Yes, you can also overclock the 1070 and land near the 1080, but $240-250 to >$400, yeah, not exactly a tempting value proposition.

For a card that is in an entirely different market segment, to pack this kind of punch, well, makes sense for AMD to be so quiet about it. Expect AMD to gain tons of marketshare everywhere outside the US where hardware isn't as cheap, $200-240 for this kind of performance is insane. Hello 4850/70, we meet again.


I'm still expecting 390/390x performance on the 29th which is quite the deal for $200-240 (plus up to date connectivity and lower power consumption). Anything more than that, I'll probably be finding my 290 a new home for the time being... or get excited for Vega next year.

 

[DisarmedDespot]

I'm trying to be a cynic. I'm still expecting 390 performance for $100 less, and I'll be happy with that.

...But the hype train is really, really tempting. I don't want to make a bet (surely someone on this forum will remember it if I did...), but if OC'd 480s on air match stock 1070s, I might make a shrine to Raja Koduri.

 

[JDG1980]

Some shops are listing retail spec for custom RX 480s...

This one is ridiculous if true.

1.6ghz and 10Gbps GDDR5..

I don't believe it though, since GDDR5 doesn't like running above 9Gbps.

It would only make sense if its the special sauce $299 model with GDDR5X..

I believe someone from AMD did confirm that Polaris 10's memory controller supports GDDR5X. It makes sense that a highly OC'd version could benefit from that.

1600 MHz would be awesome, if true. That could rival stock GTX 1070.

 

[selni]

I'm trying to be a cynic. I'm still expecting 390 performance for $100 less, and I'll be happy with that.

...But the hype train is really, really tempting. I don't want to make a bet (surely someone on this forum will remember it if I did...), but if OC'd 480s on air match stock 1070s, I might make a shrine to Raja Koduri.

Just wait for reviews/release - nearly there and it's not like buying a 1070 right now is easy anyway. It's not impossible but that's an enormous overclock if stock performance is where we're expecting and there's been way too many misleading overclocking claims from both companies in the past to believe anything without independent testing.

 

[R0H1T]

In case of Nvidia, Gaymerz matter alot, look at their financial statement.

If AMD wants to beat Nvidia in terms of perf/w, POLARIS HAS TO BEAT PASCAL WHEN BOTH OC IN TERMS OF PERF/W.

Polaris beating pascal perf/w at stock simply won't cut it anymore. As of now, Pascal is twice times perf/w of Fiji.

AMD claimed UP TO 2.8x perf/w from the previous generation ( not sure which ), so in order to beat Pascal when OCed in terms of Perf/w, Polaris maximum power consumption cannot go higher than 90W when OC ( 10% less perf than Fury X ). I expect Polaris will scare well with voltage both perf gains and power consumption. Probably why AIBs have 6 + 8 pins card lying around.
http://www.techpowerup.com/reviews/ASUS/GTX_1080_STRIX/22.html

No it doesn't, especially when you're paying 200$ for AMD & nearly 2x for the nearest competitor & still not getting twice the performance from GTX 1070.

That's what you're saying, wonder how this got lost in all the noise of Maxwell being a great OCer & when that left the perf/W by the wayside? Not 2x as the most efficient Fiji that is Nano.

Every other chipmaker claims such stuff, including your truly Intel, Nvidia, Apple, QC et al. The up to* is there for a reason cause not all apps & games are the same & with varying level of workloads you can only show the best & worst scenarios, even averaging is not possible or ideal because DX11 & DX12 can also cause huge anomalies.

Just as they have 8+8 pin for many custom 1080's & they still won't get past 2.1GHz on the majority of them

 

[Silverforce11]

The perf/w is going to make RX 480 so popular with miners. Right now, it's still very profitable on older GCN that's power hungry, but on RX 480, the minus electricity bill is going to double its profitability.

And it's cheap. I expect 4GB models will be utterly sold out with miners hoarding it to the max.

Once that's sold out, they will pay extra for the 8GB models as well, because it's only a bit more expensive. :/

 

[DisarmedDespot]

The perf/w is going to make RX 480 so popular with miners. Right now, it's still very profitable on older GCN that's power hungry, but on RX 480, the minus electricity bill is going to double its profitability.

And it's cheap. I expect 4GB models will be utterly sold out with miners hoarding it to the max.

Once that's sold out, they will pay extra for the 8GB models as well, because it's only a bit more expensive. :/

Yeah, that's what I'm afraid of. TBH I'm hoping the DAO explosion sinks Ethereum's price enough that less people hop on the mining wagon. Last thing AMD needs is another round of cards getting snatched up by miners, then the used market flooded once the craze passes.

 

[R0H1T]

The perf/w is going to make RX 480 so popular with miners. Right now, it's still very profitable on older GCN that's power hungry, but on RX 480, the minus electricity bill is going to double its profitability.

And it's cheap. I expect 4GB models will be utterly sold out with miners hoarding it to the max.

Once that's sold out, they will pay extra for the 8GB models as well, because it's only a bit more expensive. :/

At this point in time it's just better to invest in ether directly IMO, especially if all the RX 4xx sell out in the coming months due to miners.