AMD Radeon RX 480 (14nm Polaris)

[DanceswithUnix]

I'll put money on vega having an AIO cooler, and it ought to have better stepping and a beefier power supply - a 15% OC with all of that going for it shouldn't be hard. It'll probably need to beat the Nvidia 1080ti by then though, so AMD may not be able to tap into the ultra-high margin upper end of the market.

Well Polaris does seem sensitive to memory speed, so HBM2 could make all the difference. It doesn't look like the comfortable lead one would hope for though when the 1080 is already available.

In fact, if you work out the per clock speedup from the GCN1.2 of an R9 380 to an FX480 and apply that to the performance of a Fury (as that is basically GCN1.2 and should have the same number of shaders as Vega and already has plenty of memory bandwidth) adjusting for clock speed then you can probably guess the performance of Vega. But that involves maths and I've done enough today, time to fire up a game and shoot stuff

[CAT-THE-FIFTH]

OcUK have the Sapphire RX 480 Nitro on pre-order for £249.95 delivered:

https://www.overclockers.co.uk/sapph...gx-37b-sp.html

It has a custom PCB AFAIK with an 8 pin power connector,user replaceable fans and even LEDS!!

Hopefully this better and more consistent performance than the reference 8GB card.

Also:

It is panic that has no doubt being created by parties to try and hurt AMD sales.

The facts are all modern PCI-E gaming cards can and do exceed PCI-E limits on the voltage, we have not seen any failures on motherboards from this at all for past several years. All the current NVIDIA and AMD cards can draw more than PCI-E limit and no damage has being caused.

It is nothing new, but for some reason some select individuals trying to make it sound like a bigger problem than it is and one which only effects RX 480 when in reality all cards are equally effected.

I would not worry, of course some people will still be frightened by such scare mongering and as always we pride our selves on customer service so anyone who does not want to chance it can simply return their card, but cards such as 960, 970, 380, 390 etc. can also draw more power than PCI-E specification and no one made so much noise about those cards.

Also AMD have already announced a fix for those who want to stay within PCI-E regulations and the driver that resolves this shall be released on July 5th, so there really is nothing to worry about.

AMD had said a driver will be released on the 5th to sort any issues out.

Edit!!

AIB RX480 4GB cards will be released at the end of the month or early August it appears.

Yes looking around end of July / August time for AIB 4GB's.

[watercooled]

Good on Gibbo for quashing the scare-mongering and kudos to lots of the mainstream sites (Hexus included) for not jumping on the scaremongering bandwagon. Some people just can't see AMD doing well with something without coming up with some obscure metric where it's imperfect and making out like it's catastrophic.

[watercooled]

Well Polaris does seem sensitive to memory speed, so HBM2 could make all the difference. It doesn't look like the comfortable lead one would hope for though when the 1080 is already available.

In fact, if you work out the per clock speedup from the GCN1.2 of an R9 380 to an FX480 and apply that to the performance of a Fury (as that is basically GCN1.2 and should have the same number of shaders as Vega and already has plenty of memory bandwidth) adjusting for clock speed then you can probably guess the performance of Vega. But that involves maths and I've done enough today, time to fire up a game and shoot stuff

Using your estimates:
https://www.techpowerup.com/reviews/AMD/RX_480/24.html
According to TPU overall score at 1920x1080, the 380 is 63% of the 480. So 100/63*100=158.7 i.e. the 480 has 158.7% performance of the 380. So take the Fury X's 121 relative performance and multiply by 158.7% and we get about 192 which would put it above the 1080?

I've not considered clock speed, just end performance, so of course this assumes Vega 10 will match the 480's clock speed which isn't an unreasonable expectation IMO. And obviously it assumes identical architectural performance - considering this is Vega and not Polaris I'm guessing there will be some improvements?

Edit: D'oh! Took me far too long to realise the 480 has far more shaders than the 380 so ignore the above for now!

OK so a bit of a cheaty way to do it, lets estimate the performance of a theoretical 300 series with 2304 shaders: 81 relative TPU performance. So putting that back through the calculations gives us about 150 relative performance, same as a 1070. That's estimates based on estimates based on estimates though, so take it with a large pinch of salt.

Or an even simpler one assuming perfect scaling with shader count I get about 177 (which just happens to match the 1080 weirdly!). So without microarchitectural improvements and clock speed changes we're probably ballparking between the 1070 and 1080.

[kryton666]

If the partner cards are going to be £250 or more with more power hardly the low power budget option anymore is it!

[Xlucine]

Eh, it's not as if Nvidia is offering anything as modern for that much. A single 8 pin connector is hardly an onerous requirement either

[scaryjim]

... A single 8 pin connector is hardly an onerous requirement either

I don't think it's so much the need for an 8pin connector as the implication that this isn't really a forward step in perf/watt. Compared to AMD's last few $199 launch cards it's clearly a perf/watt improvement (although it's a shame we didn't have an R9 380 in there for the proper segment comparison!), although that's largely managed by brute force on a 2 node jump. But once you start comparing it to nvidia, and even Fiji, it's not so hot - it probably just about matches maxwell's perf/watt, but I suspect it's got worse perf/watt than Fiji in its Nano form, and it's clearly a lot worse than Pascal. This really isn't the remarkably efficient card we were all lead to believe, and given the leaked specs for the RX 460 I'm concerned about the performance levels it's going to manage whlst being bus-power limited.

Let's be clear here - the performance for a $199 launch-price card is great - it leaves the 7870/R9 380/GTX 960 eating its dust. But the perf/watt - this huge efficiency increase that AMD were trumpeting? That's just not there.

[cptwhite_uk]

Have we heard any reliable rumours about the RX 470 specs and release date yet?

[CAT-THE-FIFTH]

I don't think it's so much the need for an 8pin connector as the implication that this isn't really a forward step in perf/watt. Compared to AMD's last few $199 launch cards it's clearly a perf/watt improvement (although it's a shame we didn't have an R9 380 in there for the proper segment comparison!), although that's largely managed by brute force on a 2 node jump. But once you start comparing it to nvidia, and even Fiji, it's not so hot - it probably just about matches maxwell's perf/watt, but I suspect it's got worse perf/watt than Fiji in its Nano form, and it's clearly a lot worse than Pascal. This really isn't the remarkably efficient card we were all lead to believe, and given the leaked specs for the RX 460 I'm concerned about the performance levels it's going to manage whlst being bus-power limited.

Let's be clear here - the performance for a $199 launch-price card is great - it leaves the 7870/R9 380/GTX 960 eating its dust. But the perf/watt - this huge efficiency increase that AMD were trumpeting? That's just not there.

Its a C7 stepping meaning there have been issues,so that indicates to me the RX480 is probably overvolted,and pushed past its optimal performance/watt window - a bit like the R9 290/390 series were.

Also,Polaris 11 is what the RX460 is based on,and Raja Koduri said that it was actually the lead chip of the generation in a PCPER interview in the last few days,and the one they were targetting towards laptops. I also expect it will be a lower clockspeed part.

Considering the bus powered R7 360E is around GTX750 performance and performance/watt,it would be surprising if it is not any better IMHO!!

[scaryjim]

... Polaris 11 is what the RX460 is based on,and Raja Koduri said that it was actually the lead chip of the generation

...

Considering the bus powered R7 360E is around GTX750 performance and performance/watt,it would be surprising if it is not any better IMHO!!

Well, a full fat Polaris 11 would be more comforting - the AMD slides state that the RX 460 will have 14 CUs rather than 16. Wonder if the desktop is getting the lower binned dies that don't make the cut for laptop cards, or if they're simply yielding poorly and so they've had to fuse CUs off to make enough viable cards. Might mean we see an RX 460X when yields improve?

Either way, I think the RX 480 is a bit of a blind in terms of Polaris' abilities - it's a huge front end increase on a fairly minimal back end increase - the shading power is up 68% from R9 380 whilst the bandwidth only increased 25%. I guess the fact that it's pulling a ~ 60% performance boost over the R9 380 is pretty respectable at that point.

I guess we'll really find out how good Poaris is when the 470 and 460 get reviewed - given they're almost directly comparable to previous gen cards in terms of their shader and memory setups (RX 470 -> R9 380X, RX 460 -> R7 260X), the direct comparisons with those cards will tell us a lot more about what to expect from Vega and going forwards.

[CAT-THE-FIFTH]

Well, a full fat Polaris 11 would be more comforting - the AMD slides state that the RX 460 will have 14 CUs rather than 16. Wonder if the desktop is getting the lower binned dies that don't make the cut for laptop cards, or if they're simply yielding poorly and so they've had to fuse CUs off to make enough viable cards. Might mean we see an RX 460X when yields improve?

Either way, I think the RX 480 is a bit of a blind in terms of Polaris' abilities - it's a huge front end increase on a fairly minimal back end increase - the shading power is up 68% from R9 380 whilst the bandwidth only increased 25%. I guess the fact that it's pulling a ~ 60% performance boost over the R9 380 is pretty respectable at that point.

I guess we'll really find out how good Poaris is when the 470 and 460 get reviewed - given they're almost directly comparable to previous gen cards in terms of their shader and memory setups (RX 470 -> R9 380X, RX 460 -> R7 260X), the direct comparisons with those cards will tell us a lot more about what to expect from Vega and going forwards.

Interestingly the RX460 is a 4 ACE design like the RX480. The R7 260X which has the same number of shaders is a 2 ACE design.

BTW,according to AT Polaris is modified GCN1.2 - it will be interesting to see if Vega is actually a newer generation of GCN,ie,GCN1.3 as it is meant to be better performance/watt than Polaris.

[scaryjim]

... Might mean we see an RX 460X when yields improve? ...

Turns out (I stumbled across this a few minutes ago) that AMD will be using the last number to indicate the revision, so what we'll see, if we see it at all, is an RX 465.

OTOH, anyone hoping for a modicum of logic to the naming scheme will be disappointed, as it turns out the RX prefix is for a set of cards that meet technical and performance standards, and there will be versions of the 460 and 450 families that don't. So we'll get Radeon RX 460 and Radeon RX 450, but we'll also get Radeon 460 and Radeon 450 cards, that will (presumably) be slower. The slide reckons the technical cut off might be 100GB/s memory bandwidth, which would means > 6Gbps on a 128 bit GDDR5 interface. So I guess the RX 460 will be 896 shaders, 7Gbps, the RX 450 will be cut down (768 shaders, maybe?) and > 6.25Gbps, and the non-RX versions will be the same shaders but slower memory?

if that is the case, and they do release an RX 450, it means they're suggesting the second tier Polaris 11 card will be able to hit 60fps @ 1080p in eSports titles, and will be well under $99. Can't wait to see that....

[CAT-THE-FIFTH]

From what I have been gathering,the RX480 seems to settle down to sub 1.2GHZ clockspeeds. The Sapphire RX480 will be apparently running close to 1.35GHZ. That would mean possibly another 15% increase in performance there - that should be Fury level performance I suspect.

[Bagnaj97]

So I guess the RX 460 will be 896 shaders, 7Gbps, the RX 450 will be cut down (768 shaders, maybe?) and > 6.25Gbps, and the non-RX versions will be the same shaders but slower memory?

if that is the case, and they do release an RX 450, it means they're suggesting the second tier Polaris 11 card will be able to hit 60fps @ 1080p in eSports titles, and will be well under $99. Can't wait to see that....

RX460 has been confirmed at 16CUs afaik. Assuming the CUs are the same as on Polaris 10 that's 1024 shaders. That's the same number of shaders as a 7850 and the clockspeed should be higher - 1080p 60fps in eSports titles (LoL/Dota2/CS:GO) should be well within reach. Memory bandwidth will be lower, but Polaris has memory compression to recover a bit of that.

[DanceswithUnix]

Had a look at the Amazon uk top selling computer bits list. Lots of micro sd cards and memory sticks in there, and a single video card at #88: a Sapphire RX480. Must be selling fairly well then.

https://www.amazon.co.uk/gp/bestsell..._computers_0#5

https://www.amazon.co.uk/Sapphire-21...s_computers_88

[watercooled]

I don't think it's so much the need for an 8pin connector as the implication that this isn't really a forward step in perf/watt. Compared to AMD's last few $199 launch cards it's clearly a perf/watt improvement (although it's a shame we didn't have an R9 380 in there for the proper segment comparison!), although that's largely managed by brute force on a 2 node jump. But once you start comparing it to nvidia, and even Fiji, it's not so hot - it probably just about matches maxwell's perf/watt, but I suspect it's got worse perf/watt than Fiji in its Nano form, and it's clearly a lot worse than Pascal. This really isn't the remarkably efficient card we were all lead to believe, and given the leaked specs for the RX 460 I'm concerned about the performance levels it's going to manage whlst being bus-power limited.

Let's be clear here - the performance for a $199 launch-price card is great - it leaves the 7870/R9 380/GTX 960 eating its dust. But the perf/watt - this huge efficiency increase that AMD were trumpeting? That's just not there.

Assuming they've carried over power saving technologies from the Fury line I agree it doesn't look that great as they're getting very little from the node jump. The Nano is a funny one as it was I assume a fairly low-volume, conservatively-clocked part so you'd expect it to be more efficient than normal, but even vs the other Fury parts it's not massively more efficient. Assuming AMD didn't somehow regress on architectural efficiency, I could only guess the node or AMD's use of it isn't (yet?) optimal.

But from an objective point of view, it's a decent jump for AMD in that performance/price range which is what matters, I guess. Maybe we'll see a revision (485?) with a clock speed bump when the process improves?

I'm just left wondering about that Polaris power demo, which was Polaris 11 IIRC? Hopefully that will work out a bit better for them. It'll also be interesting to see if efficiency improves a bit with partner cards as it occasionally does. I remember my overclocked HIS 4850 was a fair bit more efficient than the already fairly efficient (at the time) stock card.