[Guru3d] Radeon RX 470 Benchmarks

[swilli89]

If it actually had 290 performance, they would likely advertise it that way, and not compare it to a 270X, imo.

I'm guessing that it is actually in the 270X ballpark overall.

Do you not see that the scores are double in some instances? How is that in the same ballpark overall?

 

[Thala]

I'm not a technical wizard, in fact quite the opposite. If you think the rumoured 230-265mm^2 or so Polaris 10 should have this massive deficit to the 314mm^2 Pascal while also concluding that that this deficit when compared to 28mm is equal or has lessened, not worsened, then convince me please. I want to believe.

I do not get you. What precise metric are you using to conclude a deficit? Don't you see the huge gains in perf/W or what is your issue?

Comparing die sizes for totally different clocked parts is completely moot. Of course you get away with smaller die size if you clock higher. Or you can put more ALUs into your chip and clock lower. Typically the second approach might give you a power advantage while might have a cost disadvantage.

Look at the Hitman score. Double the performance while significantly reduced power....giving a 14nm + architecture gain of more than factor 3! That's by all accounts impressive.

 

[SPBHM]

great performer for the very reasonable price and with power usage well under control, that's the AMD/ATI I used to love, not the 300W and $400+ cards one.

this card might have an easy time against the competition, even if the 960 gets a price drop I can't see it doing much against the 470.

 

[Despoiler]

No. I am not comparing Polaris to Pascal.

I am comparing Polaris to Hawaii. I am also comparing Pascal to Maxwell.

Then I am clearly seeing (keep in mind based half on rumours half on AMD's figures) that Nvidia has further pulled away from AMD.

You are comparing architectures in your premises and then end by comparing companies. That means you are comparing architectures between companies.

To echo what Thala pointed out. You should not even compare architectures within the same company between different generations in the method that you are. Die size and transistor count aren't metrics by themselves. There needs to be some other data point to make a valid comparison.

 

[kraatus77]

at least wait till 29th before declaring arch fails. perf/mm doesn't mean that much as long as they can deliver better p/w and p/$. nobody buys gpu based on how much silicon they are getting.

remember how hawaii had more p/mm than tahiti on 28nm, it can happen again with 384bit die or something like that. i know so far benchmark results seem meh from architectural standpoint but at least wait till launch.

 

[Armsdealer]

The analysis posted is absurd. The entire reason the performance per watt metric exists is so to avoid the multitude of conclusions that can be reached when looking at stock vs oc vs temp vs die size across different architectures.

AMD is ahead of nVidia in perf / watt. Full stop. Pascal is showing 60-70% better perf / watt vs maxwell, which isn't bad. This would suggest they can create a 100 w actual tdp gtx 970 but what we're seeing with p10 is 100w gtx980+ performance and 70-80w gtx 970 perf in rx 470. P11 is likely to kick ass in notebooks as a consequence.

Shareholders should be happy. It remains to be seen just how good rx480 performs for enthusiasts when wattage is dialed up.

 

[maddie]

With those charts we have seen, it looks like the 480 is 390X territory, and the 470 is 390 territory, in 3D Mark?

Isn't this another side to what Nvidia has done. Two faces of the same.

AMD gives similar performance with a big price drop.

Nvidia gives greater performance at the same price.

The low/medium budget gamers will prefer the former. Remember that in the rest of the world, AKA most gamers, the 390/390X were never as cheap as the USA.

 

[crisium]

I do not get you. What precise metric are you using to conclude a deficit?

Look at the Hitman score. Double the performance while significantly reduced power....giving a 14nm + architecture gain of more than factor 3! That's by all accounts impressive.

Die size ratios, based on the rumoured sizes. Titan X to GTX 1080 shrink appears to be similar to the 390X/290X to 480 shrink, slightly more favorable to AMD.

An alternative way to look at it, in % Polaris 10 is closer to GP104 then Hawaii is to GM200, if you are interested in a direct comparison. In other words, with similar gains, you'd expect 470 to compare to 1080 the same as 290 compares to Titan X.

Yet the performance improvements, based on this, are much smaller for AMD.

It is true that you have less room to work with with a smaller die. 960 is half a 980, yet is 57% of the size for example. So I understand if AMD does not quite match the ratio improvement, but as far as I can see they quite behind from before because the rumoured ratio is slightly better here, but the performance gain is much worse.

Don't you see the huge gains in perf/W or what is your issue?

Yes, here they have almost certainly gained ground on Nvidia for sure, although I suspect Nv still has an edge but we will find out. Wattage is a hard barrier because you can only go so high. But what about die size?

I assume the max feasible die size of both companies is relatively similar. If Nvidia can do equal performance on smaller die, then they can have the performance crown if they want it. So not just from a curiosity or monetary standpoint, isn't there a hard limit that die size efficiency reveals for the top chip if both companies pursue a similar max size? This is my concern. If this is not a barrier, then who cares die size is 100% irrelevant. Is it a realistic barrier?

 

[Ma_Deuce]

We'll see what the wattage is, but based on the purported die size and SP count, it is a complete and utter disaster, and leaves no confidence that the company is capable of competing in DX11 whatsoever in any technical metric.

You are really going full tilt on the FUD. What do you have to gain from it?

If this card matches the posted benches, it's going to be an amazing product. If these have the same cooling and throttling issues that the 1080 FE has, it's a little less to be excited about. So far the rumors are saying the opposite though.

 

[LTC8K6]

Do you not see that the scores are double in some instances? How is that in the same ballpark overall?

You really should read the whole thread before replying to a single post out of context.

 

[Erithan13]

great performer for the very reasonable price and with power usage well under control, that's the AMD/ATI I used to love, not the 300W and $400+ cards one.

this card might have an easy time against the competition, even if the 960 gets a price drop I can't see it doing much against the 470.

Please just stop, we can't have reasonable posts in this thread, gotta keep that FUD train going!

Seriously though, great point about the 960, it's never been a compelling card for most people and the 470/60 are going to make it basically irrelevant. GP106 may prove to be a great competitor, when it finally turns up....AMD not having a high end response to the 1080/70 is a problem yes, but I'm betting it's going to prove way less of a problem that NV having nothing to show against Polaris. It's entirely possible NV are waiting to see how the Polaris launch goes before drip feeding some info about GP106 although as I see it AMD have the market cornered until you can actually buy a 1060 or 1050 and that looks to be some time away.

 

[lilltesaito]

Please just stop, we can't have reasonable posts in this thread, gotta keep that FUD train going!

Seriously though, great point about the 960, it's never been a compelling card for most people and the 470/60 are going to make it basically irrelevant. GP106 may prove to be a great competitor, when it finally turns up....AMD not having a high end response to the 1080/70 is a problem yes, but I'm betting it's going to prove way less of a problem that NV having nothing to show against Polaris. It's entirely possible NV are waiting to see how the Polaris launch goes before drip feeding some info about GP106 although as I see it AMD have the market cornered until you can actually buy a 1060 or 1050 and that looks to be some time away.

Nvidia will just start talking about the GP106 within a week of the 480 coming out or they even do something like the did with the 980ti and Furyx. Would not shock me if they have the cards ready and waiting.

 

[xthetenth]

Yes, here they have almost certainly gained ground on Nvidia for sure, although I suspect Nv still has an edge but we will find out. Wattage is a hard barrier because you can only go so high. But what about die size?

I assume the max feasible die size of both companies is relatively similar. If Nvidia can do equal performance on smaller die, then they can have the performance crown if they want it. So not just from a curiosity or monetary standpoint, isn't there a hard limit that die size efficiency reveals for the top chip if both companies pursue a similar max size? This is my concern. If this is not a barrier, then who cares die size is 100% irrelevant. Is it a realistic barrier?

We really need to know performance, size and wattage, because those can be traded off against each other. If the size is reasonable and the performance is a bit low for it but the wattage is great, then it's a really good sign, and for a max speed card they could hot clock its big derivative and get even more performance.

Also, how much of the size and energy budget is going into the RAM, and how does that compare to the scaling on the 1070/1080? That's important as well, and they're paying for a disproportionately beefy memory system.

 

[crisium]

If it's 1536SP then throw everything I said out the airlock, or nuke it from orbit.

 

[4K_shmoorK]

If it's 1536SP then throw everything I said out the airlock, or nuke it from orbit.

Perhaps seal the deal with the eating of some cat food? :sneaky:

 

[Thala]

Yet the performance improvements, based on this, are much smaller for AMD.

I thought we were over this already. You compare the same useless metrics.
This is in particular true since you completely taking the clock frequency difference and thus power out of the equation. Or more precisely, one can "invest" into power to get the same performance out of a smaller chip. Consequently perf/area will go up then.

Therefore taking power out of the equation is completely misleading. If you want do decide what the better approach is, larger lower clocked or smaller higher clocked chips is you need to look into perf/W... looking into anything else in isolation is moot.

 

[Attic]

If it actually had 290 performance, they would likely advertise it that way, and not compare it to a 270X, imo.

I'm guessing that it is actually in the 270X ballpark overall.

Agree with this.

 

[xthetenth]

They're going to talk about the 270X because they're talking to 270X buyers, that is people who look to pay that much for a graphics card. They may very well talk about performance comparisons as well.

 

[4K_shmoorK]

>Article graphs show 470 1.6x-2x+ performance of 270X
>470 will have comparable performance to 270X

 

[Armsdealer]

This thread is so stupid it's actually amazing.

It insists on a ridiculous conclusion on the basis of metrics that have no meaning.

If that weren't bad enough, it does so using numbers we don't have yet.

If THAT weren't enough, it does all of that while completely ignoring that we actually for once have hard data on perf / watt, a metric people actually obsess over and incontrovertibly does have meaning!

 

[antihelten]

>Article graphs show 470 1.6x-2x+ performance of 270X
>470 will have comparable performance to 270X

I guess we can now conclude that the RX480 will have comparable performance to the 1080

 

[crisium]

I thought we were over this already. You compare the same useless metrics.
This is in particular true since you completely taking the clock frequency difference and thus power out of the equation. Or more precisely, one can "invest" into power to get the same performance out of a smaller chip. Consequently perf/area will go up then.

Therefore taking power out of the equation is completely misleading. If you want do decide what the better approach is, larger lower clocked or smaller higher clocked chips is you need to look into perf/W... looking into anything else in isolation is moot.

You're right I'm working with incomplete info. That's the whole point, working with what I have. I think it's fun to work with theoreticals and make guesswork from a mix of rumour and press releases to see if AMD is making gains (no fun allowed armsdealer).

Nv clocked their new cards 50%+ over their old, although the max OC is closer to 40%. AMD is only +20% over the 390X based on 480 numbers. That itself is the main reason for the leaked performance gains, and lower wattage does not necessarily mean they have the skies limit for frequencies. For now I am working with the numbers we have and assuming a similar overclocking margin that the competition has until proven otherwise

There are frequency limits from the architecture too that can be somewhat independent of power consumption, e.g. Pitcairn uses much fewer watts but doesn't get higher clocks than Tahiti. A lot of Shaders that can't (relatively) clock high is perhaps why we don't see better gains. Those Shaders take up die spice - frequency does not, so getting performance with more and more shaders takes up more and more space - again that's my worry the hard limit of a die size.

1600-1700Mhz+ max OC (with voltage tuning) on a typical aftermarket would shut me up fast

since that's a 40% gain over typical Hawaii and means AMD is really stressing PPW and there is plenty of room for performance per die. We shall see. Remember, I want to be convinced I'm wrong so I appreciate it (I want a big fat Vega that competes with the next Ti).

 

[Bacon1]

1600-1700Mhz+ max OC (with voltage tuning) on a typical aftermarket would shut me up fast

since that's a 40% gain over typical Hawaii and means AMD is really stressing PPW and there is plenty of room for performance per die. We shall see. Remember, I want to be convinced I'm wrong so I appreciate it (I want a big fat Vega that competes with the next Ti).

Hawaii cards were around 1100mhz factory OC'd. "Reference" was 1050 for 390x and 1000 for 390 / 290x and 947 for 290.

1600 = 45% OC over 1100, 52% over 1050
1700 = 55% OC over 1100, 62% over 1050

I have a feeling you will be disappointed in Polaris no matter what.

 

[Thala]

1600-1700Mhz+ max OC (with voltage tuning) on a typical aftermarket would shut me up fast

Look, you still do not understand. Going wider with lower clocks is a design decision deliberately done by AMD. It does _NOT_ mean you can overclock Polaris to insane levels.

My main issue with your argument is, that you are taking perf/area as basis - an irrelevant metric - and are only happy (see you 1700Mhz overclock statement) until perf/area of Polaris matches Pascal. Not going to happen.

Even worse, based on your flawed perf/area metric you claim Pascal is somehow the better architecture compared to Polaris. This could very well be the case if Pascal is better in Perf/W than Polaris. However the numbers in this very thread point to the assumption that this is a close battle.

 

[coercitiv]

Even worse, based on your flawed perf/area metric you claim Pascal is somehow the better architecture compared to Polaris. This could very well be the case if Pascal is better in Perf/W than Polaris.

Perf/W is so last year...