Official AMD Polaris Review Thread: Radeon RX 480, RX 470, and RX 460

[flash-gordon]

You guys are prejudging GloFo based on their past performance. Having played with my rx480 OCing and undervolting.. it's a great chip. This is Samsung's 14nm FF.. it's good stuff.

AMD and Nvidia are closer on relative performance/efficiency now then they were back on the same 28nm process.

Isn't TSMC process oriented for higher performance and GloFo for lower power? I bet the second is cheaper overall and AMD is pushing to the edge, hence the higher TDP.

 

[sirmo]

Isn't TSMC process oriented for higher performance and GloFo for lower power? I bet the second is cheaper overall and AMD is pushing to the edge, hence the higher TDP.

AMD had worse efficiency issues (pushing to the edge as you put it) with 28nm when both AMD and Nvidia were on TSMC's 28nm.. this looks improved on 14nm.

AMD's architecture with its command processor, ACEs, FP ratio is just less efficient in DX11 games. But the fact that they have narrowed the gap somewhat by coming to 14nm tells me the process at GloFo is just as good if anything compared to TSMC. rx480 just uses 20% more power than 1060.. and when the potential can be unlocked a la Doom Vulkan it also performs 20% better. So they are about equal when full GCN is utilized.

Apple has used both TSMC and Samsung's finfet process (although previous version of it).. these have both been improved since.

GloFo and Samsung pool their fabs for big projects (someone posted a brochure while back where Samsung offers manufacturing at GloFo for big capacity dual sourcing projects), I doubt there is a big discrepancy between the two if any.

 

[psolord]

already posted
sorry
delete
https://www.reddit.com/r/Amd/comments/4tum2s/sapphire_rx_480_nitro_benchmarks_crysis_3_w3_gta/

 

[PhonakV30]

There has been a bug in the calculation, see the comments in the thread

https://www.reddit.com/r/Amd/comments/4tz24g/rx_480_strix_oc_8_gb_review_techpowerup/

https://www.techpowerup.com/forums/threads/asus-rx-480-strix-8-gb.224311/page-2

Don't take this review , They will update.

 

[dark zero]

Fake..
In GTA5 rx480 nitro at 1420Mhz is 27% faster than stock rx480 at 1200mhz.
Clock difference is only 18%...
Same with all games BTW.

Sadly... that is false.
The RX 480 can manage to have that performance since it was badly made the reference card.

The AIB card manages to get the true power and is really dangerous...
Definately with GDDR5X it would get 10% less than GTX 1070....

 

[Stuka87]

https://www.reddit.com/r/Amd/comments/4tz24g/rx_480_strix_oc_8_gb_review_techpowerup/

https://www.techpowerup.com/forums/threads/asus-rx-480-strix-8-gb.224311/page-2

Don't take this review , They will update.

Also looks like ASUS is using the same junk cooler that they are using on the 1080 where only two heat pipes (out of 5) make full contact.

 

[tviceman]

The AIB card manages to get the true power and is really dangerous...
Definately with GDDR5X it would get 10% less than GTX 1070....

You are absolutely wrong on this and it's not even close, not one bit. The RX480 has very little headroom, and it's lackluster overall performance has nothing to do with being bandwidth constrained.

https://www.techpowerup.com/reviews/ASUS/RX_480_STRIX_OC/26.html

13% memory overclock and a 7% overclock above the reference model equals....a 6% increase in performance. 13% memory overclock is 9040mhz, only 10.6% off current GDDR5X speeds. If it was bandwidth constrained, it would have scaled higher than 6% with a 13% memory overclock and 7% core overclock.

Everyone kept saying wait for AIB custom cards, it'll overclock much better with more power delivery. But that is being proven false right now and will continue to be proven false as more AIB custom card reviews trickle in. RX 480 is an overclocking dud, every bit as bad as Fiji was. Lets stop wishing and pretending it will magically grow wings.

 

[Silverforce11]

BF3.

I already know its bandwidth limited in games I play. But I do agree, its not going to get close to 1070 in DX11. Gap is too big and Polaris doesn't like >1400mhz.

 

[SPBHM]

the asus strix review is very disapoiting, also outside of the specifics of this model I can't help but notice the bad power usage of the 480 when it comes to dual monitor and blu ray, even compared to the r7 270... bad multi monitor power draw was a clear problem with my 5850, and such a long time after that they still haven't focused on getting it under control on all models like Nvidia seems to do.

480 might be bandwidth limited (since the memory compression is clearly behind even Maxwell),

but with stuff like this

it's difficult to expect a lot more

 

[sirmo]

You are absolutely wrong on this and it's not even close, not one bit. The RX480 has very little headroom, and it's lackluster overall performance has nothing to do with being bandwidth constrained.

https://www.techpowerup.com/reviews/ASUS/RX_480_STRIX_OC/26.html

13% memory overclock and a 7% overclock above the reference model equals....a 6% increase in performance. 13% memory overclock is 9040mhz, only 10.6% off current GDDR5X speeds. If it was bandwidth constrained, it would have scaled higher than 6% with a 13% memory overclock and 7% core overclock.

Everyone kept saying wait for AIB custom cards, it'll overclock much better with more power delivery. But that is being proven false right now and will continue to be proven false as more AIB custom card reviews trickle in. RX 480 is an overclocking dud, every bit as bad as Fiji was. Lets stop wishing and pretending it will magically grow wings.

Dude I own an rx480.. with a custom cooler. I hit 1385Mhz on stock voltage.. without issue. RAM also can be OCed pretty well (2200Mhz).

With power mods the reference hit 1500Mhz on water. It can definitely hit at least 1420Mhz on air.. I am 100% sure of that.

In Firestrike I hit ~14300 graphics score, here is an OCed AIB 1060 doing about the same: https://www.guru3d.com/articles_pages/geforce_gtx_1060_review,29.html

And I am not even fully OCed on VRAM, and I haven't even modified voltages. The only way I can explain TPU's numbers is that the card is severly power throttled as they haven't modified the limits.

 

[JDG1980]

https://www.techpowerup.com/reviews/ASUS/RX_480_STRIX_OC/26.html

13% memory overclock and a 7% overclock above the reference model equals....a 6% increase in performance. 13% memory overclock is 9040mhz, only 10.6% off current GDDR5X speeds. If it was bandwidth constrained, it would have scaled higher than 6% with a 13% memory overclock and 7% core overclock.

As much as I like TPU's reviews in general, their overclocking tests are lackluster because they test with exactly one game (BF3). We have no idea how that extrapolates to other titles. Some games will do better with higher core clocks, some with more memory bandwidth.

Another thing to keep in mind is that a memory overclock doesn't always give the same results as native memory of that same speed might. Memory timings are often tuned for optimization purposes, and clock changes can mess with this. If you took a good AIB R9 290X and overclocked the RAM to 1500 MHz, it wouldn't necessarily have the same performance as a R9 390X whose memory timings in the BIOS are designed for that.

 

[Silverforce11]

Holy cow, what's going on at TPU?

https://www.reddit.com/r/Amd/comments/4tzds0/tpu_asus_strix_radeon_rx_480_strix_review/d5lljjd

Regarding OC on the Strix 480:

That's what I'm doing. With stock fan, power and voltage.
What should I change?

He then mentions he can raise the voltage higher via software but didn't do it, nor power limit or fan speed, because it would have added to noise and power usage.

LOL?

 

[littleg]

He knows what he's doing, he's not an idiot or new to this. Hell, he even wrote the volt mod for Fury iirc. If he's doing this at stock settings, unlike his previous OC tests, there's a reason.

 

[PhonakV30]

Here from www.computerbase.de

https://www.computerbase.de/2016-07/asus-radeon-rx-480-strix-test/

Edit : Max Oc : 1330Mhz ?

 

[Silverforce11]

He knows what he's doing, he's not an idiot or new to this. Hell, he even wrote the volt mod for Fury iirc. If he's doing this at stock settings, unlike his previous OC tests, there's a reason.

Exactly, and pretending to be naive about it does him a disservice.

Nobody professional would OC AMD GPU with stock fan, stock volts or stock power limit unless they want it to look total crap.

Even at Computerbase, they didn't raised the voltage, but raising the power limit allows for >10% (Pascal-like!) performance gains from OC.

https://www.computerbase.de/2016-07/asus-radeon-rx-480-strix-test/4/#abschnitt_uebertaktbarkeit

 

[littleg]

Here from www.computerbase.de

https://www.computerbase.de/2016-07/asus-radeon-rx-480-strix-test/

Edit : Max Oc : 1330Mhz ?

Bottom of this page they get 1370 with 4400 on the memory I think:

https://www.computerbase.de/2016-07/asus-radeon-rx-480-strix-test/4/

 

[PhonakV30]

So If you look at this

From 1310/4000 to 1370/4400 , You get 11% boost.

 

[Tigel]

If I have followed this thread correctly it looks like Polaris scales better with Vulkan than Pascal.
Is this just the case for Polaris+Vulkan on Windows or is AMD's architecture generally better suited to thin APIs?

Assuming the latter is the case, could this be the reason why Apple chose AMD, because they are anticipating AMD's architecture to deliver better performance with their own proprietary thin API, Metal?

 

[Silverforce11]

If I have followed this thread correctly it looks like Polaris scales better with Vulkan than Pascal.
Is this just the case for Polaris+Vulkan on Windows or is AMD's architecture generally better suited to thin APIs?

Assuming the latter is the case, could this be the reason why Apple chose AMD, because they are anticipating AMD's architecture to deliver better performance with their own proprietary thin API, Metal?

Yes. OSX has already transitioned to Metal. It's like Vulkan/DX12, sharing a lot of similarity even on the code level.

And GCN excels in OpenCL workloads which Apple loves.

 

[Tigel]

Yes. OSX has already transitioned to Metal. It's like Vulkan/DX12, sharing a lot of similarity even on the code level.

And GCN excels in OpenCL workloads which Apple loves.

Yes, I'm aware of that, I just wasn't sure wether Polaris has the edge over Pascal in all thin APIs but it looks like it has.

Two years ago it looked like mobile dGPUs would soon be a thing of the past and now it looks like they'll be more important than ever, considering the news around Intel's high end iGPUs has gone awfully quiet (maybe they bet on the wrong horse and optimized for DX11/OGL?) and Apple plans to do their "AI"/Deep learning services locally on the customer machines.

Very interesting developments all in all.

 

[3DVagabond]

If I have followed this thread correctly it looks like Polaris scales better with Vulkan than Pascal.
Is this just the case for Polaris+Vulkan on Windows or is AMD's architecture generally better suited to thin APIs?

Assuming the latter is the case, could this be the reason why Apple chose AMD, because they are anticipating AMD's architecture to deliver better performance with their own proprietary thin API, Metal?

AMD is capable of better hardware scheduling where nVidia needs software support for a lot of what GCN does natively.

 

[railven]

So If you look at this

From 1310/4000 to 1370/4400 , You get 11% boost.

Is that chart power consumption or performance / watt?

EDIT: NVM, Google Translate finally kicked in. Swear thing is hit or miss at work.

 

[dacostafilipe]

Assuming the latter is the case, could this be the reason why Apple chose AMD, because they are anticipating AMD's architecture to deliver better performance with their own proprietary thin API, Metal?

You can't compare Metal to Vulkan/DX12. Metal is "just" OpenGL with less overhead, where Vulkan/DX12 are different APIs with a different approach.

 

[USER8000]

First review for the Sapphire RX480 Nitro+ 4GB:

http://www.pcworld.com/article/3098...laris-rethought-and-refined.html?page=1#toc-1

Overclockers UK TV review the Sapphire RX480 Nitro+ OC 8GB:

https://www.youtube.com/watch?v=jTFDJODCSCc

 

[sze5003]

Surprised they didn't test against a custom 1060. Seems decent but not sure about 1440p.

They also mention that the 8bg model will be more powerful with a better overclock.