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

[Mopetar]

They'd need a pretty big step up from Tonga compression performance to equal Hawaii's bandwidth. A ~20% performance only means ~300 GB/s which is only 80% of what the 390X could do.

AMD would need another 25% improvement on top of what they achieved with Tonga. No idea whether or not something like that is possible.

 

[Bacon1]

No details on the video encode/decode part?

They were in previous slides that were leaked

https://youtu.be/5g3eQejGJ_A also talks about it a bit

 

[Maverick177]

Yep looks to be ~18-19% with Fury vs Hawaii and ~35-37% with Polaris, or almost the same gain with Polaris vs Fury that Fury had over Hawaii

That's pathetic.

 

[Stuka87]

They'd need a pretty big step up from Tonga compression performance to equal Hawaii's bandwidth. A ~20% performance only means ~300 GB/s which is only 80% of what the 390X could do.

AMD would need another 25% improvement on top of what they achieved with Tonga. No idea whether or not something like that is possible.

Need to remember that it does not need to match it. The 390X already had more memory bandwidth than it could use in the majority of cases. Being slightly below it should not have much of an impact. Plus this is not a card aimed at 4k resolutions, where memory bandwidth has the largest impact.

 

[Bacon1]

That's pathetic.

????

How is almost 40% memory compression over Hawaii pathetic??

That means @ 256 GB/s they can get 350 GB/s

 

[JDG1980]

There was some speculation on OCUK that the Primitive Discard Accelerator might make up for the low number of ROPs. Does anyone know enough about the low-level technical details of GPUs to say whether that post makes sense?

 

[JDG1980]

Also 36 CUs (2304 SPs) as expected.

I wonder what the footnote (8) is? I don't see it anywhere on the slide.

 

[jpiniero]

32 ROPs is probably to save space. AMD probably cut too deep but it's too late now. But since they are pricing it appropriately I don't see a problem with it.

 

[JDG1980]

Hopefully 32 ROPs will provide enough pixel fillrate to do upscaling to 4K in madVR while simultaneously powering two 1440p monitors in web browsing or other light tasks. My 7870 drops frames sometimes with madVR+SVP, but that usually happens when it runs out of RAM (only 2GB), which shouldn't be an issue on RX 480.

 

[Mopetar]

There was some speculation on OCUK that the Primitive Discard Accelerator might make up for the low number of ROPs. Does anyone know enough about the low-level technical details of GPUs to say whether that post makes sense?

From my understanding it makes sense, but it's a bit like hyper threading in that the tighter the code (the fewer things discarded) the less performance improvement you get.

It probably prevents cases where performance would normally tank due to spending time rasterizing things that won't be visible, but if there wasn't much of that going on to begin with, the primitive discard accelerator doesn't add much value.

I think the lack of ROPs won't matter much for 1080 gaming, but it may be part of the reason the VR score was reported as lower since the 480 be fillrate limited at higher resolutions.

 

[maddie]

Considering the fact that the 32 ROPs in Tonga outperforms the 64 ROPs in Hawaii, we should consider the possibility that a ROP is not a ROP is not a ROP.

Love the importance of adaptive clocking on this design.

 

[Silverforce11]

They'd need a pretty big step up from Tonga compression performance to equal Hawaii's bandwidth. A ~20% performance only means ~300 GB/s which is only 80% of what the 390X could do.

AMD would need another 25% improvement on top of what they achieved with Tonga. No idea whether or not something like that is possible.

What's the difference between a custom R290X that doesn't throttle and a 390X?

Not much at all, as any 290X user can attest.

290X has 320GB/s. It shows that 390X performance gains aren't from the big jump in bandwidth, because Hawaii already has more than it needs.

256GB/s with 40% more efficiency is ~350GB/s, already above the 290X class.

 

[Head1985]

What's the difference between a custom R290X that doesn't throttle and a 390X?

Not much at all, as any 290X user can attest.

290X has 320GB/s. It shows that 390X performance gains aren't from the big jump in bandwidth, because Hawaii already has more than it needs.

256GB/s with 40% more efficiency is ~350GB/s, already above the 290X class.

Yep and only reason if rx480 is slower than 390x is 32rops.I still cant belive only 32rops.Lets hope it is at 390x level with 32rops atleast.
BTW rx480 will be beast with 64rops.With fastest tessellation on market and other improvement it should be at FURYX/980TI level..Too bad it have 32rops.

 

[JDG1980]

Yep and only reason if rx480 is slower than 390x is 32rops.I still cant belive only 32rops.Lets hope it is at 390x level with 32rops atleast.
BTW rx480 will be beast with 64rops.With fastest tessellation on market and other improvement it should be at FURYX/980TI level..Too bad it is only 32rops.

I wonder if AMD's ROPs are really ginormous, with tons of transistors, and this is why AMD is so reluctant to provide enough. We've seen this happen recurrently: Tahiti (despite being a flagship card at the time) had only 32 ROPs, which was badly inadequate, and Tonga didn't fix this. Fiji was also badly shorted on ROPs, with a mere 64 when it should have had 128.

 

[Bacon1]

https://youtu.be/a3XRW67DuGo

Shows off Overclocking using the new tool, haven't watched whole thing (almost 1 hr), hard to view most of it too due to external camera and well.. no idea what he is saying

 

[Silverforce11]

Yep and only reason if rx480 is slower than 390x is 32rops.I still cant belive only 32rops.Lets hope it is at 390x level with 32rops atleast.
BTW rx480 will be beast with 64rops.With fastest tessellation on market and other improvement it should be at FURYX/980TI level..Too bad it have 32rops.

If it's ROP bound, you should compare the performance scaling drop as resolution rises.

ie, it will scale worse than 390X if you go from 1080 -> 1440p -> 4K.

If there's no worse scaling, it is NOT ROP bound.

 

[Shivansps]

What's the difference between a custom R290X that doesn't throttle and a 390X?

Not much at all, as any 290X user can attest.

290X has 320GB/s. It shows that 390X performance gains aren't from the big jump in bandwidth, because Hawaii already has more than it needs.

256GB/s with 40% more efficiency is ~350GB/s, already above the 290X class.

350GB/s is still a good 34GB/s less than a 390, for example thats more than a A10 APU has for its IGP, and that one can play games at FHD at low details...

 

[Elixer]

I wonder if AMD's ROPs are really ginormous, with tons of transistors, and this is why AMD is so reluctant to provide enough. We've seen this happen recurrently: Tahiti (despite being a flagship card at the time) had only 32 ROPs, which was badly inadequate, and Tonga didn't fix this. Fiji was also badly shorted on ROPs, with a mere 64 when it should have had 128.

So, you know more than the chip designers and the limitations that they had to work with?
Nothing is free in silicon, and you have to manage what you got.

They were giving a performance target at a specific price range, and that is what they came up with.

Sure, it would be nice to see a 1080 level of performance card for $199, but, that just isn't happening anytime soon.

 

[boozzer]

So, you know more than the chip designers and the limitations that they had to work with?
Nothing is free in silicon, and you have to manage what you got.

They were giving a performance target at a specific price range, and that is what they came up with.

Sure, it would be nice to see a 1080 level of performance card for $199, but, that just isn't happening anytime soon.

2 years from now, maybe. I would bet on 3 years because of the 199$ price tag. and I am 100% sure we will be stuck on 14/16 nm for 3+ years.

1 generational of products stretch over 3 years D: a repeat of 28nm is 100% likely.

 

[crisium]

Welp, if the shader efficiency improvement is true and it still loses/ties a 390X, then looks like we got another unbalanced design?

Where is the ROP defender guy again?

Or maybe AMD just exaggerated about shader efficiency.

We'll discover true performance tomorrow, but probably not answers to all of our questions.

 

[Silverforce11]

350GB/s is still a good 34GB/s less than a 390, for example thats more than a A10 APU has for its IGP, and that one can play games at FHD at low details...

You didn't read the post if you quote me and don't understand the logic.

290/X non-throttling cards ~= 390/X, especially if their core clocks are set the same.

290/X has 320GB/s. Hawaii does not need more than 320GB/s to achieve it's peak performance.

Thus, the 390/X bandwidth is just fluff and marketing. The gains came from non-throttling custom designs and 30-50mhz core clock bumps.

Instead of saying 390/X, just think of the chip, which is identical, Hawaii. Despite AMD claiming the chip has a different name, it's the same. Marketing differs to reality.

350GB/s effective is more than required for Hawaii class performance.

 

[Mopetar]

Considering the fact that the 32 ROPs in Tonga outperforms the 64 ROPs in Hawaii, we should consider the possibility that a ROP is not a ROP is not a ROP.

In what way? Tonga is listed as having about half the pixel fill rate as Hawaii and suggests that once you adjust for clock speed, there's no performance difference.

 

[Stuka87]

Yep and only reason if rx480 is slower than 390x is 32rops.I still cant belive only 32rops.Lets hope it is at 390x level with 32rops atleast.
BTW rx480 will be beast with 64rops.With fastest tessellation on market and other improvement it should be at FURYX/980TI level..Too bad it have 32rops.

There is no reason to increase the cost of the chip too add more ROPs. This is a 1080P card, 32 ROPs is the correct number for this resolution. Having 48 or 64 will not increase 1080P performance. It will only increase performance at higher resolutions. And since this card should never be used for 4k or the like, having more ROPs would increase the cost of the card for no reason other than to have a number on a spec sheet.

 

[Silverforce11]

Welp, if the shader efficiency improvement is true and it still loses/ties a 390X, then looks like we got another unbalanced design?

Where is the ROP defender guy again?

Or maybe AMD just exaggerated about shader efficiency.

We'll discover true performance tomorrow, but probably not answers to all of our questions.

What? If it ties a 390X, at it's perf/w, that's efficiency gained.

More performance at the same power, or the same performance at less power is "efficiency" improvements.

 

[Shivansps]

There is no reason to increase the cost of the chip too add more ROPs. This is a 1080P card, 32 ROPs is the correct number for this resolution. Having 48 or 64 will not increase 1080P performance. It will only increase performance at higher resolutions. And since this card should never be used for 4k or the like, having more ROPs would increase the cost of the card for no reason other than to have a number on a spec sheet.

This wanst marketed as VR Premium Card? You need more than 1080 perf for that.

Actually i now may understand the low Steam VR scores.