AMD: Speculation, Rumors, and Discussion (Archive)

[Deleted member 13524]

Well, so far it looks like Polaris 10 is pretty rubbish performance wise. It has the same bandwidth as GTX 1070 but with radically lower performance.

"Polaris 10 is pretty rubbish because it's similar to the much bigger GP104 model with not so much bigger bandwidth"?

Is this a troll post or some joke I missed?

 

[3dilettante]

This leaves memory bandwidth and ROPs as possible culprits. The bandwidth drop (256 vs. 320GB/s) isn't so severe, meaning that in most cases the framebuffer compression should be able to compensate that easily. Does anybody believes Polaris 10 drops the number of ROPs to 32? Could there be a change to the ROP caches (some slides mention some changes on the L2 Cache)?

Keeping Hawaii's ROP count on a 256-bit bus would produce a ROP/channel ratio higher than I've found so far for GCN GPUs.

As far as bandwidth goes, the 380X was able to match performance with the 7970 despite a larger bandwidth disparity, with at least an overclocked model.
There seems to be some evidence that AMD's DCC might not be as good as Nvidia's, but it was sufficient in its initial form to overcome a larger bandwidth disparity.

The Polaris marketing mentioned the L2 as being changed, but the render back end section was not marked as having anything new.
Non-synthetic tests seemed to have a reduced sensitivity to ROP disparities, and there's a rumored Vega specification that shows it hasn't moved beyond what Hawaii and Fiji have for ROPs.

A pixel throughput gap might be a way to counteract being equivalent or better in most other ways.

 

[Razor1]

Are you 100% sure of this?

AFR always does require double the resources (each card needs all information in their own memory), what doesn't double are things like AF, and AA.

 

[Kaotik]

Keeping Hawaii's ROP count on a 256-bit bus would produce a ROP/channel ratio higher than I've found so far for GCN GPUs.
The Polaris marketing mentioned the L2 as being changed, but the render back end section was not marked as having anything new.

Then again, 4th gen GCN is supposed to be the biggest change in GCN's history, so nothing's impossible, and "not new" isn't same as "not more per memory channel than before"

 

[Kaotik]

AFR always does require double the resources (each card needs all information in their own memory), what doesn't double are things like AF, and AA.

I think he was referring to AotS using AFR, not if AFR requires doubled resources

 

[3dilettante]

Then again, 4th gen GCN is supposed to be the biggest change in GCN's history, so nothing's impossible, and "not new" isn't same as "not more per memory channel than before"

It could be different, it just seems odd that a seemingly milder change from Tahiti to Tonga allowed for the newer architecture to win despite a larger bandwidth deficit. It would also imply more rasterizer width per channel as well, since the two match. The rasterizer block, for what it's worth, isn't marked as being changed either.

Hawaii has some other differences besides no compression, like a lower 4xFP16 pixel throughput in synthetics that might allow a lower ROP count to do better than expected.

 

[Anarchist4000]

That's simply because there's nothing to see there, Polaris doesn't support HBM-memories, since it supports GDDR5-memories, you can't fit both in the same chip.

AMD has stated that HBM is supported by the architecture, but if a variant was using it they either canned it or did something interesting with the memory controller.

 

[Jawed]

I wonder if Polaris 10 has 32 ROPs and AMD was counting on higher clocks from the change to 14nm.

 

[Deleted member 2197]

More Info on Radeon RX 460 - 470 and 480 from new Slide deck

So I just received a new slide-deck from AMD moments ago, in which they talk about the Radeon RX 460 - 470 and 480 in a PDF. The good news, It's cleared to post these, the bad news it not a lot more info; yet I did want to share some slides with you as some new stuff can be derived, extrapolated and thus confirmed from these slides

So the slide-deck basically is all about VR, but then all of the sudden mentions a nice sub-set of Polaris 10 and 11 details that I wanted to run you guys by. So Polaris 10 will end up being the GPU for the Radeon RX 480 and likely RX 470 which we all expected of course. The Polaris 11 GPU would be used for the Radeon RX 460, so far you guys are with me okay ?

AMD has presented they will release three new graphics cards, two have been added to the product stack: the mainstream Radeon RX 470 and entry level Radeon RX 460. The two lower specced cards are obviously positioned below the previously announced RX 480. The video cards all do have a low energy consumption according to AMD, thanks to a die-shrink at 15/16nm FinFET. The RX 460 can even be powered from just the PCI Express slot. And with no need for a PCI Express power connector the card thus has a TDP of less than 75 watts. According to rumors the RX 470 is based on Polaris 10, which also is the basis of the RX 480 with its 150 Watt TDP. Probably, the GPU from the RX 470 has slightly lower clocks and some shaders clusters cut away. The RX 460 is built around the slower Polaris 11 GPU. No specifications were announced. However, now we go to the next slide:

http://www.guru3d.com/news-story/sl...eon-rx-460-470-and-480-in-new-slide-deck.html

 

[Kaotik]

AMD has stated that HBM is supported by the architecture, but if a variant was using it they either canned it or did something interesting with the memory controller.

Yes, they have memory controller blocks fit for GCN, but they haven't made Polaris chips with them, since we know there's only 2 chips both carrying GDDR5

 

[Kaotik]

It could be different, it just seems odd that a seemingly milder change from Tahiti to Tonga allowed for the newer architecture to win despite a larger bandwidth deficit. It would also imply more rasterizer width per channel as well, since the two match. The rasterizer block, for what it's worth, isn't marked as being changed either.

Hawaii has some other differences besides no compression, like a lower 4xFP16 pixel throughput in synthetics that might allow a lower ROP count to do better than expected.

AMD's new slides confirm that Polaris 10 is 36 CU total (while Polaris 11 is 16 CU), maybe they've gone to 6 x 6 CU configuration for example, instead of 4 x 9 or even worse, 2 x 18?

 

[3dilettante]

At least traditionally, the architecture would trend towards having at least 8 ROPs per shader engine. Hawaii associates its render back ends to specific channels, whereas Tahiti (and probably Tonga if had all channels active) used an additional crossbar in order to allow its 32 ROPs to work with a larger number of channels that did not divide evenly.
For the 6-engine case, I wouldn't know if it would work in the other direction.

 

[CSI PC]

More Info on Radeon RX 460 - 470 and 480 from new Slide deck
http://www.guru3d.com/news-story/sl...eon-rx-460-470-and-480-in-new-slide-deck.html

It is going to be interesting to see if the 8GB will be launched at same time as the 4GB, I could be reading too much into this though from Guru3d.

The card will start at 199 USD for the 4GB model. 8GB models will also become available.
The product did not get a launch date.

I assume the lack of a launch date was applicable to the 8GB (also re-inforced by the "will also become available") as I assume we all expected the 480 launch when NDA ends June 29th.
Cheers

 

[Jawed]

Steam VR test score of 6.3:

http://videocardz.com/61064/amd-polaris-10-and-polaris-11-specifications

I have no idea how AMD thinks that's "premium VR".

 

[Kaotik]

Steam VR test score of 6.3:

http://videocardz.com/61064/amd-polaris-10-and-polaris-11-specifications

I have no idea how AMD thinks that's "premium VR".

It's above the performance for Vive's & Rift's "recommended VR spec"

 

[Jawed]

Actual performance numbers:

http://www.overclock3d.net/articles/gpu_displays/amd_release_rx_400_series_polaris_info/1

[...] thanks to the endnotes on AMD press deck.

[...]

The Polaris 10 number was calculated when the performance of two identical systems with either an R9 270X and RX 470 are compared. The R9 270W[sic] drew 180W of power while the RX 470 drew 110W.

Both GPUs were tested in Ashes of the Singularity and Hitman at high settings, where the R9 270X achieved 28.1FPS in Ashes and 27.6 in Hitman while the RX 470 achieved 46FPS in ashes and 60FPS in Hitman.

Apologies for the EDIT fest.

 

[CSI PC]

Steam VR test score of 6.3:

http://videocardz.com/61064/amd-polaris-10-and-polaris-11-specifications

I have no idea how AMD thinks that's "premium VR".

Also historically the 390 scored 7, so it is going to break some peoples expectation of it consistently bettering 390x in games.
Cheers

 

[SimBy]

"Polaris 10 is pretty rubbish because it's similar to the much bigger GP104 model with not so much bigger bandwidth"?

Is this a troll post or some joke I missed?

Some of us expected or hoped for 15% higher performance. So around Nano/Fury Pro level. Now for all we know P10 may still be an OC beast and AMD went for efficiency in reference designs and left the 'dirty job' to AIB partners. Or this leaks are all complete BS, that's also possible.

We shall see.

 

[Razor1]

I think he was referring to AotS using AFR, not if AFR requires doubled resources

Oh ok, yeah definitely uses AFR. Oxide mentioned it and also review sites mentioned it

http://www.anandtech.com/show/9740/directx-12-geforce-plus-radeon-mgpu-preview/3

Because Oxide is using a DX12 EMA AFR implementation here, this gives Ashes quite a bit of flexibility as far as GPU compatibility goes

 

[eastmen]

Also historically the 390 scored 7, so it is going to break some peoples expectation of it consistently bettering 390x in games.
Cheers

has it ? my 290 stuggles to hit 6 . So a 6.3 isn't bad imo