AMD: Speculation, Rumors, and Discussion (Archive)

[kalelovil]

Look like the cooler is way longer than the pcb..

If that cooler ever sees retail. The 3xx series were introduced with a similar cooler design, but all retail cards AFAIK ended up with custom AIB coolers.

http://wccftech.com/amd-officially-...0x-329-390-8-gb-cards-official-sight-r9-fury/

Edit: Although on second thought perhaps it will, considering this is new chip and we haven't see any non-reference cards leaked yet.

 

[jacozz]

With the information we have, there are two new chips in 2016. Polaris 10 and Polaris 11. Now, when we sort of know the performance class the RX 480 is in, what does it say about the high end segment? We know that Vega with HBM2 is coming early 2017, but with what will RTG compete in the high end in the meantime?


• Price dropped furys an furyX:es with 4 gb Vram and Dp 1.2 (really? Hope not)

• Polaris 10- Duo-card (Hope not. I never go the duo card route again unless the frame time issues is magically solved)

• Vega pushed forward to be released in october (Hope so)

• Nothing (would be kind of sad)

 

[homerdog]

Everyone's saying that the RX480 won't approach the 1070, but 5.5 vs 6.5TF isn't a huge difference IMO. Especially considering the 480 is *half* the price for the 4GB model (which is probably fine for this performance level).

 

[jacozz]

Or... Maybe, RX480 is a cut down version of Polaris 10 and the bigger brother waiting in the shadows to be leaked?

 

[pjbliverpool]

Everyone's saying that the RX480 won't approach the 1070, but 5.5 vs 6.5TF isn't a huge difference IMO. Especially considering the 480 is *half* the price for the 4GB model (which is probably fine for this performance level).

Traditionally AMD has been a fair way behind NV in performance/flop though. Fury-X is an 8.6TF card compared to Titan-X which is only 6.6 and faster. Polaris may change that of course.

 

[killeak]

Everyone's saying that the RX480 won't approach the 1070, but 5.5 vs 6.5TF isn't a huge difference IMO. Especially considering the 480 is *half* the price for the 4GB model (which is probably fine for this performance level).

Yes, but the 390X has 5.9TF and the 980ti 5.1TF. Comparing 2 GCN cards usign FLOPs it kind of makes sense, but you can't compare it the same way to a Maxwell/Pascal. I would expect that the real performance difference would be much higher than the difference in FLOPs.

 

[Rootax]

And flops are not everything... But, 1070 has a 150w tdp too, with I guess much better performances, so I wonder if GoFlo process is real bad or Polaris "leaks" a lot...

 

[lanek]

What's the source of this list?

I am a bit disappointed if the "Primitive Discard Accelerator" turned out to be a marketing name for conservative rasterization. This naming doesn't make much sense. I was hoping that AMD takes a big leap forward in their primitive processing performance. Nvidia is much faster rejecting invisible triangles.

Hopefully "Memory Compression" means that writes from compute shaders also support delta compression. It would be awesome. I already have some ideas how to exploit that (in sparse data structures).

I actually had a conversation regarding to instruction pre-fetch just a few weeks ago in Twitter. I didn't find any public documents describing GCN (1.0-1.2) instruction pre-fetch. This is important to know if you want to build a "jump table" style shader system. Sort all shaders by GPR count and bucket them to GCN occupancy classes (http://amd-dev.wpengine.netdna-cdn.com/wordpress/media/2014/05/gcn_vgpr_table.png). This results in 10 different shaders in total, allowing you to execute any amount of different shaders in just 10 compute dispatches. Nice trick for tiled/clustered lighting for example (especially when combined with deferred texturing). Simple pre-fetch (load first N instructions of each shader) is not a good idea for shaders like this. Pre-fetch after the jump would of course work just fine.

Well, the list come from WCCF.. should have maybe not adding it this way, ( was 5h in the morning ) but seems concern effectively triangles rejecting, hence the "general " name. Its a global name for all the primivites performances improvement. We will need to wait for all GCN4 "features" improvement. The name is pretty indicating what it do so..

 

[homerdog]

Yes, but the 390X has 5.9TF and the 980ti 5.1TF. Comparing 2 GCN cards usign FLOPs it kind of makes sense, but you can't compare it the same way to a Maxwell/Pascal. I would expect that the real performance difference would be much higher than the difference in FLOPs.

Oh I didn't realize the 390X was almost 6TF. Hopefully Polaris will be better in the FPS/Flop department.

 

[OlegSH]

Yes, but the 390X has 5.9TF and the 980ti 5.1TF. Comparing 2 GCN cards usign FLOPs it kind of makes sense, but you can't compare it the same way to a Maxwell/Pascal. I would expect that the real performance difference would be much higher than the difference in FLOPs.

980 Ti has 6.2 teraflops with its typical ~1100 Mhz boost frequency for reference card and 5.6 TFlops with base frequency which is nonexistent in real games, it also beats Fury X by 1.5X in ROP count, rasterizer blocks count, triangle setup count, etc, etc, so I hardly believe RX 480 could be on par with such monster as 980 Ti

 

[Infinisearch]

Well it seems the 480's clock speed is now public. 1266mhz

http://www.techpowerup.com/223043/amd-radeon-rx-480-clock-speeds-revealed-clocked-above-1-2-ghz

 

[Entropy]

And flops are not everything... But, 1070 has a 150w tdp too, with I guess much better performances, so I wonder if GoFlo process is real bad or Polaris "leaks" a lot...

Rx480 has a single six-pin molex, so 150W is the maximum that can be delivered to the card. You can safely assume that actual in-use power draw is significantly lower.

 

[Ext3h]

I am a bit disappointed if the "Primitive Discard Accelerator" turned out to be a marketing name for conservative rasterization. This naming doesn't make much sense. I was hoping that AMD takes a big leap forward in their primitive processing performance. Nvidia is much faster rejecting invisible triangles.

No source for this, just my intuition. "Primitive Discard Accelerator" might be partially(!) a software solution, based on the GeometryFX pre-processor they recently presented.

 

[Kaotik]

Or... Maybe, RX480 is a cut down version of Polaris 10 and the bigger brother waiting in the shadows to be leaked?

If we assume the old 232mm^2 leak is correct, about what size would it be on 28nm? That could give some clues on whether there's more units in there, or not

 

[lanek]

No source for this, just my intuition. "Primitive Discard Accelerator" might be partially(!) a software solution, based on the GeometryFX pre-processor they recently presented.

I was under the feeling that this Discard accelerator was mostly related to a change on the hardware side, who will imply new software features ( as presented during the first Polaris architecture presentation some month ago ) ( But honestly, i think we will need the review and presentation for it, along that the consoles developers analysis (remember other GCN in depth analysis ).

Not quite sure about the "GeometryFX", it is a "code" library, not much related to a direct change, evolution in the architecture http://gpuopen.com/geometryfx-1-2-cluster-culling/
http://gpuopen.com/gaming-product/geometryfx/

mbcnt
http://gpuopen.com/fast-compaction-with-mbcnt/

On the software side, GeometryFX (code extensions to DX ), can be used for cull triangles ( in some cases only i think ). ( work with GCN1.0 too )

Make me thing to that too http://gpuopen.com/unlock-the-rasterizer-with-out-of-order-rasterization/
( This make me ask me, if thoses "codes " was used by developpers before )

I dont see much a reason to cite it in an architecture improvement slide if it was just "geometryFX"

 

[Alexko]

Which is going to be ironic when the percentage of 480s actually ever hooked up to a VR headset will turn out to be the exact inverse of that 33/40 ratio.

121%?

 

[Silent_Buddha]

If we assume the old 232mm^2 leak is correct, about what size would it be on 28nm? That could give some clues on whether there's more units in there, or not

I suspect there's an X version waiting in the wings. It'd be strange if there was only a Rx 480 and no Rx 480x at some point. Ever since they've moved from the xx50/xx70 naming schemes, they've almost always had a ### and ###X version. I think someone also mentioned that during the keynote it was mentioned the price stack for Polaris would go from 100 to 300 USD. So there's likely another unannounced card higher up in the price stack (as well as 460 below it).

Perhaps they haven't finalized the clocks on it yet. Or perhaps the yields are too bad and they can't launch the X version. Or they are waiting to see what 1060 is like? Perhaps they are waiting as long as possible in hopes that more Furys will sell before they announce something that would be uncomfortably close to it?

Regards,
SB

 

[CarstenS]

If the GPUs in Rajas AotS demo were only 50% saturated (what? their AC engines? Their shader cores? Their ROPs?), I wonder how power consumption compared to the single 1080 - given that every effort in Polaris has been geared towards efficiency.

 

[Deleted member 13524]

This WILL be my first dGPU since a long time. My last dGPU is GeForce 400MX.....
Hopefully the 8GB variant is closer to 200 than 300 ($250 would be fine).

When Raja was mocking nvidia's $700 price point, he compared it to a <$500 Crossfire solution from AMD. If you think that would be a 2*480X with 8GB each (the $199 4GB version, would be called a <$400 solution), then the 8GB version is definitely $250 or less (probably $249, actually).

 

[SimBy]

Well easy way to see this what is the utilization rate of current AMD 2 card setups in AOTS?

Considering AotS mGPU scales poorly I would say utilization is very low.

http://www.anandtech.com/show/10067/ashes-of-the-singularity-revisited-beta/4