AMD: Speculation, Rumors, and Discussion (Archive)

[Ext3h]

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

Correct me if I'm wrong, but it would still count as an architectural improvement if the culling was improved by adding additional microcode to the last pre-FS pipeline stage, wouldn't it?

Pre-cull and compact on the CU, prior to stressing the fixed-function units in the rasterizer?

 

[CarstenS]

I wonder what good this will do. Especially in terms of energy efficiency and latency. Pure and simple triangle setup (and culling) has not been a major concern for Radeon cards in a long time IIRC.

 

[xpea]

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...

Or simply they had to boost the clock on Polaris to be more competitive with Pascal, and they ended with worst tdp than the initial target...

All in one, except the welcome aggressive price, I'm not impressed at all by Polaris. Of course, more details on the uarch are needed for a final judgment but the overall picture is not rosy. I have the feeling that GP106 won't have any difficulties to spoil Polaris and dominate the segment like gtx950/960 did (lower tdp, better perf/w, and the superior nvidia brand that will be a huge factor in this mainstream category, not even talking about gtx1080 halo effect)

 

[3dilettante]

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.

Not something like collecting batches of primitives and only submitting after they fill up with non-discarded geometry?
http://www.google.com/patents/US20140292756


On the topic of Polaris currently, the board's power budget does take it into the realm of Hawaii's power density.
At this point, it's likely a mm2 game with Nvidia's silicon, and there's likely more fixed overhead with ancillary silicon and the memory interface's area and power that Polaris has to pay into before it can get to execution resources. The power savings seem to resemble Nvidia's not getting 2x per transistor either due to clock compensating for a die diet.

 

[xEx]

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.

I don't know if i'm wrong but doesn't PCIE 2.0 is able to deliver 150W and PCIE3.0 175W?

 

[Razor1]

no, the pci-e bus delivers 75 watts.

More stuff on the rx4xx line

http://videocardz.com/60824/amd-polaris-10-gpu-in-radeon-rx-400-series

 

[Kaotik]

I don't know if i'm wrong but doesn't PCIE 2.0 is able to deliver 150W and PCIE3.0 175W?

No, PCIe 1.x, 2.x and 3.x slots can all provide only 75W

 

[Deleted member 13524]

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

To be honest, there are no two equal cards being shown in that DX12 explicit multi-gpu test, so the poor scaling could be the fault of a load distributor in its infancy.

Moreover, those tests are in 4K, which means a crapload of draw calls, so maybe it's bottlenecked by the CPU? (scary that it would be bottlenecked by a 6-core Ivybridge at 4GHz, though).

 

[Razor1]

they are getting around a 50% boost for SLi

Here is Xfire and Sli

 

[Kaotik]

To be honest, there are no two equal cards being shown in that DX12 explicit multi-gpu test, so the poor scaling could be the fault of a load distributor in its infancy.

Actually we do have one such bench: http://www.techspot.com/article/1137-directx-12-multi-gpu-geforce-radeon/page2.html
They have 390X + 390X performing better than 980 + 390X, but worse than 980 Ti + Fury X

 

[jacozz]

More stuff on the rx4xx line
http://videocardz.com/60824/amd-polaris-10-gpu-in-radeon-rx-400-series

Ok now I get it.
The "big brother" based of Polaris 10 is going to be named "RX490"
The cut down version is "RX480"
The cut down of the cut down version is "RX470"

 

[SimBy]

they are getting around a 50% boost for SLi

Here is Xfire and Sli

So around 35% for CF. I'm surprised SLI actually scales better in AotS.

 

[CarstenS]

Moreover, those tests are in 4K, which means a crapload of draw calls,[...]

Are those per pixel?

 

[Love_In_Rio]

Well RX 480 at 150 watts vs Nvidia 1070 GTX at 150 watts with the former with a performance similar to a 980 and the later to a 980TI tells us than Polaris architecture with respect to Pascal has almost not moved a pinch from where it stood at 28 nm with respect to Nvidia Maxwell architecture efficiency.Of course they have to play the price card, at least until 1060 is out.

 

[Infinisearch]

Moreover, those tests are in 4K, which means a crapload of draw calls, so maybe it's bottlenecked by the CPU? (scary that it would be bottlenecked by a 6-core Ivybridge at 4GHz, though).

Yeah what does resolution have to do with draw call count?

 

[Deleted member 13524]

Yeah what does resolution have to do with draw call count?

I'm assuming higher resolution = larger amount of "zoom out" that the game allows = more units being rendered. It happens on some RTS IIRC, but I could be wrong for this specific game.

 

[Orion]

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

Hmmm, 31% performance increase versus single fury? I'd thought it was 66% increase, from what I skimmed elsewhere.

If similar holds for 480 that means that 31% performance increase would beat a 1080 by 6%. Or that a 1080 would be 25% faster than a single 480 in amds ashes bench. A 1080 is also about 25% faster than a 1070 in some benches.

 

[Infinisearch]

I'm assuming higher resolution = larger amount of "zoom out" that the game allows = more units being rendered. It happens on some RTS IIRC, but I could be wrong for this specific game.

Isn't the aspect ratio locked so the same things are in view?

 

[Deleted member 13524]

Well RX 480 at 150 watts vs Nvidia 1070 GTX at 150 watts with the former with a performance similar to a 980 and the later to a 980TI tells us than Polaris architecture with respect to Pascal has almost not moved a pinch from where it stood at 28 nm with respect to Nvidia Maxwell architecture efficiency.Of course they have to play the price card, at least until 1060 is out.

So you think pricing has to do with power consumption alone? The fact that the Polaris 10 chip is 30% smaller, may use cheaper memory and cheaper PCB components are nothing compared to power consumption and that's why they have to play the price card?

Besides, you'll only see actual power consumption comparisons in side-to-side reviews, not with the TDP numbers whose actual meaning seem to vary greatly between IHVs and even between different generations of the same IHV.

 

[el etro]

I think that GCN cards are rated for high TDPs because under compute loads they fully stress the GPU, making efficiency being different(but good), and makimg power consumption going up. So a RX 480 may consume ~110W on gaming loads and consume around the TDP on GPGPU loads.


Nvidia instead rates their cards with average power consumption, once they feel secure that in no plausible condition the cards will have thermal/power problems. This TDP rating become a bit less agressive in Pascal, versus Maxwell.