AMD: Speculation, Rumors, and Discussion (Archive)

[SimBy]

That's 11% higher than predicted based on AotS CF benchmark which put it exactly at 390 level. Could be an OC model.

 

[AlNom]

It's (c)hip to be square.

too many sixes with a 384bit bus
384bit x 6666mhz = 320GB/s

Or rather, 0.30ns.

 

[trinibwoy]

So according to those charts the 1070 is 90% more $$$ for 44% more performance. And that's using the mystical $379 price.

If the 480 hits the right level of absolute performance for 1080p it can inhale a lot of the market.

 

[Gipsel]

A 1.83 scaling in a game is good.
But the context is comparing the 480 to the 390x where we can use the same number as a theoretical comparable (I really doubt the 390x scales that well) to see really how well the 480 does compare to previous cards.

I would take that 1.83 scaling number with quite some salt. Robert Hallock clearly messed up some numbers there and I had the feeling he didn't know exactly what he was talking about. At least he used the utilization and scaling as synonymes (already back then, when it was at 51% for the light batches or even a single one). Assuming a 83% GPU load is a real number in this AotS run, it is probably reported as average over the two cards. That would put an upper limit of 1.66 to the scaling. That would also be better in line from the multi GPU scaling in the tests I'm aware of (usually <=65% performance gain from a second card). That would bridge already more than half of the performance gap to the 390X. And of course one could make the argument that AotS may be not the right benchmark to demonstrate the architectural improvements in Polaris.
Anyway, if the clock of 1.26GHz holds some water, the arithmetic/texture performance should be basically a wash with the 390X, if the 480X can hold this speed (better cooling of custom cards needed?), even if we neglect any improvements with GCN v4. 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)?

 

[Grall]

Are you using a custom AIB 390x or reference?

Fully custom PCB/cooler, ASUS DCII, it's standard two-slot wide, and it does get pretty loud when the fans are turning at speed.

 

[CSI PC]

Fully custom PCB/cooler, ASUS DCII, it's standard two-slot wide, and it does get pretty loud when the fans are turning at speed.

Ah ok, looks like the 2.5 wide designs were needed then.
Shame as I thought overall the 390x generally overcame the heat/noise issue of the 290x, especially looking at reviews but those were probably 2.5/3 wide slot.
Cheers

 

[CSI PC]

I would take that 1.83 scaling number with quite some salt. Robert Hallock clearly messed up some numbers there and I had the feeling he didn't know exactly what he was talking about. At least he used the utilization and scaling as synonymes (already back then, when it was at 51% for the light batches or even a single one). Assuming a 83% GPU load is a real number in this AotS run, it is probably reported as average over the two cards. That would put an upper limit of 1.66 to the scaling. That would also be better in line from the multi GPU scaling in the tests I'm aware of (usually <=65% performance gain from a second card). That would bridge already more than half of the performance gap to the 390X. And of course one could make the argument that AotS may be not the right benchmark to demonstrate the architectural improvements in Polaris.
Anyway, if the clock of 1.26GHz holds some water, the arithmetic/texture performance should be basically a wash with the 390X, if the 480X can hold this speed (better cooling of custom cards needed?), even if we neglect any improvements with GCN v4. 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)?

Robert clarified this as an edit, so this is further information he received and put right, so I would say it is more robust than just generally being able to dismiss it as wrong.

//EDIT: To clarify this, the scaling from 1->2 GPUs in the dual RX 480 test we assembled is 1.83x. The OP was looking only at the lowest draw call rates when asking about the 51%. The single batch GPU utilization is 51% (CPU-bound), medium is 71.9% utilization (less CPU-bound) and heavy batch utilization is 92.3% (not CPU-bound). All together for the entire test, there is 1.83X the performance of a single GPU in what users saw on YouTube.
The mGPU subsystem of AOTS is very robust.

If Robert and co at AMD cannot get this right with a clarification update, then this raises questions about any AMD benchmark tests-setups in general and the worth/validity of presenting their cards with those results against the competition..
Thanks

Edit:
I would like to think that if Robert Hallock was wrong , then he would had edited that post a 2nd time with an update to the clarification I quoted.
This is in an official AMD reddit section.

 

[Gipsel]

Robert clarified this as an edit, so this is further information he received and put right, so I would say it is more robust than just generally being able to dismiss it as wrong.


If Robert and co at AMD cannot get this right with a clarification update, then this raises questions about any AMD benchmark tests-setups in general and the worth/validity of presenting their cards with those results against the competition..
Thanks

Edit:
I would like to think that if Robert Hallock was wrong , then he would had edited that post a 2nd time with an update to the clarification I quoted.
This is in an official AMD reddit section.

I read the whole thing at reddit. As I said, he uses the GPU usage and scaling as synonymes in several places there. That really makes me wonder, what truth it holds (does AotS even report utilization numbers as quoted? since when are there numbers for a single batch [and not light ones]?) and if he understands exactly what he writes. I don't care if it is an official thing or not. It's fishy in any case.

 

[CSI PC]

So according to those charts the 1070 is 90% more $$$ for 44% more performance. And that's using the mystical $379 price.

If the 480 hits the right level of absolute performance for 1080p it can inhale a lot of the market.

Well you need to use the 8GB 480 model as a comparison.
Assuming that is $239 it would make the mythical/mystical/drumming fingers while waiting for a cheap 1070FE around 58% more expensive.
The 8GB model could come a little under that figure or a bit more (doubt it myself).
Cheers

 

[CSI PC]

I read the whole thing at reddit. As I said, he uses the GPU usage and scaling as synonymes in several places there. That really makes me wonder, what truth it holds (does AotS even report utilization numbers as quoted? since when are there numbers for a single batch [and not light ones]?) and if he understands exactly what he writes. I don't care if it is an official thing or not. It's fishy in any case.

I appreciate where you are coming from but it is easy for them to take the single GPU figure, compare to the mGPU and have the gain, and I read it like that.
The rest of the info on utilisation/etc I take with a pinch of salt, but the 1.83 figure is not hard for them to work out.
TBH I see the the 1.83 gain being fine in mGPU as this could be achievable with ideal support and game design.
Could still be pie in the sky as you say, and that would mean we have zero figures apart from the recent released 3DMarks benchmark at videocardz with their source allegedly being an AMD partner (so possibly slight OC): http://videocardz.com/61005/new-amd-radeon-rx-480-3dmark-benchmarks
But here we would expect the 480 to be stronger than previous GCN due to reported improvements in the architecture.
The 390 should read 89% and not 79%
Cheers

 

[Deleted member 13524]

I'm thinking the $199 4GB model is more of a placeholder for AMD's intended direction for the RX480 and the 8GB will be quite a bit more expensive.

Remember Polaris is supposed to go up to $300.

 

[Kaotik]

I'm thinking the $199 4GB model is more of a placeholder for AMD's intended direction for the RX480 and the 8GB will be quite a bit more expensive.

Remember Polaris is supposed to go up to $300.

They can't in any possible scenario justify over $50 premium for the 8GB version, and even $50 is stretching it

 

[Deleted member 13524]

They can't in any possible scenario justify over $50 premium for the 8GB version, and even $50 is stretching it

Well they did say the Crossfired RX 480 setup that was beating the GTX 1080 in AotS would cost less than $500. Remains to be seen if those results were achieved with the 4GB or 8GB versions (can't see why not, since DX12 explicit multi-adapter doesn't require the cards to have the exact same VRAM content). I mean 2*RX 480 4GB ($398) is definitely less than $500.

I actually asked that question (4GB or 8GB RX480 in the AotS comparison) to Robert Hallock in a reddit AMA he did after the computex announcement. He didn't answer.

 

[Razor1]

yeah 250 would be the max for an 8 gb version, AIB's will charge more for out of the box overclocked versions, that will probably be up to $300.

Well less that 500 could mean 498 lol, 249 per card?

 

[Anarchist4000]

6 TF Polaris likely wouldn't fit into a console power envelope (for reference, XBO uses 110-120 watts for the entire machine when gaming). It's coming out at the end of 2017. It's more likely to be based on Vega.

For reference, I'm expecting PS4 Neo to fit within a similar power envelope as PS4 (~135-145 watts total when gaming). And that's using a 4.4 TF GPU likely based on Polaris.

Regards,
SB

It seems likely that traditional envelope may be changing. Larger power supply just to run up to 100W over thunderbolt for a future VR headset would make a lot of sense. As for the XBO they mentioned two new models with a high and traditional performance and price.

But that statement seems wrong since you are suggesting that Vega will provide slightly better performance than Polaris but much less power consumption? I think is more plausible to have a customize version of polaris, maybe 40CUs with a excavator mk2(?) cores.

thinking about it would it be viable to make a polaris chip with HBM for ultimate perf/watt? I think that would consume much less than a vega chip but i really have no idea if it is viable to do such a huge change.

At <150W 8GB of GDDR5 would be a full quarter or more of total power. Cutting that 30W in half would be a significant reduction.

Supposedly all the cards out there have varying default clocks at the moment, which will be corrected by final BIOS or driver.

They might not have provided drivers for the final product to enable everything either. Really easy way to prevent leaks that can't easily be worked around. In theory some, if not all, of their tweaks would require significant compiler changes. The bigger issue for partners is just the cooler and TDP, implementing features that likely lower power usage shouldn't be a problem.

I'm thinking the $199 4GB model is more of a placeholder for AMD's intended direction for the RX480 and the 8GB will be quite a bit more expensive.

Remember Polaris is supposed to go up to $300.

Fairly sure AMD said somewhere it was 199/229 for each model with the series falling in to 100-300 range. The bottom of that is easily the 460/470 parts. The $300 is either a liquid cooled midrange, which some partners demoed, or high OC parts. The problem was the liquid cooling part is that it's rather unjustified for a midrange product.

I'm still a bit surprised we haven't seen anything with HBM1. Maybe HBM2 is coming early, but it opens up some technical possibilities I'd think would serve a midrange market well. Lower power consumption being one of them, which is huge for mobile markets. Pairing up ships being another possibility. 4GB isn't necessarily a limit for those markets either.

EDIT: Guess another possibility would be a GDDR5X part. That would also require a 480X actually being constrained by bandwidth.

 

[MDolenc]

can't see why not, since DX12 explicit multi-adapter doesn't require the cards to have the exact same VRAM content

Careful. Just because it's possible to implement an explicit multi-adapter scheme that does not require a per adapter copy of pretty much all the resources doesn't mean such a scheme is implemented in AotS. In fact AotS still implements just classic AFR which does require doubling resources. It's not like textures and geometry used in even frames won't be used in odd frames...
So no memory doesn't just magically double.

 

[Deleted member 13524]

In fact AotS still implements just classic AFR which does require doubling resources.

Are you 100% sure of this?

 

[Jawed]

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.

Is there some excuse such as "compute heavy" game graphics is heavier on bandwidth, therefore cards need more bandwidth for future games?

 

[Kaotik]

I'm still a bit surprised we haven't seen anything with HBM1. Maybe HBM2 is coming early, but it opens up some technical possibilities I'd think would serve a midrange market well. Lower power consumption being one of them, which is huge for mobile markets. Pairing up ships being another possibility. 4GB isn't necessarily a limit for those markets either.

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.

 

[Kaotik]

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.

Is there some excuse such as "compute heavy" game graphics is heavier on bandwidth, therefore cards need more bandwidth for future games?

Did I miss the moment in time where bandwidth became the main player on video card performance?
Polaris 10 is around 100mm^2 smaller chip (that's around 2/3rds of the size for chips this size) with less theoretical compute power than GTX 1070 (assuming the 1266 MHz clocks are the final clocks), given that NVIDIA has been always(?) getting more performance in games per theoretical TFLOP, did you expect something else?