AMD: Speculation, Rumors, and Discussion (Archive)

[spworley]

If AMD concluded that a very temporary, very minor margin hit on a few 4GB parts was an acceptable trade-off for meeting their $200 marketing claims, then why not?

Especially if the channel is supplied with mostly 8GB and not 4GB SKUs. This provides the $199 anchor price to help perception and demand, but forcing an upsell to the more profitable $239 8GB in practice. The deliberate dichotomy already worked for launch reviews, which mostly benchmarked the 8GB variant.

NVidia did the same price anchor switch with FE models. $379 retail for GTX 1070. But in practice only the FE edition is available, for $449 retail. Yes, list price OEM models are "coming soon". Yep. Real soon. As soon as the cream is fully skimmed, and not before.

 

[lanek]

Especially if the channel is supplied with mostly 8GB and not 4GB SKUs. This provides the $199 anchor price to help perception and demand, but forcing an upsell to the more profitable $239 8GB in practice. The deliberate dichotomy already worked for launch reviews, which mostly benchmarked the 8GB variant.

NVidia did the same price anchor switch with FE models. $379 retail for GTX 1070. But in practice only the FE edition is available, for $449 retail. Yes, list price OEM models are "coming soon". Yep. Real soon. As soon as the cream is fully skimmed, and not before.

Its not the first time that at launch, first samples are the same with different bioses, it was too the same for the small variant even with disabled SP in bios.. its just a question of channel availability and production lines. ( 7950 flashable to 7970 bios, 290 flashable to 290x, 5850 flashable to 5870 )

I absolutelly not see what it have to do with the Nvidia FE / MSRP prices ... In fact, it is completely the invert situation.... Whatever is the avaibility of the 4Gb version yet, it is there. And you willl find them on shelf at a moment or another, something who will never happend with the 1080-1070 "MSRP" one. ( outside maybe for fix the OEM price ( before appllied the deduction )

Happy "future" buyer of the 4GB version if it is indeed a 8GB one they have pay at a lower price.

Ha, and the report is confirmed by 4GB version buyers, not by "reviewers" ...

http://www.guru3d.com/news-story/ea...80-4gb-cards-actually-do-have-8gb-memory.html


I can imagine that the "cheaper" version have been send on some chanel country circuit when 8Gb version availability have been send more widely in US, EU ..

 

[CarstenS]

As already said, one can program the VRM Controller to bias the balancing of the phases. You can instruct the VRM Controller to draw more from specific phases than from other ones. And that is not a guess, it's an example given in the data sheet for the 4phase version of the controller (didn't find the complete datasheet for the 6phase version in public sources, but it supports the same features according to the first 2 pages which are accessible):

And there also may be the option to simply switch off phase 5 and 6 at normal loads, if they are connected to the slot and not the 6pin plug (it's somewhat of a 50:50 chance right now if they didn't put some thought in it).

Even it would still be massively OP and in need for a nerf, cost-down wise, you could cycle through 3 extra phases (one for the memory, two for the GPU) in order to keep temperatures down and power consumption lower. VRMs running very hot, especially when aiming to build a quiet blower-style card, would not be very beneficial, I think. Adding to that: You would blow the hot air from the GPU (temperature target 80+ °C) right over the VRMs.

 

[Razor1]

more than 2 year in the making, still come out like this. I can't imagine any reason for this than how low the resources was to the engineering team. I think it would be much better to AMD to focus in few product, few well made products at time. Like a 480 alone and then few months after a 460 and then focus everything in Vega and so on. Its better to have few well made product than a whole line up of almost well made products.(Im talking from the perspective that AMD have really low resources so it would be better to focus it on few product to maximize the engineering)

Depends on how they set up their teams, I'm sure they had two design teams, one for Polaris and one for Vega, so we can't really preclude to think they don't have enough resources to make a compelling product at least not yet. Polaris by its performance/mm isn't a bad product, it seems it came up short on power utilization though......

Vega can be totally different in this regard, just have to wait and see.

 

[Grall]

No you can't pre order in newegg or other stores at the moment(just cheked)

I'm not speaking exclusively of this one particular product. Online tech does allow you to place orders of merchandise which is not in stock; if Newegg doesn't work like that then that's on them, because they choose to work that way. Or there's something bigger afoot here, a recall, or AMD placing a hold sale order or whatever.

 

[xEx]

I'm not speaking exclusively of this one particular product. Online tech does allow you to place orders of merchandise which is not in stock; if Newegg doesn't work like that then that's on them, because they choose to work that way. Or there's something bigger afoot here, a recall, or AMD placing a hold sale order or whatever.

Which is exactly my guess

 

[Deleted member 13524]

Of course, that means pushing more power draw onto the 6-pin connector, so that would bring that one (further, possibly) out of spec. So not exactly a superduper solution on the whole. Best would be to magically grow an 8-pin connector instead for 150W headroom on the molex connector, but that one is not doable in a software patch for sure!

Let's be honest though, the 8pin-for-150W requirement is just aesthetic in most practical cases.
I have yet to encounter a PSU whose 8-pin connectors don't come from 6 wires and the additional 2 grounds are literally shunted from the connector itself.
I wonder how many non-terrible PSUs (rated above say 350W) would underperform from pulling sustained 150W or even 200W from a 6-pin connector, let alone the sub-100W we're discussing for the RX 480.



As for the 4 + 2 phases, is there any real reason why the PCIe's 2 phases can't handle most of the power? If the VRMs are as high-end as reports say, then why not?
Maybe the 4 phases on the PEG is there to test the waters for MXM-B cards where the connector can withdraw up to 200W.
AMD is low on resources so it might be advantageous for them to perform R&D testing on production hardware, even if it means losing some money on the components themselves. OEMs can then distribute their phases the way they like through their own custom versions.

 

[lanek]

Bullzoid measured the 6 Pin to have 3 grounds instead of 2 + 1 sensing. It allows the same current as the 8 pin.

theres no 6pin or 8pin on actual GPU, they are all 8pin. it was the same with 8+6pin gpu's ( its not an hazard if you see the first "custom" RX480 with 8pin ).. The PCB is exactly the same, there's no custom PCB who wil be made before a while. So the 8pin tracing is just " physically reduced " to 6pin but the tracing are the same.

there's no cost difference to make a 8 or 6 pin PCI express. so in general they just do an 8pin version.

 

[CSI PC]

Even it would still be massively OP and in need for a nerf, cost-down wise, you could cycle through 3 extra phases (one for the memory, two for the GPU) in order to keep temperatures down and power consumption lower. VRMs running very hot, especially when aiming to build a quiet blower-style card, would not be very beneficial, I think. Adding to that: You would blow the hot air from the GPU (temperature target 80+ °C) right over the VRMs.

Yeah that is my concern as well, ironically it is overspec in most ways but unfortunately the cooling system is too weak and crude.
It would result back to the situation of the boost algorithms reducing voltage-power target.
But it may be part of an overall solution if at all feasible, because I am pretty sure they do not want performance to drop noticeably as that will be a real PR nightmare with news reporting of review samples being faster than retail even though I do not think this is deliberate, it will be a perception thing that sticks.

Just to add.
IMO quite a lot of technical variables for AMD to handle with the already released 480 cards and not all of them are going to be resolved to a satisfactory level for everyone or even possible specs on all connections.
Cheers

 

[el etro]

I'm thinking they will solve the power consumption problem via driver, both changing the power delivery scheme to push more power via PCI-e conector, but also being more aggresive with the voltage tunning with the boards, maintining the performance level. Is a very robust tune to do on driver, much bigger than what they did with R290x issue.

I think they could impress more the reviewers doing more agressive voltage tunning on reference cards, is not the first GCN gen. where i see really great undervolting results with Radeons.

 

[Anarchist4000]

VRMs running very hot, especially when aiming to build a quiet blower-style card, would not be very beneficial, I think. Adding to that: You would blow the hot air from the GPU (temperature target 80+ °C) right over the VRMs.

Engineering is actually the opposite of that. Hotter components, not more heat production, are easier to cool. If the card were designed to operate at 100C it might passively cool itself. Same logic applies to low RPM, quiet fans. The VRMs they used are so over designed I doubt heat will ever be an issue for them.

I'm not speaking exclusively of this one particular product. Online tech does allow you to place orders of merchandise which is not in stock; if Newegg doesn't work like that then that's on them, because they choose to work that way. Or there's something bigger afoot here, a recall, or AMD placing a hold sale order or whatever.

Fairly sure when I checked Newegg the other day they listed backorder dates for the 2nd and eventually 7th. Not receiving deliveries daily or on Sunday or a national holiday doesn't seem unreasonable. Vendors appear to still be receiving inventory, they just clear them quickly. Best Buy is sold out of the cards for example but still selling built systems with 480s in them. If there were any sort of recall I'd expect that to change.

As for the 4 + 2 phases, is there any real reason why the PCIe's 2 phases can't handle most of the power? If the VRMs are as high-end as reports say, then why not?
Maybe the 4 phases on the PEG is there to test the waters for MXM-B cards where the connector can withdraw up to 200W.
AMD is low on resources so it might be advantageous for them to perform R&D testing on production hardware, even if it means losing some money on the components themselves. OEMs can then distribute their phases the way they like through their own custom versions.

Each VRM can probably handle 150W at >70C. So long as a 480 doesn't pull down North of 300W two of those VRMs could power the entire board if everything were 12V. Not sure why some people are so hysterical about this allegedly unsolvable problem.

Yeah that is my concern as well, ironically it is overspec in most ways but unfortunately the cooling system is too weak and crude.
It would result back to the situation of the boost algorithms reducing voltage-power target.

The cooling system could be more than ample. It just depends on the operating temperature. Cooling is always based on delta from ambient. Hotter components are easier to cool from a heat dissipation perspective. Stick your PC in a freezer and your cooling system won't have to work nearly as hard.

 

[pharma]

AMD denies more than 36 shader clusters at Polaris 10
http://www.3dcenter.org/news/amd-dementiert-mehr-als-36-shader-cluster-bei-polaris-10

 

[lanek]

Engineering is actually the opposite of that. Hotter components, not more heat production, are easier to cool. If the card were designed to operate at 100C it might passively cool itself. Same logic applies to low RPM, quiet fans. The VRMs they used are so over designed I doubt heat will ever be an issue for them.




Each VRM can probably handle 150W at >70C. So long as a 480 doesn't pull down North of 300W two of those VRMs could power the entire board if everything were 12V. Not sure why some people are so hysterical about this allegedly unsolvable problem.


The cooling system could be more than ample. It just depends on the operating temperature. Cooling is always based on delta from ambient. Hotter components are easier to cool from a heat dissipation perspective. Stick your PC in a freezer and your cooling system won't have to work nearly as hard.

- Effectively AMD have allways over engineering the VRM and power delivery parts, hence why for overclocking with "non" standard cooling systtem,, it was offtly better to just take the reference gpu's instead of the custom one ( outisde some specific sauces as the MSI custom pcbs but who was available way later ). its not much a desire to use "high end component quality, but it reduce the cost of developement and testing by dont have 3-4 different level of quality ( hence why even middle end and low end have offtly "over than needed" power system delivery. Looking at the specs of it, you could find the same components on a 600$ + gpu's.

I have allways put my gpu's under H20 coolling ( i dont remember the last time i have make run a new gpu's more than 5 minutes under his aircoooler ), I allways buy reference ones for a simple reason, i couldnt care less about the fancy partners air cooling block, and i want the reference PCB+VRM, power delivery ), so i have no bad surprise then with bad Mossfets, Caps, or VRM quality. ( With AMD gpu's ofc )

- I can ensure you that thoses VRM could take way more than 300W without an hi-cups ...

 

[CSI PC]

As for the 4 + 2 phases, is there any real reason why the PCIe's 2 phases can't handle most of the power? If the VRMs are as high-end as reports say, then why not?
Maybe the 4 phases on the PEG is there to test the waters for MXM-B cards where the connector can withdraw up to 200W.
AMD is low on resources so it might be advantageous for them to perform R&D testing on production hardware, even if it means losing some money on the components themselves. OEMs can then distribute their phases the way they like through their own custom versions.

Less phases in basic terms means more ripple, more heat (high-side state longer due to duty cycle requirements), worst transient response, potentially more noise, potentially greater loss, larger capacitors and better rating inductors.

I would need to dig it out, but the big steps in terms of optimal multiphase improvements is 2 going to 4, and 4 going to 6 phases, these of course being true phase and not doubled.
Cheers

 

[Anarchist4000]

AMD denies more than 36 shader clusters at Polaris 10
http://www.3dcenter.org/news/amd-dementiert-mehr-als-36-shader-cluster-bei-polaris-10

This entire topic has been missed with all the power discussion. Shader clusters aside, I haven't seen much in the way of what is inside a cluster. Historically AMD hasn't counted scalars, but they could be playing a larger role with Polaris and there could be more of them.

- Effectively AMD have allways over engineering the VRM and power delivery parts, hence why for overclocking with "non" standard cooling systtem,, it was offtly better to just take the reference gpu's instead of the custom one ( outisde some specific sauces as the MSI custom pcbs but who was available way later ). its not much a desire to use "high end component quality, but it reduce the cost of developement and testing by dont have 3-4 different level of quality ( hence why even middle end and low end have offtly "over than needed" power system delivery. Looking at the specs of it, you could find the same components on a 600$ + gpu's.

- I can ensure you that thoses VRM could take way more than 300W without an hi-cups ...

That 150W was downrated per VRM at high operating temperatures. Spec sheet only specified 70C and I ballparked 100C or more. Perfectly balanced they could probably push over 900W with that thing. So I don't doubt they could power whatever they wanted.

 

[CarstenS]

Engineering is actually the opposite of that. Hotter components, not more heat production, are easier to cool. If the card were designed to operate at 100C it might passively cool itself. Same logic applies to low RPM, quiet fans. The VRMs they used are so over designed I doubt heat will ever be an issue for them.

I think we're talking two different things here.

 

[Clukos]

I'm curious to see if people at oc.net manage to unlock a 4gig card into an 8 gig with a bios update. Since most people with 4gig 480s are reporting the same amount of memory modules on the board.

 

[silent_guy]

I'm curious to see if people at oc.net manage to unlock a 4gig card into an 8 gig with a bios update. Since most people with 4gig 480s are reporting the same amount of memory modules on the board.

4GB and 8GB will always have the same amount of DRAM chips. The question is whether or not the chips have the same capacity.

 

[lanek]

I'm curious to see if people at oc.net manage to unlock a 4gig card into an 8 gig with a bios update. Since most people with 4gig 480s are reporting the same amount of memory modules on the board.

As said upper, the numbers of modulles are offtly now the same with GDDR5.. just the capacity of each modues is different ( tthis is like that for years... different on the past, when, modules ( chips ) was even on the up and down of the gpu's ) .. 4x2GB modules or 4x1GB modules .

 

[pharma]

A user called bertieX over at Dutch website Tweakers.net reported that he purchased a 4GB PowerColor RX 480. The card has 8 memory chips and when he watched the GDDR5 ICs closely, he noticed Samsung K4G80325FBHC25 chips, these are confirmed 8Gbit. With eight of them that does make a total of 64 Gbit and thus indeed there is 8 GB of GDDR5 memory present on the PCB.

The card however is recognized and utilizing 4GB as model, this means that the remaining 4 GB is disabled in the BIOS.

http://www.guru3d.com/news-story/ea...80-4gb-cards-actually-do-have-8gb-memory.html