[WCC] AMD Rolling Out New Polaris GPU Revisions With 50% Better Perf/Watt

[Det0x]

Link to rumors can be found here:
http://wccftech.com/amd-polaris-revisions-performance-per-watt/

AMD has reportedly started rolling out new revisions of its Polaris 10 & Polaris 11 GPUs with a 50%+ improvement in performance per watt. Polaris 10 and Polaris 11 are AMD’s latest GPUs powering the company’s midrange RX 480 & RX 470 as well as mainstream RX 460 graphics cards. The new GPU revisions will reportedly debut first in the embedded market and will deliver the same and/or slightly better performance compared to the previous revisions but at significantly lower power consumption.

AMD’s New Polaris GPU Revisions Are 50%+ More Power Efficient

This significant improvement in power efficiency has been attributed to an improvement to the 14nm metal mask layers of both Polaris GPUs by AMD in combination with a more refined binning process. The new revision of Polaris 10 that’s will go into the embedded market will bring down the typical board power from 150W as is found in the RX 480 to less than 95W. The updated revision of Polaris 11 will not only bring down the power from 75W to less than 50W but it will also improve clock speeds and bring up the compute throughput to 2.5 TFLOPS from 2.15.

The new revisions will be available immediately to the embedded market under the E9550 and E9260 product names for Polaris 10 and 11 respectively. The updated revisions are believed to be coming to the mobile market in the form of mobility Radeon RX 400 series graphics cards in the coming months. It’s not clear yet whether AMD will also introduce the new revision to the desktop market. However, in the mobile market where power efficiency rules supreme this could prove to be a pivotal change for the company’s competitiveness in notebooks.

AMD Radeon RX 485, RX 475 & RX 465 Graphics Cards Possibly In The Works
Some industry sources are speculating that AMD could introduce the new Polaris 10 and Polaris 11 revisions as part of the company’s graphics refresh next year. To be part of the new product stack alongside the brand new powerful enthusiast-class Vega 10 and Vega 11 GPUs in what we assume we’ll be called the RX 500 series. There’s evidence to believe that this is plausible. The Radeon Technologies Group had revealed earlier this year that the new naming scheme introduced with the RX 400 series had been designed with room for updated revisions in mind, e.g. RX 485, RX 475 and RX 465 graphics cards.

However, as the desktop isn’t a power critical platform the company is more likely to choose to leverage the higher power efficiency to boost performance by raising clock speeds and maintaining similar TDP levels of current RX 400 series graphics cards. AMD’s new flagship graphics chip code named “Vega 10” is scheduled to debut at the end of the year. Another high-end “Vega 11” GPU is scheduled to launch next year to slot it in between Vega 10 and Polaris 10.

 

[Det0x]

If we take this rumor at face value, then suddenly this review by "JayzTwoCents" makes alittle more sense..

This RX 480 DOESN'T SUCK! @ https://www.youtube.com/watch?v=zWASNajSdpg

XFX RX 480 GTR on air: 1288 MHz@ 1.05v around 85-90W in Heaven benchmark, 1475 MHz@1.185v - 133W.
Water Cooled RX 480: 1266 MHz@1.05v around 105W in heaven benchmark, 1470 MHz@1.185v - 150W, and unstable/crashes.

Gigabyte G1 from the same reviewer:

125W at stock clocks. And 180-190W when OC'ed.

Something is just not right here, guys...

I do think that XFX RX 480 must use at least new revision of the process, or be made on Samsung process. Then it is logical.

 

[RussianSensation]

If we take this rumor at face value, then suddenly this review by "JayzTwoCents" makes alittle more sense..

This RX 480 DOESN'T SUCK! @ https://www.youtube.com/watch?v=zWASNajSdpg

If I were running AMD, I'd release a 15-20% faster clocked RX485 refresh (or call it RX580, etc.) on the desktop at 140-150W TDP, and use the more efficient revisions for mobile dGPU OEM design wins. I feel that having a card that outperforms GTX1060 3-6GB by 7-10% is a lot more important on the desktop than having an RX 480 card with 95-100W TDP. If they push RX 470 down to $159-169 against upcoming $149 1050Ti, position RX 480 8GB at $199-209 against GTX1060 3GB and bring out a 15-20% faster RX485 at $239-249 against GTX1060 6GB, that would make all 3 of those AMD cards superior to NV's offerings, forcing NV to drop prices and/or do a refresh of Pascal.

The rumoured specs for Vega 10 230W have 12Tflops against 7Tflops RX 580 with 130W TDP. That's 230W/130W = 77% increase in power usage for a 71% increase in Tflops. What concerns me is that there isn't a GTX1080Ti (or 2080Ti GP102) competitor on the road-map yet. It seems Vega 10 was designed to take on GTX1070/1080 and it seems a 275-300W Vega chip is missing for 2017.

 

[Det0x]

If I were running AMD, I'd release a 15-20% faster clocked RX485 refresh (or call it RX580, etc.) on the desktop at 140-150W TDP, and use the more efficient revisions to for mobile dGPU OEM design wins. I feel that having a card that outperforms GTX1060 3-6GB by 7-10% is a lot more important on the desktop than having an RX 480 card with 95-100W TDP. If they push RX 470 down to $159-169 against upcoming $149 1050Ti, position RX 480 8GB at $199-209 against GTX1060 3GB and bring out a 15-20% faster RX485 at $239-249 against GTX1060 6GB, that would make all 3 of those AMD cards superior to NV's offerings, forcing NV to drop prices and/or do a refresh of Pascal.

Could already be underway..

http://wccftech.com/amd-radeon-rx-470-price-cut-for-gtx-1050-ti/

While AMD set the MSRP to $179 US (RX 470), models listed on retail sites are on sale for prices close and beyond $200 US...
...cut the prices of their RX 470 cards down to $169 US and also force it down to AIBs.

The rumoured specs for Vega 10 230W have 12Tflops against 7Tflops RX 580 with 130W TDP. That's 230W/130W = 77% increase in power usage for a 71% increase in Tflops. What concerns me is that there isn't a GTX1080Ti (or 2080Ti GP102) competitor on the road-map yet. It seems Vega 10 was designed to take on GTX1070/1080 and it seems a 275-300W Vega chip is missing for 2017.

Yeah the name seems to indicate as much.. Hence the "Fury Pro" name, and not XT/Rage/Fury/XT PE/whatever..
Have a feeling we haven't "seen" the highest SKU Vega10 yet

 

[vissarix]

kinda hard to believe the 50% better perf/watt with a revision, maybe if it was 10nm and an improved architecture i would have believed that...maybe they just lowered the clocks and voltage...

 

[happy medium]

so these are expected next year? Why are they waiting so long?
They are getting spanked by Nvidia now.

 

[Arachnotronic]

NVIDIA's Tesla P4 uses a low clocked GP104 to get a 5.5TFLOP (with boost clock enabled) product in a 50W power envelope. I am surprised WCCFTech didn't make up some nonsense about how this is a new revision of Pascal with >2x the perf/watt of GTX 1080...

http://images.nvidia.com/content/pdf/tesla/184457-Tesla-P4-Datasheet-NV-Final-Letter-Web.pdf

Anyway, it's obvious that these are embedded parts that will simply not run at max boost clock the whole time, which is how they are able to claim much lower power consumption with the same max TLOPs. This isn't a "new revision" of Polaris, just as Tesla P4 isn't using a "new revision" of GP104.

 

[ConsoleLover]

kinda hard to believe the 50% better perf/watt with a revision, maybe if it was 10nm and an improved architecture i would have believed that...maybe they just lowered the clocks and voltage...

Why is it hard to believe? The launch RX cards clearly had some issues with the 14nm process and instead of delaying for 3-4 MONTHS and loosing to Nvidia by default by being too late to the party, they decided to launch early with what they had.

Now that they've had few months to improve the 14nm process and improve the binning process, they can take full advantage of the smaller 14nm process. Essentially giving AMD even better than Nvidia levels watt performance.

 

[kraatus77]

just look at power to teraflops ratio (not performance) of both pascal and pol10, they are already similar. just that p10's perf/tflops is less so perf/w goes down compared to pascal. so what it means ? p10 is not power-hungry. just that it's not delivering performance it should. so result is lower p/w vs pascal . "50% more power efficiency" its just not possible. 50% more performance is possible via ipc increase but its not happening here since it's the same chip.

this is just bs hype from wccftech just like 1.6ghz thingy. they just read 1 thing and make their own news to hype and get clicks from amd/nvidia fanbois.

i'm saying this again, just ban that website and don't allow any posts referring to them as source. simple as that.

 

[Arachnotronic]

this is just bs hype from wccftech just like 1.6ghz thingy. they just read 1 thing and make their own news to hype and get clicks from amd/nvidia fanbois.

i'm saying this again, just ban that website and don't allow any posts referring to them as source. simple as that.

You win this thread, couldn't agree more.

 

[Arachnotronic]

Why is it hard to believe? The launch RX cards clearly had some issues with the 14nm process and instead of delaying for 3-4 MONTHS and loosing to Nvidia by default by being too late to the party, they decided to launch early with what they had.

Now that they've had few months to improve the 14nm process and improve the binning process, they can take full advantage of the smaller 14nm process. Essentially giving AMD even better than Nvidia levels watt performance.

You don't get 50%+ perf/watt improvements from process revisions, sorry. They'd be lucky to get 10% out of process enhancement alone, and that would probably require more substantial changes than...well, nothing.

 

[HurleyBird]

Whether or not there is a new revision, WCCTech seems like they're speculating and passing it off as something authoritative, no surprise.

And whether or not there is a new revision, binning also has to be significantly tighter on the mobile and embedded parts.

If there is a new revision, don't expect +50% perf/watt. It could happen, but it's not likely.

 

[kraatus77]

You win this thread, couldn't agree more.

I just don't get it, if anything the amd fans should be more pessimistic about these kind of rumors because if it turns out be false, it only ruins amd's image even more. and in last 4-5 years i've seen more hype and performance rumors about amd than intel/nvidia. and most of those are hyped by these type of websites.

why don't people learn ? you are only increasing expectations by hyping.

back to topic, most i can see is 5-10% lower power, or frequency bump @ same power. but this 50% less power is simply not possible.

for science -
does anyone have gtx1070? can you run gpu at same clocks as rx470 and measure power consumption? i want to know something. thanks.

 

[swilli89]

Why is it hard to believe? The launch RX cards clearly had some issues with the 14nm process and instead of delaying for 3-4 MONTHS and loosing to Nvidia by default by being too late to the party, they decided to launch early with what they had.

Now that they've had few months to improve the 14nm process and improve the binning process, they can take full advantage of the smaller 14nm process. Essentially giving AMD even better than Nvidia levels watt performance.

TBH there was probably 20% efficiency already on the table due to AMD dealing with poor bins at launch and overly conservative voltages. Tons of posts out there where people undervolt 480's and even get high clocks while doing so, greatly increasing efficiency. I doubt they could get 50% based on just a revision but I could definitely believe there are big improvements. Good news!

 

[Phynaz]

Pakistan. The center of the Technology World.

Or not.


Why do people give that site any credence?

 

[dark zero]

If I were running AMD, I'd release a 15-20% faster clocked RX485 refresh (or call it RX580, etc.) on the desktop at 140-150W TDP, and use the more efficient revisions for mobile dGPU OEM design wins. I feel that having a card that outperforms GTX1060 3-6GB by 7-10% is a lot more important on the desktop than having an RX 480 card with 95-100W TDP. If they push RX 470 down to $159-169 against upcoming $149 1050Ti, position RX 480 8GB at $199-209 against GTX1060 3GB and bring out a 15-20% faster RX485 at $239-249 against GTX1060 6GB, that would make all 3 of those AMD cards superior to NV's offerings, forcing NV to drop prices and/or do a refresh of Pascal.

The rumoured specs for Vega 10 230W have 12Tflops against 7Tflops RX 580 with 130W TDP. That's 230W/130W = 77% increase in power usage for a 71% increase in Tflops. What concerns me is that there isn't a GTX1080Ti (or 2080Ti GP102) competitor on the road-map yet. It seems Vega 10 was designed to take on GTX1070/1080 and it seems a 275-300W Vega chip is missing for 2017.

They are planning to do to counter Pascal Refresh.
Also, the RX 485 is supposed to have GDDR5X instead of normal one.

 

[kawi6rr]

They are planning to do to counter Pascal Refresh.
Also, the RX 485 is supposed to have GDDR5X instead of normal one.

If this card does launch and is under $300 I will have found my next upgrade.

 

[Bacon1]

The XX5 was posted a long time ago

http://www.guru3d.com/news-story/amd-possibly-will-be-using-xx5-revisions-e-g-amd-radeon-rx-495.html

 

[Glo.]

Khalid just went to Anandtech forum, read what I was posting and made whole article about the topic.

He even didn't bothered to check how it fares to efficiency of Polaris, the RX 480 GTR. It is 15% more efficient, not 50%.

I have to quote Didier Drogba. This is fu***ng disgrace. Disgraceful journalism. Nail to the coffin of that site.

 

[LTC8K6]

Would have been much better to have waited and simply introduced these cards instead of the others, if they are actually so much more efficient.

Also, such a rumor would surely make some people wait for these cards instead of buying now.

 

[raghu78]

You don't get 50%+ perf/watt improvements from process revisions, sorry. They'd be lucky to get 10% out of process enhancement alone, and that would probably require more substantial changes than...well, nothing.

Polaris chips run better with undervolting. AMD have basically pumped more voltage to qualify as many dies as possible. With changes to process and a new chip revision combined with better binning I think we could see a 30-40% perf/watt increase. If AMD bring G5x on Rx 485 , a 50% perf/watt increase is possible.

 

[Piroko]

just look at power to teraflops ratio (not performance) of both pascal and pol10, they are already similar. just that p10's perf/tflops is less so perf/w goes down compared to pascal. so what it means ? p10 is not power-hungry. just that it's not delivering performance it should. so result is lower p/w vs pascal . "50% more power efficiency" its just not possible. 50% more performance is possible via ipc increase but its not happening here since it's the same chip.

What? You're mixing completely different things with eatch other. What does IPC have to do with power consumption?

You don't get 50%+ perf/watt improvements from process revisions, sorry. They'd be lucky to get 10% out of process enhancement alone, and that would probably require more substantial changes than...well, nothing.

...Explain GTX 580.

This should clearly be taken as a rumor, but it's not the weirdest I've heard.

 

[Arachnotronic]

What? You're mixing completely different things with eatch other. What does IPC have to do with power consumption?

...Explain GTX 580.

This should clearly be taken as a rumor, but it's not the weirdest I've heard.

GTX 580 was a chip known as GF110, a distinct chip from GF100. It also didn't show up 3 months after GF100 did.

WCCFTech is preying on the AMD enthusiast base to generate clicks, even if they print unsourced, untrue garbage. They have also shown to be willing to prey on the NVIDIA and Intel fan bases but they realize that pro AMD headlines are the most profitable, hence the greater volume of AMD product promotion compared to the other two companies I mentioned.

 

[Arachnotronic]

Polaris chips run better with undervolting. AMD have basically pumped more voltage to qualify as many dies as possible. With changes to process and a new chip revision combined with better binning I think we could see a 30-40% perf/watt increase. If AMD bring G5x on Rx 485 , a 50% perf/watt increase is possible.

Pascal can be undervolted for dramatic perf/watt gains, too.

 

[Piroko]

GTX 580 was a chip known as GF110, a distinct chip from GF100. It also didn't show up 3 months after GF100 did.

I expected more from you. Ryan Smith noted that GF110 had no more than 6 months of development time after the bad rep launch of GF100 and most of its internal changes are copied from GF104. It runs circles around GF100. The only variable is the third transistor type that has an unknown impact (I would assume a rather small one or else it would've been used by marketing).