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

[dark zero]

wccctech rumor = made up by wccftech based on one misrepresented, out of context fact somewhere multiplied by all the unmet expectations of who always want more

Actually it has some basis... and comes from the embedded market this time... so he is only theorizing about the refresh (which AMD already hinted it), so that refresh is not as crazy to expecting.

Also considering some aspects:
- RX 460 is NOT full Polaris 11 is a direct hint that the full one will come soon.
- RX 480 is memory speed bottlenecked in some games is another hint.

However, that 50% of improvements is a joke, so is not to being taken seriously, but otherwise in other aspects it might have some truth.

A refresh is likely to happen before Vega lands.

 

[krumme]

Asuming the througput is indicative of freq for p11 embedded as Ryan writes there is something weird going on.

http://www.anandtech.com/show/10710/amd-announces-embedded-radeon-e9260-e9550

I was under the impression a metal layer respin was more done for yield? And 1400 is high. Boost or not doesnt matter here.

Wasnt a metal layer 3 to 4 months? While a transistor layer is more like 1 year.

Clearly there is lot of process variation at play at gf and 480 is standard oc to the moon. My 470 runs 1340 at a low 1087mv others run only 1266 but at 950mv. Its crazy stuff.

If we see eg say 25% efficiency improvement its imo not so much an improvement as prior gen but more a fix of several issues.

Metal layer respin in combination with less process variation mean both higher freq and lower voltage due to a completely new more agressive binning and bios control.

 

[dark zero]

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

And can go for phones and are better cryptominer than AMD cards /sarcasm

 

[TimCh]

Actually it has some basis... and comes from the embedded market this time... so he is only theorizing about the refresh (which AMD already hinted it), so that refresh is not as crazy to expecting.

Also considering some aspects:
- RX 460 is NOT full Polaris 11 is a direct hint that the full one will come soon.
- RX 480 is memory speed bottlenecked in some games is another hint.

However, that 50% of improvements is a joke, so is not to being taken seriously, but otherwise in other aspects it might have some truth.

A refresh is likely to happen before Vega lands.

50% seems quite incredible, but remember that a 50% increase in efficiency is "only" a 33.3% decrease in power consumption.

Sendt fra min SM-G930F med Tapatalk

 

[Piroko]

I was under the impression a metal layer respin was more done for yield?

From a physical perspective a metal layer respin is done to improve signal quality and fix internal routing issues. Nowadays there's also stuff like integrated inductors in the thicker metal layers, I think AMD patented internal resonators a couple years ago? It's hard to tell what exactly a metal layer respin will do from an external viewpoint.

Wasnt a metal layer 3 to 4 months? While a transistor layer is more like 1 year.

Yeah, I believe so too. Metal layers 6 months in normal planning and probably 3 to 4 months if rushed, transistor layers up to a year.

 

[krumme]

We need some of the guys from cpu to chime in but as i understand better signal quality means better SN ratio giving oportunity for lower V and also higher freq?
I just think its years back when we saw a metal respin giving remotely those kinds of benefit. Perhaps the original metal layers was not fit for the process? Meaning amd didnt know gf process well enough or didnt have the right tools or more likely the right specs? Or the process ended different than anticipated? Lol there is a gazilion possibilities.

 

[Abwx]

We need some of the guys from cpu to chime in but as i understand better signal quality means better SN ratio giving oportunity for lower V and also higher freq?
I just think its years back when we saw a metal respin giving remotely those kinds of benefit. Perhaps the original metal layers was not fit for the process? Meaning amd didnt know gf process well enough or didnt have the right tools or more likely the right specs? Or the process ended different than anticipated? Lol there is a gazilion possibilities.

Given the timeline it s a respin of the power deliveries.

The supply voltage inside the GPU is not clean, serial resistances and inductances of the power lines limit the frequency at wich the logics can be switched, because as frequency is increased those parasistic elements become the real loads attached to the transistors, hence noise in the signal path will increase accordingly rendering the circuit unfunctional, at the limit it s no more a GPU but a bunch of transistors switching power in inductive loads (the supply routings) and providing no more any coherent signals between logical blocks...

 

[PPB]

Well, Vishera-K did exist and was a really good improvement on the same node for the same uarch, regarding the voltage/clock curve and power efficiency in the end.

But we have to think that revision came almost 2 years later than the first Vishera iteration. So probably this 'news' is just hyping a new lineup based on super binned chips. Nothing to see here.

PS: If someone asks me "would a Polaris new stepping a la Vishera, with improved voltage curve, more OC headroom and better perf-watt paired with GDDR5X be awesome". Totally, it would be a really exciting product and probably the last non-HBM card I would buy. But probably such a SKU would fit better for a RX 580 than the rumored RX 485 implied in the article.

 

[ZGR]

Why is it so hard to contemplate a 50% perf/watt increase on the same process node?

We are not talking about maintaining or increasing the overall performance, but rather increasing efficiency and taking a hit in performance. It could be a combination of finer voltage tuning, better binning, and a cut in maximum clockspeed. These will be mobile first chips plain and simple.

 

[ConsoleLover]

Latest reports have it at over 60% efficiency, which in essence means around 39% power draw reduction, so from 150W to around 90W at the same clocks. If AMD increases clocks though and offers a 15% increase in performance for the same price at 150W that would be a best seller.

GTX 480 8GB at $240 that is 15% faster at the same power draw would be amazing.

 

[krumme]

Why is it so hard to contemplate a 50% perf/watt increase on the same process node?

We are not talking about maintaining or increasing the overall performance, but rather increasing efficiency and taking a hit in performance. It could be a combination of finer voltage tuning, better binning, and a cut in maximum clockspeed. These will be mobile first chips plain and simple.

If clocks goes down 20% is not that hard to contemplate but what is the value then? Competing with small nv 128bit gpu.

I am sure its the mobile chips comming. 4 month late. If we look at p11 embedded we have 1400MHz. Boost or not nothing but a metal respin enables that imo. How many p11 runs 1400 now? Its not at all viable with the current gpu.

At same clocks even 40% is insane.

 

[TimCh]

If clocks goes down 20% is not that hard to contemplate but what is the value then? Competing with small nv 128bit gpu.

I am sure its the mobile chips comming. 4 month late. If we look at p11 embedded we have 1400MHz. Boost or not nothing but a metal respin enables that imo. How many p11 runs 1400 now? Its not at all viable with the current gpu.

At same clocks even 40% is insane.

40% increase in performance per watt, is a 29% drop in power. It is a huge drop but not completely impossible.

Sendt fra min SM-G930F med Tapatalk

 

[Piroko]

Read a bit more about it:

Seems like the XFX card JayZ has tested has a different Bios (modified by XFX) and a different power controller. It's possible that the card only reported core power draw (110W reference) instead of board power draw (~160W reference). A core power draw of 85~90W with max boost clocks is still impressive (~20% better), but in the realms of chip lottery. We might still see a tendency towards these better performing chips in the future if we go by past GloFo tendencies (process is slowly improved+stabilized over time) but certainly no 50% jump from that.

AMDs embedded GPUs have always had much lower TDP, probably as a combination of better binning and lower average clocks. That alone can't hold up as a source.

 

[krumme]

No but the 1400 is what makes it interesting.

 

[Erenhardt]

Read a bit more about it:

Seems like the XFX card JayZ has tested has a different Bios (modified by XFX) and a different power controller. It's possible that the card only reported core power draw (110W reference) instead of board power draw (~160W reference).

Seems you didn't read enough. All 480 (reference/custom) report only core power draw. The reported power is not a whole card power consumption.

Still his XFX takes less watts than other 480s, runs on lower voltages, and overclocks way better and at lower power, even compared to superior watercooled solution.

 

[Mopetar]

50% seems high, but it's possible if you make some assumptions:

1) The GF process was not mature or had serious issues at Polaris launch. This seems like a given at this point.

2) We're comparing the best case for the new chips against the worst for the current chips. There are some 480s that throttle at stock settings and others that get a good OC while undervolting. It isn't 50% over the chips that are hitting better power or performance numbers.

3) AMD has a better cooling solution. If the chips don't run as hot and need as much voltage, cooling is easier.

 

[krawcmac]

There is this one thing that bothers me. We are talking about drop o power consumption from 150W to 95W for R9 480. Isn't it more like 35% drop not 50%? Or am I mathematically challenged?

 

[Glo.]

There is this one thing that bothers me. We are talking about drop o power consumption from 150W to 95W for R9 480. Isn't it more like 35% drop not 50%? Or am I mathematically challenged?

The MSI Afterburner reads GPU die power consumption. Typically, in normal circumstances we have seen that GPU die power consumption is 110-120W for RX 480 GPUs. The rest is consumed by the board.

In XFX RX 480 GTR the power consumption of the die is 85-95W, with peaks in 100-105W. That is drop by up to 15%.

As for Embedded GPUs the key words here are: Up to...

 

[Det0x]

There is this one thing that bothers me. We are talking about drop o power consumption from 150W to 95W for R9 480. Isn't it more like 35% drop not 50%? Or am I mathematically challenged?

While iam only making up numbers now, it could look something like this:

150/95 = ~1.57 = 57%

Or we could say you play a game @ 100 fps, with 3 different cards, drawing 3 different amounts of power. (95,120 and 150 watts during load)

100 fps / 150 watt = ~0.67 watt per fps

100 fps / 120 watt = ~0.83 watt per fps

100 fps / 95 watt = ~1.05 watt per fps

1.05 watt per fps / 0.67 watt per fps = 56% increase in efficiency (best case)

1.05 watt per fps / 0.83 watt per fps = 26% increase in efficiency (probably a more down to earth figure)

It all depends on perspective..

ps, please don't flame my "math", it was only made as an visual example

 

[sxr7171]

Can be done if clock speed limited for mobile. They aren't talking about absolute performance increases at the top.

 

[ElFenix]

1.4Ghz for Polaris 11 should've been the lede but Khalid has to hype AMD to the moon. He should get paid by AMD if he isn't already.

why? setting expectations too high doesn't help amd.

 

[TrentN]

Xbox Scorpio: 6 TFLOPS, 320GBGB/s, unknown amount and type

PS4 Pro: 4.12 TFLOPS, 8GB GDDR5, unknown bandwidth

Microsoft gets close to 50% improvement with the same 14nm technology. The new GCN architecture appears to be the answer to this riddle.

 

[HurleyBird]

Or a bigger chip + more mature process. I don't doubt that Scorpio will use a more advanced GCN iteration which will impart some amount of benefit, but comparing the raw numbers of Scorpio to PS4 pro isn't hardly proof of anything. At least not with other numbers like wattage or die size.

 

[TrentN]

Mature process + new GCN architecture + bigger chip. There is never going to be proof when we are guessing.
But, an important clue is that Microsoft said that they decided to delay the Scorpio 1 year behind the PS4 Pro to wait for something. They did not delay for bigger chip, and I don't think they delayed for mature process.
I agree a new architecture almost never increases performance 50%. Maxwell architecture increased performance 35% and AMD says that ZEN architecture increases performance 40%.
My guess, Microsoft is using a larger die to gain a lot of performance, but requires the new GCN circuits to keep power of the larger chip below their power budget.

 

[Headfoot]

Why is it so hard to contemplate a 50% perf/watt increase on the same process node?

We are not talking about maintaining or increasing the overall performance, but rather increasing efficiency and taking a hit in performance. It could be a combination of finer voltage tuning, better binning, and a cut in maximum clockspeed. These will be mobile first chips plain and simple.

If BenchLife or Anandtech reported a 50% increase in perf/watt I'd be asking "How are they doing it?" Because WCCFTech reported it, it's likely made up