AMD: Speculation, Rumors, and Discussion (Archive)

[3dilettante]

No wonder 20nm was a waste of time - FinFET is pretty underwhelming. Oh well, fingers-crossed that optical computation gets here in the next 10 years.

The general comparison for that sort of iso-performance versus iso-power is from a circuit performance standpoint, generally referencing a CPU. If AMD wants a 2x transistor count device at the same clocks as the previous generation, the power consumption of the silicon would be within the current bounds.
On the other hand, AMD has more examples where the GPU silicon was pushed past its sweet spot at 28nm, so if Nvidia manages to any amount of generational improvement architecturally AMD might face the exact same pressures to do so again. FinFETs have a less impressive curve past their sweet spot versus 28nm, as one of the graphs showed.
The FinFET nodes are generally what 22/20nm should have been, and so it's a case of non-ideal process scaling yielding what amounts to a decent single-node transition being stretched out over two.

 

[Razor1]

That said, it doesn't seem like this card is targeting the GM107's power envelope. It's targeting the GM108's (Geforce 940M) power envelope at a performance close to the desktop GTX 950.
This would also mean that its performance is not between Pitcairn and Tonga, but rather between Bonaire and Pitcairn.
One of the slides says "Goal: Console-caliber performance in a thin-and-light notebook". That's why they compared it against a GTX 950, because it's the card from nVidia that performs closer to the PS4's Liverpool with 18 CUs and 32 ROPs. Plus, the GM107 with a 55W lowest TDP can't really fit inside a thin-and-light notebook (say, Macbooks and XPS 13).


Coincidentally, the GM108 is the chip coming in the Surface Book's keyboard. I'm guessing it would be in Microsoft's best interest in refreshing the SBook's dedicated GPU towards something that would match the XBone in performance, with little to no impact on battery life.

Firstly the 970 m can fit into the thin and light notebooks, take a look at razer blade, its .70 thickness notebook with a 970m, I have one, its probably the best thin notebook on the market, much better than any macbook or dell or hp out there performance wise. So the GM107 fitting into a ultra thin notebook will be easy to do lol.

They are comparing it to the gtx 950 because its the sale "level" or cost chip (MSRP), I think the frame rate lock is on because it saves AMD a good deal of power, without showing anything about true performance of their chip vs the competition. Because if they did it will give ammunition for nV to extrapolate possible performance figures and at this early of a stage, it wouldn't be prudent to do that because we saw what happened Fiji and the 980ti, conspiracy theory I know, but the way nV launched the Ti and timing and performance it was just too many coincidences.

Lastly MS will not introduce new GPU's into an already made product such as a surface book, because that would be requiring a total redesign from a thermal point of view for an ultra thin product. Which if you have ever seen a surface book 4 opened up, you will know there will be no way around at least no easy route so its a compete redesign. They may introduce a new version of the surface book by that time I think 5 might be a good number

 

[Razor1]

I'm talking about the implications for enthusiast discrete: utterly miserable. 30% more performance from the node change, before architectural improvements, after 5 years is just horrible.

Well that might not be what we see in the enthusiast discrete GPU's, we might see a bigger gain in performance without the power savings..... I mean why would they not try to go for the 250 watt power limit on the high end boards?

 

[Jawed]

"25-30% performance increase at the same power"

 

[Razor1]

well I sure hope not that would be a disappointment indeed

 

[Deleted member 13524]

Firstly the 970 m can fit into the thin and light notebooks, take a look at razer blade, its .70 thickness notebook with a 970m, I have one, its probably the best thin notebook on the market, much better than any macbook or dell or hp out there performance wise.

So the GM107 fitting into a ultra thin notebook will be easy to do lol.

Thin and light in 2016 means a laptop/convertible weighing 1-1.3Kg, thickness around 10mm and usually a screen size of 13.3". Also, they carry tiny ~50W power adapters at less than 150g.

As well built as your Razer Blade may be, it still weighs 2Kg, it's 18mm thick, not to mention a much larger 150W power supply at 350g.

No one would ever mistake your Razer Blade for a Dell XPS 13 or an Asus UX305.

"25-30% performance increase at the same power"

As far as I know, what they usually mean with "performance increase" when talking about new process nodes is that if you were to make an optical shrink of a given chip, you could clock it 25-30% higher without spending more power (e.g. the G92 to G92b transition allowed for ~15% higher clocks).
Of course, increasing clocks alone has never shown to be the best compromise, and IHVs prefer to increase transistor count and increase performance-per-clock a lot higher.
The 28nm->FinFet transition is supposed to allow at least twice the number of transistors within the same area. Even if the consumption-per-die-area at the same clocks ends up being somewhat worse in FinFet (I have no idea about this), the sheer increase in execution units is bound to be rewarding.

 

[Razor1]

err do those even use discrete GPU's, I don't remember any of those types of notebooks using discrete GPU's, at least not from Dell or HP or Apple. Why would that want to put in a discrete GPU into those types of notebooks? There is no use for them in them. Thats why I didn't even think about those notebooks, cause they haven't been using discrete in those for years now.

BTW the previous razer book did have a 750m in it.....

And I think you are talking about ultra portable notebooks, not thin notebooks.

 

[iMacmatician]

By the way, I added "Polaris" to the thread title to reflect AMD's official architecture name.

 

[ImSpartacus]

Lastly MS will not introduce new GPU's into an already made product such as a surface book, because that would be requiring a total redesign from a thermal point of view for an ultra thin product. Which if you have ever seen a surface book 4 opened up, you will know there will be no way around at least no easy route so its a compete redesign. They may introduce a new version of the surface book by that time I think 5 might be a good number

We see a new Surface Pro every year, so it's pretty likely that we'll get a new Surface Book in 2016 that could have a significantly better GPU. MS wasn't afraid of getting semi-custom GDDR5 940M, so it makes sense that they will actively pursue superior GPU performance in their power envelope.

 

[Razor1]

replacing a CPU that is pin compatible and has standardize settings is a bit different than the GPU modules. This is why many OEM's once they make certain design choices when making a product it takes time for them to switch to another. If and when AMD is ready to to make a custom module for the Surface, than MS has to validate and do their thing which also takes time, testing and what not. So from initial showing of a full compatible module to actual release of the end product lets say 6 months or so. That means right now AMD has to be ready to put something into the Surface to show up soon, Its a bit too soon.

Also the 940m surface book 4 was launched along with the entire line of Surface book 4, it wasn't put in later.

 

[silent_guy]

"25-30% performance increase at the same power"

But that's just the pure process improvement, right?
AMD didn't get the architectural perf/W upgrade that Nvidia got with Maxwell, so the real perf/W increase should be much higher for Polaris. Probably 100% or more. Not bad.

 

[Infinisearch]

"25-30% performance increase at the same power"

That's just taking into account process advantage right? Not including architecture tweaks right?

 

[iMacmatician]

From the PC Perspective article linked earlier:

How is Polaris able to achieve these types of improvements? It comes from a combination of architectural changes and process technology changes. Even RTG staff were willing to admit that the move to 14nm FinFET process tech was the majority factor for the improvement we are seeing here, something on the order of a 70/30 split. That doesn’t minimize the effort AMD’s engineers are going through to improve on GCN at all, just that we can finally expect to see improvements across the board as we finally move past the 28nm node.

 

[Jawed]

But that's just the pure process improvement, right?

Yes.

AMD didn't get the architectural perf/W upgrade that Nvidia got with Maxwell, so the real perf/W increase should be much higher for Polaris. Probably 100% or more. Not bad.

I can't see how they can get past 50% gain, when 70% of the gain is from process, amounting to "25-30%" and 30% of the gain (70/30 split) comes from architecture.

 

[gamervivek]

This might end up in apple's laptops first and only comes to desktop in mid-2016. Alongside the bigger chips.

 

[Infinisearch]

But that's just the pure process improvement, right?

Damn ninja.

 

[Razor1]

Yes.


I can't see how they can get past 50% gain, when 70% of the gain is from process, amounting to "25-30%" and 30% of the gain (70/30 split) comes from architecture.

Well not saying you are incorrect, or anything like that since we are speculating, I just can't see AMD not making any more that 30% performance increase per tier, if this is the case, I don't even think it would be worth upgrading to such a product, unless power is a real issue for some users.

Of course still don't know any thing about new GPU/API features yet.

 

[3dilettante]

Yes.


I can't see how they can get past 50% gain, when 70% of the gain is from process, amounting to "25-30%" and 30% of the gain (70/30 split) comes from architecture.

To reiterate my earlier point, and with further context from the Pcper article, the graph that is referenced by that claim is an fmax plot rather than overall graphics performance.

 

[silent_guy]

I can't see how they can get past 50% gain, when 70% of the gain is from process, amounting to "25-30%" and 30% of the gain (70/30 split) comes from architecture.

My bad: I didn't notice the part in Ryan's article about the majority of the gains being process related. Carry on.

 

[Jawed]

I didn't understand why you said "FinFETs have a less impressive curve past their sweet spot versus 28nm, as one of the graphs showed." when it appears to be the inverse to me: FMax increases to its maxima at nearly 3x power cost of the sweetspot on 28nm versus about 2.5x on FinFET.

I put the sweet spots at 1.5 and 1 units along Relative Power for 28nm and FinFET, respectively.

---

70/30 split and the 25-30% gain from process are both statements referenced to actual GPUs. The graph can be derived from an actual chip, since it's up and running.

GTX 950 has a base clock of 1024MHz and boosts to 1188MHz. AMD appears to spending huge amounts of transistors to get "equivalent performance" at 850MHz. The NVidia chip has a 20% clock advantage, so AMD must be using more transistors for "par performance". GM108 has a base of 1029MHz, if we want to talk about dies of roughly the same size (NVidia doesn't need to crash to a lower clock for performance per watt at the low end).

This appears to be exactly what we see when comparing AMD and NVidia at 28nm.

Put another way: on 28nm NVidia has a graph that looks something like the yellow line (sweet spot is at a higher frequency and at less power) and AMD is on something like the dotted-blue line. It appears that we'll see the same situation repeat on FinFET.