AMD: Speculation, Rumors, and Discussion (Archive)

[CarstenS]

Reviews tend to not think about Tonga when reviewing the 980, and in the reverse. The card that gets significantly higher FPS being hobbled by a CPU that draws more power tends to weigh the comparison in a manner more complementary to Tonga than not.

The former is quite understandable, considering GM204 was out before Tonga (in the Wintel-space). Your latter assumption I do not really understand: The 980 is 75-86% faster as the 380 in the data presented which requires the CPU to feed it faster, thus consuming more power (from what I could gather out of AT's recent R9 Nano benchmarks it'2 an OC'ed 4960X), thus contributing more to the total power figure. If the CPU was left out of the equation, we'd have the 980 consume even less power than R9 380.

That is, If we assume, the 980 was not slowed down by the CPU - which we cannot know.

 

[3dilettante]

If the CPU was left out of the equation, we'd have the 980 consume even less power than R9 380.

That would mean that a whole-system figure would obscure more of the 980's superiority, making Tonga look "better" as far one can look better with a performance deficit such as that.

It makes reality less flattering for Tonga, and potentially less impressive for an architecture claiming scaling numbers using it as a baseline.

 

[iMacmatician]

I thought I'd make a new thread for Vega since AMD's roadmap shows Vega as being distinct from Polaris.

To estimate the perf/W and dates of the upcoming architectures, I placed a scatterplot over a larger version of this slide, with the x-axis on the thin white line, and adjusted the values and axis scales until the data points aligned with (the centers of) the red squares.

With a base value of 1.00 for the 28 nm GPUs square, I obtained 2.56 for the Polaris square, 3.84 for the Vega square, and 5.11 for the Navi square (to 2 decimal places). I think that is overprecision for this multi-year roadmap with no scale on the y-axis, but the significance of these values is that the latter three satisfy a 1:1.5:2 ratio almost exactly. So I'm guessing that perhaps Vega and Navi are targeted/estimated to have about 1.5x and 2x the perf/W of Polaris, respectively.

Since Polaris will be released in mid 2016, the roadmap places Vega in late 2016 / early 2017 and Navi in late 2017 / early 2018. There will also be two Vega chips: Vega 10 and Vega 11.

 

[silent_guy]

I wonder if AMD/SK´s close relationship over many years, if this wasn't seen and worked around already. I'm really just going on what the roadmap says about Vega and what Nvidia says about HBM2 in their presentation though.

There really isn't anything to work around: page activations are the only way to access data in a DRAM. There is no way around that.

I think the Vega vs Polaris perf/W difference is there due to HBM vs GDDR and maybe because they had some addition time to optimize other on-chip logic. (They say new architecture, but that's probably another incremental thing.)

 

[CarstenS]

That would mean that a whole-system figure would obscure more of the 980's superiority, making Tonga look "better" as far one can look better with a performance deficit such as that.

It makes reality less flattering for Tonga, and potentially less impressive for an architecture claiming scaling numbers using it as a baseline.

So you're constructing a worst case scenario basis? Ok.

 

[3dilettante]

So you're constructing a worst case scenario basis? Ok.

It would be a very conservative scenario that would give Tonga some benefit of the doubt.
The question would be how poorly would X need to scale with FinFET before it loses to Polaris. Perhaps a better letter would be N, but I have been eyeballing what the factor would look like even between 2.5X Tonga and what Hawaii, Fury, and Fury Nano achieved.

 

[Jawed]

We've approached the point where memory density is no longer a constraint on what developers can do in games. There are no cries for more than 8GB of VRAM in gaming. With HBM2, 8GB is effectively the minimum amount of memory for performance or enthusiast cards and 16GB won't be any sweat.

That could imply that power-constrained ultra-high-performance memory could spend transistors on radically different architectures - e.g. a hybrid of SRAM and DRAM.

Surely there must be more depth in this topic than a graph.

 

[Jawed]

[...] between 2.5X Tonga and what Hawaii, Fury, and Fury Nano achieved.

A proper characterisation consisting of changes in power consumption for clock deltas with baselines of 80%, 90%, 100% and 110% of official boosts clocks for both 380X and 980 would be pretty interesting.

e.g. the power consumption delta for 380X given a 10% overclock from stock clocks is P, yet a 10% overclock from 80% of stock clock results in a power delta of P/2 (totally imaginary). Same-percentage overclocks from lower clocks, obviously, should have a smaller delta: the question is how do the competing chips' curves compare to each other.

I expect AMD's is pretty ugly.

 

[ImSpartacus]

It looks like AMD is finally adopting a codename system similar to Nvidia's where they have one "generation" codename (e.g. Kepler, Polaris) and then particular GPU codenames that link back to that generational codename (e.g. GK104, Polaris 10).

However, where we have a good idea where a "G*104" part will slot in Nvidia's lineup, will we have a similar guidepost with something like "* 10" or "* 11"?

I heard second hand that Koduri said that the numbers are just a raw count of the given GPUs in that generation and don't denote anything. That would make sense since AMD would probably want to use Vega 10 to fill out a different part of their lineup than Polaris 10 and so on.

 

[SimBy]

Wouldn't it be easier to extrapolate from actual Polaris demo running Star Wars capped at 60fps? We already know how it compares to GTX 950 in identical system. All you really need is a good estimate of system minus GPU power consumption.

 

[Adored]

I'm pretty sure that a lot of AMD's power issues on 28nm came from the variability of the process, and due to economical/yield reasons they have quite a large variability in their cards. I don't think we can take any one GPU result as gospel here.

@SimBy I did that before, came up with a conservative estimate at 40W for Polaris the demo. I now believe I may have been too conservative now though and 30W is likely to be closer, if not less.

 

[Clukos]

When you say variability you mean leakage?

 

[Adored]

Yes sorry I meant leakage, I tried to edit it but can't still.

 

[Adored]

Also in that pic I assumed a very good PSU at 90% efficiency, however it was pointed out later that due to the tiny power draw of the Polaris system especially, 85% or less is far more likely - so the draw is very likely to be less than 40W for the part shown.

 

[silent_guy]

I'm pretty sure that a lot of AMD's power issues on 28nm came from the variability of the process, and due to economical/yield reasons they have quite a large variability in their cards.

I'm pretty sure that's just not the case.

You're dismissing architectural efficiency and low level power optimizations, yet they are crucial.

With Maxwell being targeted for mobile chips as well at the time (and having much more manpower at its disposal to work on the menial task that is needed for low level optimizations), it was essential for Nvidia to get both of those right.

AMD put its money on HBM to get some short term power relief and probably used the next gen process to get architecture and low level stuff right. Both decision are rational given the constraints they had.

perf/W for Polaris should be to be very similar to Pascal. (Though I still question were that Vega bump will be coming from...)

 

[Adored]

I'm pretty sure that's just not the case.

You're dismissing architectural efficiency and low level power optimizations, yet they are crucial.

With Maxwell being targeted for mobile chips as well at the time (and having much more manpower at its disposal to work on the menial task that is needed for low level optimizations), it was essential for Nvidia to get both of those right.

AMD put its money on HBM to get some short term power relief and probably used the next gen process to get architecture and low level stuff right. Both decision are rational given the constraints they had.

perf/W for Polaris should be to be very similar to Pascal. (Though I still question were that Vega bump will be coming from...)

It's obvious that AMD didn't do the same low-level transistor improvements that Nvidia did on 28nm. Their engineers are talking about leakage in videos though, should be a hint that something wasn't right for them on 28m,

It makes sense from an economical pov and AMD did it before with Brazos and various other CPU parts I believe (Llano maybe?). Some of their 28nm stuff undervolts massively as well, Hawaii for one.

HBM on Fury was more likely about getting it done before 14nm rather than a last ditch attempt to save power on 28.

I agree that Pascal and Polaris is all about architecture though and I'm also scratching my head at the Vega increase.

 

[Grall]

Some of their 28nm stuff undervolts massively as well, Hawaii for one.

Interesting. Can undervolting of these cards be done without hardware modifications?

You have any links to pages that discuss potential gains (reduction in volts, thermal output, and so on) from doing this? Seeing as I have two R390Xes burning power day and night this would be quite interesting to me.

 

[Clukos]

Interesting. Can undervolting of these cards be done without hardware modifications?

You have any links to pages that discuss potential gains (reduction in volts, thermal output, and so on) from doing this? Seeing as I have two R390Xes burning power day and night this would be quite interesting to me.

Thanks to pro googling skills, voila!

from this thread: http://www.overclock.net/t/1561704/official-amd-r9-390-390x-owners-club/1460

 

[Adored]

My own Tonga (285) undervolts by 100mV also but it's not a great overclocker.

You should be able to do it with a prog like MSI Afterburner or Trixx.

 

[silent_guy]

In the end, AMD and Nvidia are using the same process from the same fab. And likely the same standard cells provided to them
by TSMC. The chance that AMD hit leakage issues and Nvidia did not are not very high.